Escolar Documentos
Profissional Documentos
Cultura Documentos
OF MEDICAL DEVICES
by
Sharareh Taghipour
A thesis submitted in conformity with the requirements for the degree of Doctor of
Philosophy
Sharareh Taghipour
Doctor of Philosophy
University of Toronto
2011
ABSTRACT
For decades, reliability engineering techniques have been successfully applied in many
inspection and optimization models are developed and widely used to achieve maintenance
excellence, i.e. the balance of performance, risk, resources and cost to reach to an optimal
solution. However, the application of all these techniques and models to medical devices is new.
Hospitals, due to possessing a large number of difference devices, can benefit significantly if the
optimization techniques are used properly in the equipment management processes. Most
research in the area of reliability engineering for medical equipment mainly considers the
devices in their design or manufacturing stage and suggests some techniques to improve the
reliability. To this point, best maintenance strategies for medical equipment in their operating
We aim to address this gap and propose methods to improve current maintenance
strategies in the healthcare industry. More specifically, we first identify or propose the criteria
ii
which are important to assess the criticality of medical devices, and propose a model for the
prioritization of medical equipment for maintenance decisions. The model is a novel application
according to their criticality. The devices with high level of criticality should be included in the
Then, we propose a method to statistically analyze maintenance data for complex medical
devices with censoring and missing information. We present a classification of failure types and
establish policies for analyzing data at different levels of the device. Moreover, a new method for
trend analysis of censored failure data is proposed. A novel feature of this work is that it
considers dependent failure histories which are censored by inspection intervals. Trend analysis
Finally, we introduce some assumptions based on the results of the analysis, and develop
several new models to find the optimal inspection interval for a system subject to hard and soft
failures. Hard failures are instantaneously revealed and fixed. Soft failures are only rectified at
inspections. They do not halt the system, although they reduce its performance or productivity.
The models are constructed for two main cases with the assumption of periodic inspections, and
periodic and opportunistic inspections, respectively. All numerical examples and case studies
presented in the dissertation are adapted from the maintenance data received from a Canadian
hospital.
iii
To My Parents
iv
ACKNOWLEDGEMENTS
Dr. Dragan Banjevic for giving me the opportunity to do my PhD, and for their excellent advice,
insight, support and encouragement during the course of my doctoral studies. This thesis would
not have been completed without them. Words are inadequate to express my special appreciation
to Dr. Dragan Banjevic, who helped me patiently with his excellent knowledge all throughout
I gratefully thank Professor Dionne Aleman and Professor Tom Chau (from the
University of Toronto) who served as the members of my research committee, and provided me
with valuable comments on and insights into both the research and the presentation of the thesis.
I am very thankful to Professor David Coit (from Rutgers the State University of New
Jersey) and Professor Timothy Chan (from the University of Toronto) who served as the
Many thanks go to Mr. John Leung from Toronto General Hospital, who gave me access
to the maintenance data of medical devices and devoted his time and expertise to answer my
v
questions about the data. The data he provided me made it possible to present the application of
I wish to express my appreciation to Dr. Elizabeth Thompson from C-MORE lab, for her
Thanks also to Mr. Neil Montgomery from C-MORE lab for his useful comments on my
I also benefited from the advice, experience, and friendship of a great group of friends
and colleagues at the C-MORE lab during these past years. My thanks especially go to Dr. Ali
Zuashkiani, research associate, and Dr. Nima Safaei, postdoctoral fellow at C-MORE.
I wish to acknowledge the Natural Sciences and Engineering Research Council (NSERC)
of Canada, the Ontario Centre of Excellence (OCE), and the C-MORE Consortium members for
Last, but not least, I would like to express my deepest gratitude to my family, especially
my parents, who always believed in me and supported me. I would have achieved for less
without their continuous support. They have been my greatest mentors, and I owe all my success
to them.
Sharareh Taghipour
vi
PREFACE
Most of the chapters presented in this dissertation are the extensions of the following
journal and conference papers, which are published, accepted or submitted during my PhD
studies:
Journal Papers
Data for Complex Medical Devices. Quality and Reliability Engineering International 27(1):
71-84. This paper won the Best Student Paper Award 2010 of the American College of
for a Complex Repairable System. Reliability Engineering and System Safety 95(9): 944-952.
[Used in Chapter 5]
vii
dx.doi.org/10.1057/jors.2010.106). This paper received the 3rd place at the Best Student
Paper Award 2011 of the American College of Clinical Engineering (ACCE). [Used in
Chapter 3]
Repairable System Subject to Hidden Failures. IEEE Transactions on Reliability. 60(1): 275-
5. Taghipour S and Banjevic D (2011). Trend Analysis of the Power Law Process Using
6. Taghipour S and Banjevic D (2010). Optimum Inspection Interval for a System Under
[Used in Chapter 5]
7. Taghipour S and Banjevic D (2011). Maximum likelihood estimation from interval censored
recurrent event data. Computational Statistics & Data Analysis (submitted in March 2011).
[Used in Chapter 4]
viii
Papers Published in the Conference Proceedings
10. Taghipour S, Banjevic D, and Jardine A K S (2010). An Inspection Optimization Model for a
System subject to Hidden Failures. Proceedings of the 2010 ICOMS Asset Management
11. Taghipour S and Banjevic D (2011). Trend Analysis of the Power Law Process with
Censored Data. Proceedings of the Annual Reliability and Maintainability Symposium 2011.
[Used in Chapter 4]
ix
TABLE OF CONTENTS
1. INTRODUCTION....................................................................................................................1
x
TABLE OF CONTENTS
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data ...............63
4.2.3. Proposed Policy for Analyzing Soft and Hard Failures .........................................69
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring ...............81
xi
TABLE OF CONTENTS
5.3.3. Minimal Repair and Replacement of Soft and Hard Failures .......................160
7. REFERENCES .....................................................................................................................185
8. APPENDICES ......................................................................................................................203
APPENDIX B- Calculating the lower bound for the total score value .................................209
xii
TABLE OF CONTENTS
APPENDIX K- Upper bound for V (t ) Vk (t ) ...............................................................224
k 1
INDEX .........................................................................................................................................307
xiii
LIST OF TABLES
Table 3.2a. Assessment of the devices with respect to “Function”, “Mission”, “Age” and
“Risk” .................................................................................................................. 43-44
Table 3.2b. Assessment of the devices with respect to “Recalls” and “Maintenance
requirement”, and total, normalized and transformed criticality scores.............. 44-45
Table 3.3. Assessment of the devices with respect to “Mission criticality” sub-criteria ........ 45-46
Table 4.1. The total number of scheduled and non-scheduled work orders and findings at
inspection .....................................................................................................................64
Table 4.2. Qualitative tests results extracted from scheduled and non-scheduled work orders ....65
Table 4.3. Quantitative tests results extracted from scheduled and non-scheduled work orders ..66
Table 4.4. PM checks results extracted from scheduled and non-scheduled work orders ............66
Table 4.5. Soft and hard failure frequencies at scheduled and non-scheduled work orders .........69
xiv
LIST OF TABLES
Table 4.6. Estimated parameters of Weibull distribution for times between events –
system level..................................................................................................................75
Table 4.7. Estimated 95% confidence limits for the Weibull distribution parameters .................76
Table 4.8. An overview of the results obtained from statistical analysis at different levels .........77
Table 4.9. Sample data for three units’ failure and censoring events ...........................................91
Table 4.11. Estimation and testing results obtained for audible, housing and battery units .........93
Table 4.15. Total number of iterations in the complete EM, modified EM and the mid-points
method..........................................................................................................................97
Table 5.1. Parameters of the power law intensity functions and costs........................................116
Table 5.3. Parameters of the power law intensity functions, costs and rj ( x) parameters for
Table 5.5. Parameters of the power law intensity functions, and costs for components with soft
and hard failure ..........................................................................................................151
Table 5.6. Expected cost calculated for 1, 2,...,12 in Figure 5.14 .........................................152
Table 5.7. Costs and parameters of the probability function for minimal repair of components
with soft failure ..........................................................................................................158
Table 5.8. Expected costs calculated for 1, 2,...,12 in Figures 5.15 and 5.16 .......................158
Table 5.9. of the power law intensity functions, and costs for components with soft
xv
LIST OF TABLES
Table 5.12. The costs for different components with hard failure ..............................................173
Table A.1. Pairwise comparison matrix for the grades of criterion “Function” .........................204
Table B.1. Calculating the minimum total score value ....................................................... 209-210
Table C.1. Components/features that can be ignored since they have no failure reported .........211
Table C.2. Components/features that have more failures at non-scheduled work orders (hard
failures) ......................................................................................................................211
Table C.3. Components/features that have more failures at scheduled work orders
(soft failures) ..............................................................................................................212
Table F.1. Estimated parameters of Weibull distribution for times between events – hard failures
xvi
LIST OF TABLES
Table F.2. Estimated parameters of Weibull distribution for times between events – soft failures
LA=11.08 (systems are degrading) ............................................................................217
Table F.3. Estimated parameters of Weibull distribution for times between events – Alarms
failures (hard) LA=4.13 (systems are degrading) ....................................................218
Table F.4. Estimated parameters of Weibull distribution for times between events –
Table M.1. Estimated expected number of minimal repairs and uptime and their standard errors
calculated for the components with soft failure, for 1, 2,...,12
(used in Fig. 5.13)............................................................................................ 231-232
Table M.2. Expected number of minimal repairs and replacements, and the uptimes and their
standard errors calculated for the components with soft failure, for 1, 2,...,12
(used in Figures 5.15 and 5.16) ........................................................................ 234-235
Table P.1. Audible signal dataset for 80 units’ failure and censoring events ..................... 241-242
Table P.2. Housing/Chassis dataset for 38 units’ failure and censoring events.................. 243-244
xvii
LIST OF FIGURES
Figure 2.2. Acquisition iceberg (source: Cheng and Dyro, 2004) ................................................11
Figure 2.3. Major tests and actions performed during a device’s life cycle .................................16
Figure 4.3a. The policy for analyzing failures at the system level ...............................................70
Figure 4.3b. The policy for dealing with “no failure” devices at the system level.......................70
Figure 4.4. The policy for dealing with hard failures – component level .....................................71
Figure 4.5. The MCF plot for the failures at the system level (0.5*MTBF).................................74
xviii
LIST OF FIGURES
Figure 4.6. Scheme of hard failures and scheduled inspections in a cycle ...................................81
Figure 4.8. Lower and upper bounds for failure times and the last censoring time ......................83
Figure 4.10. Damaged sealed lead acid battery caused by battery overcharging .........................94
Figure 5.4. Total costs for different inspection frequencies (first model) ..................................117
Figure 5.5. Total costs of scheduled inspections for different inspection frequencies (threshold
model) ........................................................................................................................123
Figure 5.6. Total costs of scheduled inspections for different inspection frequencies (threshold
model); For penalty cost cP $1700 , n 9 is the optimal number of inspections
Figure 5.9. Total costs for different inspection intervals ( ) – finite horizon ...........................135
Figure 5.10. Total costs for different inspection intervals ( ) – infinite horizon ......................140
Figure 5.11. 5 is obtained with downtime penalty costs from the last column
of Table 5.4 ................................................................................................................141
Figure 5.12. Scheme of scheduled and non-scheduled inspections in a cycle ............................142
Figure 5.13. Hard and soft failures in the first inspection interval .............................................146
Figure 5.14. Expected costs for different inspection intervals ( ), assuming soft and hard
failures are minimally repaired..............................................................................153
xix
LIST OF FIGURES
Figure 5.15. Expected costs for different inspection intervals ( ), assuming soft failures are
either minimally repaired or replaced ...................................................................159
Figure 5.16. Expected costs for different inspection intervals ( ), assuming soft failures are
either minimally repaired or replaced with different replacement costs ...............160
Figure 5.17. Expected costs for different inspection intervals ( ), assuming soft failures are
either minimally repaired or replaced ...................................................................167
Figure 5.18. Expected costs for inspection intervals ( ), assuming soft and hard failures are
either minimally repaired or replaced with the replacement costs given in Table
5.11 ..........................................................................................................................168
Figure 5.19. Expected costs for different inspection intervals ( ), assuming preventive
replacement of component with hard failure, with the costs given in the third
column of Table 5.12................................................................................................174
Figure 5.20. Expected costs for different inspection intervals ( ), assuming preventive
replacement of component with hard failure, with the costs given in the last
column of Table 5.12................................................................................................174
Figure F.1. The MCF plot of hard failures .................................................................................216
xx
LIST OF APPENDICES
APPENDIX B- Calculating the lower bound for the total score value .................................209
APPENDIX K- Upper bound for V (t ) Vk (t ) ...............................................................224
k 1
xxi
LIST OF APPENDECES
xxii
ABBREVIATIONS
CT : CAT Scan
EM : Expectation-Maximization
xxiii
ABBREVIATIONS
PM : Preventive Maintenance
xxiv
1. INTRODUCTION
Nowadays from 5,000 to more than 10,000 different types of medical devices can be
found in an average to large sized hospital. Hospitals and healthcare organizations must ensure
that their critical medical devices are safe, accurate, reliable and operating at the required level of
performance. To achieve these objectives, hospitals must establish and regulate a Medical
and it should be reviewed and improved continuously in order to keep up with the pace of
healthcare organizations.
Maintenance excellence is the balance of performance, risk, resource inputs and cost to reach to
Although maintenance strategies and techniques have been significantly improved in the
last two decades, most of hospitals and healthcare organizations do not benefit from maintenance
1
1. INTRODUCTION 2
could be also loss-making likewise inadequate level of maintenance. The time, which is spent
(JCAHO) introduced standard EC.6.10 (JACAHO, 2004), hospitals in US have started adopting
their maintenance programs to put their maintenance resources where most needed. This
standard allows hospitals to not have schedule inspection or maintenance tasks for certain pieces
or types of medical equipment, if these tasks are not needed for safe and reliable operation
However, in Canada, most, if not all healthcare organizations include all their medical
equipment in their maintenance program and just follow manufacturers’ recommendations for
organizations have difficulty in identifying specific risks and applying optimal risk reduction
Moreover, even though the use of reliability engineering tools is well established, their
application to the medical industry is new. Most research in this area merely suggests how to
assess or improve the reliability of devices in their design or manufacturing stages. To this point,
best maintenance strategies for medical equipment in their operating context have not been
considered.
The research reported in this thesis aims to address this gap and proposes methods to
prioritization model which can be used to decide what medical devices should be included in the
1. INTRODUCTION 3
optimization models, we start with the actual maintenance data received from the Computerized
Maintenance Management System (CMMS) of a hospital. We learn about the nature of the data,
inspection and maintenance actions, and different types of failures. The results of our
observations lead us to find out what methods or models needed to be developed, and what
We establish policies to analyze the data statistically due to the presence of censored or
incomplete data, and propose a new method for trend analysis of such data. Then, we make some
assumptions according to the results of statistical analysis, and develop several inspection
The three main contributions of this research can be listed briefly as follows:
- A new model proposed to decide what equipment to include in the MEMP of a hospital.
- Models proposed for reliability and trend analysis of failure data in the presence of censoring.
- Several inspection optimization models developed for a repairable system under periodic
inspections.
devices, their typical life cycle, and inspection and maintenance management. In Chapter 3, we
develop a model for the prioritization of medical equipment for maintenance decisions. This
hospital according to their criticality. Devices with lower criticality scores can be given lower
priority in the MEMP. However, for devices with high scores, a root cause analysis should be
performed to find the main causes of having high criticality number. Appropriate actions such as
"preventive maintenance", "user training", "redesign the device", etc. should be taken to reduce
1. INTRODUCTION 4
their risk. Moreover, some guidelines for selecting appropriate maintenance strategies for
different classes of devices are proposed. The information on 26 different medical devices is
proposed model.
complex medical devices with censoring and missing information. We present a classification of
different types of failures and establish policies for analyzing data at system and component
levels taking into account the failure types. As a case study, we conduct the reliability analysis of
a general infusion pump. Moreover, in Chapter 4 a method for trend analysis of failure data
subject to censoring is developed. Trend analysis investigates any decreasing or increasing trend
in the historical failure data for a system or a group of similar systems. Most trend analysis
methods proposed in the literature assume that the failure data is statistically complete and
failure times are known; however, in many situations, such as hidden failures, failure times are
subject to censoring. We assume that the failure process follows a non-homogeneous Poisson
process with a power law intensity function, and the failure data are subject to left, interval and
right censoring. We use the likelihood ratio test to check for the trends in the failure data. More
specifically, we use the EM algorithm to find the parameters that maximize the data likelihood in
the case of no trend and trend assumptions. A recursive procedure is used to calculate the
expected values in the E-step. The proposed method is applied to several components of a
repairable system. We assume that the components of the system are subject to soft and hard
1. INTRODUCTION 5
failures. Hard failures are either self-announcing or the system stops when they take place; so
they are fixed instantaneously. Soft failures are unrevealed and can only be detected at
inspections; they do not stop the system functioning, although they reduce the system’s
performance.
First, several models are developed based on the simpler assumption that the system is
periodically inspected and the soft failures are rectified only at the periodic inspection moments.
We start with the assumption of minimal repair of components with soft failure, and construct
two models that respectively take into account the downtime of a component, and the number of
soft failures exceeding a pre-defined threshold. Then, the model with downtime consideration is
extended to two optimization models over finite and infinite time horizons, which assume
In the more complicated models, we assume that the soft failures can also be rectified at
non-scheduled inspections, which are performed due to the occurrence of hard failures. We call
the non-scheduled inspections “opportunistic inspections” for soft failures. We develop several
models based on the assumption of minimal repair or/and replacement of components with soft
and hard failures. Moreover, an optimization model is proposed for the case that components
with hard failure can be preventively replaced at periodic inspections according to some age
dependent probabilities. The main contribution of all these models is in calculating the excepted
The conclusions of the research and some guidelines for the future research are given in
Chapter 6.
2. INSPECTION AND MAINTENANCE OF MEDICAL DEVICES
Medical devices are fundamental components of modern health services used for
diagnosis, treatment, and monitoring of patients. They are progressively being deployed to
increase the capabilities of health diagnostic and treatment services. On the other hand, the
potential to manage and maintain medical equipment in most developing countries remain rather
weak (World Health Organization, 1998). It is required to have practical methods and powerful
management strategies to meet the challenges of ever increasing number and use of medical
devices.
Basic medical equipment is widely used in the healthcare facilities. This essential
equipment is supportive to provide primary healthcare to the public (Cheng, 2004a). World
Health Organization (WHO) classifies essential medical equipment in four main categories. The
list given in each category includes the devices required for a specified health service delivery.
The type of equipment is significantly dependent on the local health practice, physical
6
2.1. Essential Medical Equipment 7
Diagnostic imaging equipment is used to take pictures, which help physicians to diagnose
o Ultra-sound equipment
Ultra-sound equipment may be less frequently used, and is usually added if the budget
is available in a hospital.
2. Laboratory equipment
o Microscope
o Blood counter
o Analytical balance
o Colorimeter/spectrophotometer
o Centrifuge
o Water bath
o Incubator/oven
o Refrigerator
o Portable electrocardiograph
o External defibrillator
o Respirator
o Suction pump
o Diathermy unit
o Delivery table
o Electrical generator
o Refrigerator
Individual hospital authorities decide which type and what number of these devices are
safe and reliable operation of medical equipment and promote its effective utilization (Stiefel,
2009). This program defines procedures and policies to manage activities related to medical
equipment, from their selection and acquisition to decommission. MEMP ensures that devices
can provide reliable and accurate information to clinicians, operate safely for patients, and are
The life cycle of medical devices should be thoroughly considered for effective
management. Deficiencies in managing each stage of the life cycle, especially in the earlier
phases can cause more problems in the succeeding stages. For example, if the maintenance
capabilities are considered during acquisition stage, it can hinder the challenges that might be
A typical life cycle of medical equipment has the stages shown in Figure 2.1 (Cheng and
Dyro, 2004; also World Health Organization, 1998). Proper management of each phase can have
Planning
medical equipment are clearly outlined. This can significantly minimize the problems arising
from the contracts, spare parts and maintenance of the equipment (World Health Organization,
1998). For example, considering the skill level of operators, ensure that only appropriate
technology is acquired. Most costs incurred during the life cycle of a device are hidden from
2.2. Medical Equipment Management 10
view (Figure 2.2). For proper planning of a large number of devices, a full consideration of all
Planning
Acquisition
Installation, Commissioning
and Acceptance
Maintenance
Replacement or Disposal
Acquisition
Evaluation and procurement (Harding and Epstein, 2004a; 2004b) are two main aspects
of the acquisition phase. Evaluation process includes safety, performance and maintainability
assessment of devices. Moreover, the models and manufacturers of equipment are standardized.
In the procurement process, it is emphasized that the supplier must supply operating and service
manuals, and must provide operation and service training and essential spare parts.
Incoming devices should be checked carefully for possible damages in the shipment
process, conformity with the purchase order, and all required accessories, spares, and documents.
standardization (Barerich, 2004; Cohen and Cram, 2004). Inventory entries should include all
In-house technical staff or the suppliers can perform installation and commissioning
stage. In the latter, in-house staff should monitor the process and record it in the equipment
service history.
Proper training of users and operators assures effectiveness and safety of medical
In-house technical staff should act as a link between user and supplier and monitor the
Maintenance
When a medical device is old and its spares run out of supplies, it should be replaced and
The management of each stage mentioned above can be enriched if the resources are
added to enrich the maintenance element of the management plan (Cheng and Dyro, 2004).
engineers should comply continuously with two primary Joint Commission medical equipment
organizations to manage safety and security risks. Standard EC.02.04.03 presents guideline to
inspects, tests, and maintains medical equipment. The elements of performance for these two
equipment.
written inventory of selected equipment categorized by physical risk associated with use
and equipment incident history. The organization evaluates new types of equipment
before initial use to determine whether they should be included in the inventory.
3. The organization identifies the activities for maintaining, inspecting, and testing for all
strategies based on the type of equipment. Strategies must include defined intervals for
4. The organization identifies frequencies for inspecting, testing, and maintaining medical
5. The organization monitors and reports all incidents in which medical equipment is
suspected in or attributed to the death, serious injury, or serious illness of any individual,
6. The organization has written procedures to follow when medical equipment fails,
7. For organizations that provide the technical component of advanced diagnostic imaging
and elect to use The Joint Commission CMS imaging supplier accreditation option (Joint
and frequencies to maintain the reliability, clarity, and accuracy of the technical quality
Standard EC.02.04.03: The organization inspects, tests, and maintains medical equipment.
1. Before initial use of medical equipment on the medical equipment inventory, the
2. The organization inspects, tests, and maintains all life-support equipment. These
4. The organization conducts performance testing of and maintains all sterilizers. These
5. The organization performs equipment maintenance and chemical and biological testing of
Medical devices are often complex repairable systems consisting of a large number of
upon failure, can be restored to satisfactory performance by any method except replacement of
Medical devices usually undergo several types of tests/inspections during their life cycles
Acceptance Test
A series of qualitative and quantitative tasks designed to verify the safety and
and standards.
Operational Check
Visual and operational check of the equipment’s safety and functionality typically
performed at the beginning of the day or work period, or just before using equipment on a
patient.
2.3. Preventive Maintenance of Medical Equipment 16
A set of qualitative and quantitative tasks designed to verify the safety and performance
of each piece of equipment by detecting potential and hidden failures and taking appropriate
actions.
After accomplishing the acceptance test for a newly received device, SPIs are scheduled
to be performed periodically. If any problem is found at inspection, corrective actions are taken
to restore the device or its defective parts to an acceptable level. In addition, a set of failure
preventive actions may be taken to prevent future failures and/or restore device function; these
include part replacement, calibration, lubrication, etc. to address age or usage related
deterioration.
When a device fails while it is in use, the operator reports the problem, and again
appropriate actions (corrective maintenance) are taken. When the repair of a device is no longer
technically feasible or cost effective, replacement becomes the best or the only option (Atles,
2008). Figure 2.3 describes major tests and actions performed during a device’s life cycle.
Figure 2.3. Major tests and actions performed during a device’s life cycle
2.4. Computerized Maintenance Management Systems 17
an organization that assist in planning and management functions required for effective
maintenance (Gulati and Smith, 2009). A CMMS to a medical device is similar to an electronic
medical record to a patient. It provides the information required for assets management, resource
compliance (Cohen, 2008). A CMMS is essential in most healthcare organizations due to the
Joint Commission for the Accreditation of Healthcare Organization requirements under the
inventory of all equipment in the MEMP. The inventory should also include the devices that are
serviced by other organizations, but they still must be tracked. Each device must have an
equipment control number, which is labeled to the device. Equipment control number can be the
hospital asset or property number, or can be an independent number assigned to the device. The
inventory should be frequently updated as new devices are added. Without an accurate inventory
system, tracking equipment maintenance and repair, alerts and recalls is almost impossible
The core functionalities of a CMMS consist of the modules shown in Figure 2.4 (Cohen,
2008; also see Bagadia, 2006). More details of the CMMS modules are explained in Section
2.4.1.
2.4. Computerized Maintenance Management Systems 18
Inventory control
Inventory control is the core module of a CMMS. It allows an operator to track the
inventory movement, i.e. moving in and out of an item from the inventory or from a location to
another (Sapp, 2010). To proper planning of service and repair/replacement of each device,
healthcare organizations should know the quantity, type, age and the other information related to
the device.
Work order or work request is an electronic document used to schedule routine inspection
and maintenance. Work order management system is the heart of the CMMS. Using this module
2.4. Computerized Maintenance Management Systems 19
a work order can be created, followed up and completed. It stores all corrective and preventive
maintenance requests. It keeps track of the initial customer request, device information, requestor
information, date and time of the request, nature of the problem, its urgency, and a summary of
the assistance provided so far. All activities performed to a work order should be clearly
If the CMMS is equipped with an invoice system, a bill is issued and documented for the
work order. The work, which is performed, by vendors or external service providers should be
Scheduling/planning
specifies the most efficient way to perform the work, the schedules and the required resources.
Several scheduling procedures are usually considered in the CMMS. It includes periodic or fixed
scheduling, floating scheduling, and synchronized scheduling by feature such as device type,
location, specialty, and others (Cohen, 2008). Periodic scheduling is scheduling of an activity,
such as preventive maintenance periodically, regardless the time of the last action. Floating is
scheduling of an activity based on the last time the action has been completed for the device, and
the conditions of the device. Synchronized scheduling by feature allows to schedule actions
Vendor management
In some hospitals, the most sophisticated equipment such as MRI, scanner, analyzers,
etc., are maintained and services by the original equipment manufacturer (OEM). A good CMMS
should integrate the vendor and in-house work to have complete histories of all equipment.
2.4. Computerized Maintenance Management Systems 20
Vendor management system should also include the contracts and purchase management
subsystems to allow for recording the contracts with the external service providers and purchase
Parts management
Most organization uses a just-in-time process for ordering repair parts. Parts usually are
divided into three categories: stock parts, contract parts, and noncontracted parts (Cohen, 2008).
Stock parts are those, which are purchased and kept in stock by an organization to be used when
needed. They are not usually immediately assigned to a specific work order. Contract parts are
those included in a prepaid contract. Prepaid contracts are made with external service providers
for a certain period. Noncontracted parts are those, which are purchased usually just in time for a
Preventive maintenance
orders, prioritizes them based on some given criteria, and manipulates them until they are
accomplished.
Labor
Tracking labor resources can be performed using this module. It includes the information
Purchasing
Purchasing module is to initiate the requisition of parts and materials against a work order
and track the delivery and cost data of the item when the part or material arrives (Sapp, 2010).
Budgeting
2.4. Computerized Maintenance Management Systems 21
Budgeting module is integrated with the planned resources (labor hours, parts and
materials) usage on the work orders. It includes the labor, parts and materials rates to calculate
The maintenance and inspection data are usually available in the CMMS of a hospital,
stored in either scheduled or non-scheduled work orders. Scheduled work orders are used for
routine tests (SPIs); however, when a device fails or has a defective part, a non-scheduled work
order is requested to fix the problem. Both scheduled and non-scheduled work orders include the
basic information of a device and a test checklist designed for a particular class of device. The
checklist contains qualitative and quantitative tests; technicians or clinical engineers should use
this list to ensure that all necessary tests and checks are accomplished. Figure 2.5 shows a sample
device. For example, for a general infusion pump, these include testing its chassis/housing, line
cord, battery/charger, etc. Quantitative tests include measuring parameters of a device to check
whether the parameters are in control. Grounding resistance and maximum leakage currents are
A work order presents all PM checks and actions, such as cleaning, lubricating or
replacement of a device or its parts. A work order is also created for an acceptance test of a
Figure 2.5. A typical work order (designed for a general infusion pump)
“P” (pass), “F” (failed), or “N” (not applicable) are the possible results of each qualitative
or quantitative test on the work order. When a test is performed on a component, and it is found
to be non-defective, the result is “P”; however, in the case of failure of a part, the result is shown
as “F”. Since a general test checklist is designed for a class of equipment, some qualitative or
quantitative tests may not be applicable to particular devices in that class; for these devices, the
test result is given as “N”. Quantitative and qualitative tests are specifically performed for
components/features of a device, but PM checks show the actions performed at the device level.
periodic SPI of devices. SPI intervals differ from 6 to 12 months depending on the device type
and risk level (U.S. FDA/CDRH, 2005). Class III (high risk) devices such as defibrillators should
2.5. Concluding Remarks 23
be inspected every 6 months, and class II (medium risk) devices like ECGs should be inspected
annually. However, the optimality and even the necessity of these recommended intervals are
healthcare organizations must identify maintenance activities to maintain, inspect, and test all
medical equipment on the inventory. They must employ appropriate strategies, including the
intervals for maintenance activities. However, not much research has been presented in the
literature to address proper strategies and the methods for implementing them. In the research
reported in this dissertation, we propose several strategies, which can be used to meet some
devices in the MEMP, methods for trend analysis of maintenance data, and finally several
models to find the optimal periodic inspection interval for a repairable system such as a medical
device. The proposed models and methods are discussed in the rest of the thesis.
3. PRIORITIZATION OF MEDICAL DEVICES
The ever-increasing number and complexity of medical devices demands that hospitals
establish and regulate a Medical Equipment Management Program (MEMP) to ensure that
critical devices are safe and reliable and that they operate at the required level of performance.
As fundamental aspects of this program (Stiefel, 2009) inspection, preventive maintenance, and
engineering departments all around the world including Canada, Australia, and United States
have begun to employ more efficient and cost-effective maintenance strategies. Gentles et al
[http://CESOData.ca accessed 27 April 2010] have begun to develop a unique database to collect
comparative data on inventory and maintenance of the most critical devices used in hospitals
across Canada and the United States. This project will provide a large statistical failure data set
which could be used to establish optimum intervals for routine maintenance scheduling.
24
3.1. Literature Review 25
equipment and address methods, which have been used for a long time in other industry
for determining the maintenance requirement of a physical asset in its operating context through
a thorough and rigorous decision process, as shown in Figure 3.1 (Jardine and Tsang, 2006):
Steps 2 to 5 in Figure 3.1 show the process of Failure Mode and Effect Analysis (FMEA).
The results of FMEA are used to select appropriate maintenance tactics using RCM logic for the
various functional failures. Asset criticality analysis is the first step of applying RCM in an
organization, especially when a large number of different devices exist and the worst problems in
considered in a particular context. For example, the importance or criticality of a failure mode
depends on the combined influences of several factors such as severity, probability, detectability,
cost and timing, and all these factors play a part in determining the amount of attention that a
failure mode requires (JACAHO, 2005). Asset criticality is a function of the operational impact
to the organization’s mission due to the loss, damage, or destruction of an asset (Vellani, 2006).
Dekker et al (1998) define the equipment criticality as a function of the use of equipment, rather
than of equipment itself and explain how a certain device may be in one case critical and in
another auxiliary.
3.1. Literature Review 26
Significant and critical assets should be identified and prioritized, and many techniques
have been developed for criticality assessment of devices. Most use some variation of the
probability risk number or PRN (Moubray, 1997), a product of the probability of failure of an
asset, severity of the consequence of the failure, and detectability of the failure:
definition of risk differs from that used in RCM. After running an evaluation on medical devices,
clinical engineers decide which should be included in the MEMP of the hospital based on their
risk scores.
Fennigkoh and Smith (1989) proposed a risk assessment method to group medical
devices on the basis of their Equipment Management (EM) numbers, or the sum of the numbers
assigned to the device’s critical function, physical risk, and required maintenance:
Devices with an EM number above a critical value ( 12) are considered to have critical
risk and thus are included in inspection and maintenance plans. In 1989, the Joint Commission
(Fennigkoh and Smith, 1989) and eventually in 2004 approved it as the standard (EC6.10)
(JACAHO, 2004). This standard allows hospitals not to perform scheduled inspection or
maintenance tasks for certain pieces or types of medical equipment, if these tasks are not needed
for safe and reliable operation (Wang, 2006). Since then, Fennigkoh and Smith’s method or its
many variations have been used by clinical engineers (Rice, 2007). Ridgway (2009) in his recent
paper emphasizes that preventive maintenance can provide a benefit for just a relatively few
devices, and a significant number of repair calls are made due to random failures of device’s
3.2. Model for Prioritization of Medical Devices 27
components. Wang and Rice (2003) propose simplified version of gradient risk sampling and
Clinical engineers believe that risk is not the only inclusion criterion, however, even
though it is the most important one (Hyman, 2003). Other criteria which reflect the needs and
hazard notice, and recall history (Wang and Levenson, 2000; Ridgway, 2001). Moreover, current
maintenance strategies employed in hospitals have difficulty identifying specific risks and
medical devices and establish guidelines for selecting appropriate maintenance strategies.
(Yoon and Hwang, 1981). Analytical Hierarchy Process (AHP) (Saaty, 1980, 1990), a MADM
methodology used widely by practitioner and researchers (Leung and Cao, 2001), is a theory of
measurement through pairwise comparisons which relies on the experts judgments to obtain
priority scales (Saaty, 2008). AHP, briefly, is a three-step process. 1- It decomposes a complex
problem into a hierarchy, in which the overall decision objective lies at the top and the criteria,
sub-criteria and decision alternatives are on each descending level of the hierarchy (Partovi et al,
1989) compose of specific factors. 2- Decision makers then compare each factor to all other
3.2. Model for Prioritization of Medical Devices 28
factors at the same level of the hierarchy using a pairwise comparison matrix to find its weight or
relative importance. 3- The optimal solution is the alternative with the greatest cumulative
Two types of comparisons can be employed in the AHP: absolute and relative
measurements. Absolute measurement is applied to rank the alternatives in terms of the criteria
In absolute comparison, alternatives are compared with a standard in one’s memory that
has been developed through experience. In relative measurement, alternatives are compared in
alternatives does not reverse when new alternatives are introduced, or the old ones are deleted;
however, the priority of alternatives may change by altering the existing set of alternatives
AHP has been widely used in many applications involving decision making (Vaidya and
Kumar, 2006; Ho, 2008), and is often used for prioritizing alternatives when multiple criteria
must be considered (Modarres, 2006). Fong and Choi (2000) and Mahdi et al (2002) utilize AHP
for selecting contractors. Ramadhan et al (1999) use AHP to determine the rational weights of
pavement priority ranking factors, and Bevilacqua and Barglia (2000) utilize it for maintenance
strategy selection in an Italian Oil refinery. Labib et al (1998) propose a model to help take a
maintenance decision using AHP. Simpson and Cochran (1987) use AHP to prioritize
construction projects to assure that most needed projects receive funding when the budget is
limited. Al Harbi (2001) presents the AHP as a potential decision making method for use in
project management.
3.2. Model for Prioritization of Medical Devices 29
AHP to solve it. The objective is to identify and include the more critical devices in the
equipment management program of a hospital, and investigate in details the reasons of having
high criticality scores to take appropriate actions, such as “preventive maintenance”, “user
training”, “redesigning the device”, etc. when reducing the criticality score is applicable and
manageable.
The first step in applying AHP is to construct the hierarchy structure of the goal, namely,
prioritization of medical devices. All required criteria for assessment of devices must be
identified and placed at the appropriate level of the hierarchy (Saaty, 2008). Figure 3.2 shows a
Prioritization of
Goal
Medical Devices
c1 Criteria c2 c3 c4 c5 c6
Function Mission Age Risk Recalls and Maintenance
Criticality Hazard Alerts Requirements
The assessment criteria lie at the second level of the hierarchy structure. Relative
measurement method is used for pairwise comparison of the assessment criteria and for
determining their relative importance or weights with respect to the goal. In other words, the
weight of each criterion is determined by comparing its relative contribution to the goal
3.2. Model for Prioritization of Medical Devices 30
(prioritization of medical devices) with other assessment criteria. Therefore, if a new criterion is
added or an existing one is deleted from the hierarchy, all criteria should be reassessed to find
The alternatives or medical devices compose the third level of the hierarchy. The
objective is to assign a criticality score for every single device participating in the model.
However, the large number of alternatives (devices) makes their pairwise comparison with
respect to all criteria almost impossible. Moreover, medical devices are dynamic, i.e. devices are
added to or removed from the inventory over time, so to solve this problem, we suggest an
absolute measurement technique for ranking alternatives. By using absolute measurement, each
device is assessed with respect to each criterion and is given the most descriptive grade without
comparing it with other devices. Thus, our proposed model will use both relative and absolute
To be able to assess a device with respect to a criterion, the criterion’s grades and their
associated intensities should be defined in advance. The grades are possible categories or classes
of a criterion. For example, “old”, “average”, and “new” can be considered as three classes of a
device’s age. The definition of each class should be decided and concurred by the decision
makers. The decision makers may consider a device as new when its actual age is 25% of its
expected life span. The grades and their descriptions used in this thesis, are either obtained from
the available standards and literature, or proposed in this research and approved by the clinical
engineers.
Since the grades are subjective, each should be assigned an intensity value indicating its
score or importance with respect to the criterion. Quantifying the grades is a necessary step,
because when a device is assessed with respect to a criterion and is assigned the most descriptive
3.2. Model for Prioritization of Medical Devices 31
grade, then it is the assigned grade’s intensity, which participates in the prioritization model. In
order to determine more accurate intensity values for the grades, we propose to use the relative
measurement method to pairwise compare the grades with respect to their criterion. Employing
this method makes it possible to avoid assigning arbitrary intensities for the grades and having
After defining the grades and intensities for all criteria, the model is ready for use to
assess the devices. Each device is compared with respect to each criterion and is assigned the
Identify all sufficient, efficient and independent criteria and sub-criteria for criticality
assessment of devices.
Determine weighting values for all criteria and sub-criteria using relative measurement
method.
Set up grades and determine intensities for each criterion using relative measurement
method.
Evaluate alternatives (devices or failure modes) with respect to each criterion, and assign
the most descriptive grades using absolute measurement method; the assigned grade’s
n
CS i w j sij , (3.3)
j 1
3.2. Model for Prioritization of Medical Devices 32
maximum criteria, w j is the weight of the j th criterion, sij is the score of the i th device
n
with respect to the j th criterion, w
j 1
j 1.
In our proposed model (Figure 3.2), six criteria are identified at the top level. Some of
these should be divided into sub-criteria; we divide “Mission criticality” into “Utilization” and
“Availability of alternative devices”. Figures 3.3 and 3.4 show associated “Mission criticality”
c2 Mission Criticality
c21 c22
Utilization Availability of
Alternative Devices
nf
c4 Risk = Risk of failue mode k
k 1
Alternatives
Failure Mode 1 Failure Mode 2 Failure Mode 3 Failure Mode nf
The criteria suggested in this chapter include some proposed criteria in the literature for
MEMP inclusion of medical devices. For example, “Function”, “Physical Risk”, and
“Maintenance requirements” are suggested by Fennigkoh and Smith (1989). To assess the failure
consequences of a device’ failure modes we include in our model “Physical Risk” as “Safety and
environment” criterion. Wang and Levenson (2000) suggest replacing “Function” by “Mission
Critical” in the Fennigkoh and Smith’s model, and also to take into consideration the
C1. Function
The function of a device is the main purpose for which it is to be used. The Medical
Devices Bureau of Health Canada recognizes four classes of medical devices based on how the
device is represented for use by the manufacturer (Health Canada, 1998). Class I devices present
the lowest potential risk and Class IV present the highest. This classification implicitly represents
the function of a device. For example, a life support device such as a defibrillator is considered
as a class IV device with high risk of failure (death of a patient) if the device fails. However, this
classification does not explicitly describe the function of a device. Moreover, risk or
consequence of a device failure should not be confused with its function, thus we propose in our
function categories. The proposed categories are an adaptation of Fennigkoh and Smith (1989)
Mission criticality or operational impact describes the extent to which a device is crucial
to the care delivery process of a hospital (Wang, 2006). For example, Magnetic Resonance
Imaging (MRI) equipment might be extremely significant according to the mission of a hospital
but less critical in terms of its function or potential risk through use. Wang (Atles, 2008) suggests
classification of devices in three groups (Critical, Important, and Necessary) according to their
mission criticality. In our model, mission criticality depends on utilization and availability of
C21. Utilization
Utilization shows the total hours a device is used on average in a hospital (hours per day
or days per week or weeks per year). In this model, we consider the “average hours a device is
used per week” as the utilization criterion. “Utilization” can be defined as a function of average
usage and the number of patients served per unit time. Obviously, incorporating the “Number of
Although not always true, with decreased backup and fewer similar devices at hand, the
situation of a device in the care delivery process becomes more critical, especially when it is in
high demand and there is only one such device in the hospital.
It should be noted that having several similar devices does not always mean high
availability of alternative devices: the demand per unit time of these devices is also important. If
there are several similar devices in a hospital but all are highly utilized, if either fails, there is
less chance that others can be used as substitutes. Therefore, “Availability of alternative devices”
3.2. Model for Prioritization of Medical Devices 35
can be considered as a function of the number of similar or backup devices and their demand per
unit time.
C3. Age
Age score is based on the actual age of a device and its predictable life span. The life
span for a group of similar devices can be obtained from the literature. In general, ten years is the
C4. Risk
Risk is one of the most important criteria in criticality assessment of a device but cannot
be simply considered as a single number assigned to a device. Rather, the risk of a device should
be an aggregate of all risk values estimated from the actual failures, which have occurred in the
device. All failure modes and their associated frequencies, consequences, and detectabilities
should be extracted or estimated from data history and device maintenance work orders. The risk
value can then be estimated as a function of frequency, consequence, and detectability for each
failure mode. In short, the risk of the device is the total risk of all its failure modes.
Failure frequency is the likelihood of a failure occurrence. The VA National Center for
Patient Safety has designed “Healthcare Failure Modes and Effects Analysis (HFMEA)”
specifically for healthcare (DeRosier et al, 2002). Their suggested “Failure Frequency” rating
C42. Detectability
3.2. Model for Prioritization of Medical Devices 36
Failure detectability is the ability to detect a failure when it occurs. In the proposed
To find the total consequences of each failure mode, its operational, non-operational, and
safety and environment impacts should be assessed. These three categories are the conventional
failures consequences used in the RCM terminology (Moubray, 1997), so we keep this level
although both operational and non-operational criteria have only one related sub-criteria in our
model, as follows.
C431. Operational
hospital.
C4311. Downtime
Downtime is generally the average time that a device is out of service. However, for
medical devices it is also important to consider the number of patients who have to wait while
the device is down. We therefore suggest the average of the total waiting time of all patients as
“Downtime” sub-criteria, considering both the number of patients and the length of time they
C432. Non-operational
Non-operational consequence of a failure is its direct inspection and repair cost, including
Cost of repair is the sum of labor and spare part(s) costs incurred by fixing a failure or
defect. Grade classification (i.e. high, medium, and low) for cost of repair depends on the budget
Safety and environment consequences of a failure are critical impacts that should be
taken into account. These are described in Table A.9 (Fennigkoh and Smith, 1989).
The number and class of recalls and the number of hazard alerts that may occur for a
device are important criteria in prioritization of medical devices. U.S. Food and Drug
Administration (FDA) guidelines categorize recalls into three classes according to the level of
pneumatic, or fluidic often requires the most extensive maintenance. A device is considered to
have an average maintenance requirement if it requires only performance verification and safety
testing. Equipment that receives only visual inspection, a basic performance check, and safety
Once all criteria and sub-criteria have been identified, their relative importance can be
determined with respect to their goal or their upper-lever criterion using Saaty’s eigenvector
3.2. Model for Prioritization of Medical Devices 38
technique (Saaty, 1980, 1990, and 2008). In other words, relative AHP is employed to determine
criteria and sub-criteria weighting values. Table 3.1 shows the weighting values calculated for all
criteria and sub-criteria in the model. The values given here represent expert opinion; results may
Sub-Criteria Sub-Criteria
Main Criteria (Weight) Sub-Criteria (Weight)
(Weight) (Weight)
c1-Function (0.45)
c21-Utilization
(0.70)
C2-Mission criticality (0.10) c22 -Availability of
alternatives devices
(0.30)
C3-Age (0.06)
c41-Failure frequency
(0.30)
c42-Detectability
(0.24)
c431-Operational c4311- Downtime
(0.16) (1.00)
C4-Total risk (0.16)
c432-Non-
c4321-Cost of repair
c43-Failure operational
(1.00)
consequence (0.46) (0.08)
c433-Safety and
environment
(0.76)
C5-Recalls and hazard alerts
(0.16)
C6-Maintenance requirement
(0.07)
We suggest using the absolute measurement technique for ranking of medical devices due
to their large number and dynamic nature. To employ absolute measurement, qualitative
(descriptive) grades for each criterion are constructed. Then, to find grade intensities, the grades
are pairwise compared according to their corresponding criterion. If a grade with highest
intensity is assigned to a device with respect to a criterion, the criterion for this device should
contribute with its full capacity (criteria’s weight) to its upper-level criterion or goal. It means
that this intensity should have value of 1. Therefore, the intensities should be divided by the
maximum intensity that can be obtained for each criterion. Finally, using absolute AHP, each
alternative is evaluated with respect to each criterion and is assigned an appropriate describing
The grades’ intensities for the assessment criteria proposed by current classification
models in the literature can be criticized, especially with respect to inconsistency in logical
reasoning. For example, in the model proposed by Fennigkoh and Smith (1989), “Death” and
“No significant risk” have score values of 5 and 2 as the major and minor failure consequences,
respectively; however, this does not appropriately reveal their severities ratio. Therefore, in our
model, we propose using AHP to find the intensities of the criteria’s grades.
In the AHP hierarchy, the qualitative grades and intensities should be determined for
criteria/sub-criteria upon which the alternatives are directly evaluated (criteria with no sub-
criteria). Appendix A describes the qualitative grades and intensities of criteria/sub-criteria taken
from the literature or designed for this study. To demonstrate how the intensities can be
calculated for a criterion, the required steps are explained for the criterion “Function” in
Appendix A.
3.2. Model for Prioritization of Medical Devices 40
Our model is now ready to rank medical devices. Each device should be assessed with
respect to every covering criterion, the lowest level criterion or sub-criterion connected to the
alternatives (Saaty, 2008) and assigned an appropriate grade. The score of a device at a criterion
that has sub-criteria is then the sum product of the sub-criteria’s weights and their grades’
intensities assigned to the device. This part of the AHP process is called “synthesizing” (Saaty,
2008). Therefore, the total score for a device can be obtained as an absolute value from the
weighted sum of the main criteria and their assigned intensities for the device. In order to easily
prioritize and classify devices according to their score values, the absolute total score values
should be normalized by dividing the total score values by the maximum of all devices’ score
values.
To estimate a device’s “Risk”, its failure modes should first be assessed with respect to
the risk criterion. In other words, alternatives in “Risk” hierarchy (Figure 3.4) are the failure
modes of a device. The “Risk” for a device is the sum of risk values of all its failure modes, that
is
nf
RF RFk
k 1
(3.4)
The risk of a failure mode can be calculated using weighted sum of risk criteria, that is
where RFk is the risk of failure mode k of the device, k 1,2,..., nf (the total number of
devices’ failure modes). Gk FF , Gk FD , and Gk FC are the intensities assigned to the failure mode
k of the device with respect to “Failure Frequency”, “Failure Detectability”, and “Failure
3.2. Model for Prioritization of Medical Devices 41
Consequence”. c41 , c42 , and c43 are the weighting values of Gk FF , Gk FD , and Gk FC ,
respectively. The weighted sum allows us to take into account the magnitude of risk factors.
The grades’ descriptions presented in Tables A.7 to A.11 are used to assign appropriate
grades to a failure mode. For example, if a particular failure mode of a device occurs once in 2 to
5 years, “uncommon” grade is assigned to its failure frequency. The failure history of all possible
failure modes of a device should be extracted from the computerized maintenance management
system (CMMS) and analyzed to select appropriate grades. A module can be implemented in the
CMMS to automatically extract and analyze the history of a failure mode. Tables A.7 to A.11
are either adapted from the FMEA proposed for medical devices (DeRosier et al, 2002), or are
In order to compare the risk values of all devices, we need an upper bound for the risk
criterion. After calculating the risk values for all devices, we divide them by the maximum
existing risk value, so the device with the highest risk value will have a risk ratio of 1.
Obviously, when the number of device is just one, the obtained risk value will be 1, but this
Instead of the weighted sum, heuristic reasoning can be used to deduct the grade of a
device/failure mode with respect to a criterion from its sub-criteria’s grades assigned to that
reasoning such as fuzzy logic can be employed (Rzevsky, 1995). In this case, the rules of
inference should be established before assessing the alternatives. Rules are knowledge
expressions in the form of conditions and actions. A rule consists of an IF-statement and a
satisfied, the THEN-statement can consequently be concluded. For example, the consequence of
a failure mode can be deducted from the grades of its sub-criteria as follows:
In general, rules represent the experts’ knowledge, and a rule-based expert system
(Durkin, 1994) is a system in which the knowledge is stored in the condition-action rules.
In our model the criteria may be also considered as either static or dynamic. For example,
function is a static feature of a device, which does not change over time, while other criteria such
as age or risk are dynamic. Therefore, the prioritization model should be applied to devices
periodically to adjust their criticality scores according to changes in the score values of the
dynamic criteria. For example, a recently purchased device may not have enough failures
reported in the CMMS, but gradually some failure records will be added for it to the system.
New failures contribute to more precise estimation of the total risk value of the device and they
The CMMS of the pilot hospital has the information of about 11,365 non-imaging and
2,241 imaging physical units (devices) and all of them should be prioritized. Thus, presenting
our complete results would require much more space, which is not possible in this chapter. We
therefore present a simplified example to illustrate the model’s application in the prioritization of
CMMS. The selected devices are similar to the devices listed and discussed in (Fennigkoh and
Smith, 1989) and (Wang and Levenson, 2000). We used the assessment given in these papers for
the selected devices with respect to some of our proposed criteria such as “Function”, “Mission”,
We extracted multiple failure modes for each device selected for the study, and assessed
them with respect to the risk criteria. Tables 3.2 (a, b)-3.5 demonstrate the approach. As shown
in Tables 3.2 (a, b), Infant Incubator, Defibrillator, and Intra_aortic Balloon Pump have the
highest normalized criticality score of 1.000, 0.964, and 0.943, respectively. For these devices,
the main contribution to the critically score comes from their high score values with respect to
“Function”, “Mission”, and “Risk”. Infant Incubator is also assigned a “Low” grade for
“Recalls” which is an important criterion with a weight of 0.16 in the prioritization hierarchy
structure. Surgical and Exam Lights have the lowest criticality scores, so they can be excluded
from the maintenance management program of the hospital. Additional comments on the results
11 Patient Diagnostic
Automatic X-Ray Processor 1.000 Old (1.00) 0.269
(0.16)
12 Patient Diagnostic
Ultrasound Machine 0.802 Average (0.43) 0.576
(0.16)
13 Mobile Hypo/Hyperthermia
Therapeutic (0.21) 0.340 Average (0.43) 0.718
Unit
14 Almost New
Infusion Pump (A) Therapeutic (0.21) 0.760 0.886
(0.17)
15 Multi-channel Patient Diagnostic Almost New
0.538 0.695
Electrocardiograph (0.16) (0.17)
16 Patient Diagnostic
Fetal Monitor 0.340 Old (1.00) 0.610
(0.16)
17 Almost Old
Centrifuge Analytical (0.13) 0.340 0.598
(0.67)
18 Almost New
Cardiac Cath Harp Therapeutic (0.21) 0.538 0.432
(0.17)
19 Scale for patient care Miscellaneous (0.11) 0.165 Old (1.00) 0.817
20 Patient Diagnostic
Blood Pressure Modules 0.340 Old (1.00) 0.328
(0.16)
21 Almost Old
Computer Terminal Analytical (0.13) 0.298 0.737
(0.67)
22 Water Bath Circulator Analytical (0.13) 0.207 Old (1.00) 0.524
23 Sterilizer Miscellaneous (0.11) 0.760 New (0.12) 0.399
24 Patient Diagnostic Almost New
Ultrasound Doppler 0.298 0.792
(0.16) (0.17)
25 Almost Old
Surgical Lights Therapeutic (0.21) 0.340 0.432
(0.67)
26 Exam lights Miscellaneous (0.11) 0.207 Average (0.43) 0.677
Table 3.2a. Assessment of the devices with respect to “Function”, “Mission”, “Age” and “Risk”
Alternative
No Device Name Utilization Mission
Devices
1 Automatic X-Ray Processor High (1.00) Low (1.00) 1
2 Blood Pressure Modules Medium (0.34) Medium (0.34) 0.34
3 Cardiac Cath Harp Medium (0.34) Low (1.00) 0.538
4 Centrifuge Medium (0.34) Medium (0.34) 0.34
5 Computer Terminal Medium (0.34) High (0.20) 0.298
6 CT Scanner High (1.00) Medium (0.34) 0.802
7 Defibrillator High (1.00) Low (1.00) 1
ECG Physiological Telemetry
8 High (1.00) Low (1.00) 1
Unit
9 Exam Lights Low (0.15) Medium (0.34) 0.207
10 External Pacemaker High (1.00) Medium (0.34) 0.802
3.2. Model for Prioritization of Medical Devices 46
Table 3.3. Assessment of the devices with respect to “Mission criticality” sub-criteria
Device risk =
Detectability
Consequenc
Frequency
mode Risk
Total risk
risk/max
Failure
Total
Device Name score
e
environment
Consequence
(Downtime)
Operational
Safety and
Centrifuge
- Broken speed control knob Medium Medium Inappropriate therapy 0.2172
- Centrifuge intermittently drops in speed while running Medium Medium Inappropriate therapy 0.2172
- Machine turns on but motor does not run Medium Medium Delayed Treatment 0.1640
Computer Terminal
- Mouse locked up Low Low No consequence 0.1044
- Overflow errors caused data to be missed in analysis Medium High Inappropriate therapy 0.2796
- Tape speed errors Medium Medium Delayed Treatment 0.1640
CT Scanner
- CT cable loose Low Low No consequence 0.1044
- System could not recognize a scan as raw data High Medium Delayed Treatment 0.2840
- Wire harness for CT dislodged Medium Low Delayed Treatment 0.1600
Defibrillator
- Cable constructed Low Low Delayed Treatment 0.1424
- Cable tie on paddle handle Medium Medium Delayed Treatment 0.1640
- Intermittent break in power cord Medium Medium Death 0.8176
ECG Physiological Telemetry Unit
- Broken terminal connection Medium Medium Delayed Treatment 0.1640
- Soot water damage from fire High Medium Delayed Treatment 0.2840
- Telemetry does not detect lead off Medium High Inappropriate therapy 0.2796
Exam lights
- Bulb broken inside socket Low Low No consequence 0.1044
- Lamp holder broken Low Low No consequence 0.1044
- Loosed swivel head Medium Low Inappropriate therapy 0.2132
External Pacemaker
- Broken clip Low Low Delayed Treatment 0.1424
- Pace lights visible only when pacemaker is lying down Medium Low Injury 0.3120
- Pulse generator failure High High Death 1.0000
Fetal Monitor
- Monitor not powering-up Medium Medium Delayed Treatment 0.1640
- Paper was not moving Low Low No consequence 0.1044
- The unit kept losing its configuration High Medium Inappropriate therapy 0.3372
Gamma Camera 0.2796
- Fatal timeout problems Medium High Inappropriate therapy 0.3372
3.2. Model for Prioritization of Medical Devices 51
Surgical Lights
- Broken cover cat Medium Medium Inappropriate therapy 0.2172
- Light drifting Low Low Inappropriate therapy 0.1956
Treadmill
- CRT display tube goes blank Medium Low Delayed Treatment 0.1600
- Does not indicate speed Medium Medium Inappropriate therapy 0.2172
- Poor contact in ECG lead connector block Low Low Delayed Treatment 0.1424
Ultrasound Doppler
- Damaged LCD leads Medium High Delayed Treatment 0.2264
- Loose battery contact Medium Medium Delayed Treatment 0.1640
- Not picking up the signal High Medium Inappropriate therapy 0.3372
Ultrasound Machine
- Cable holder fell off Low Low No consequence 0.1044
- Diagnostic software not working Medium Low Delayed Treatment 0.1600
- system locks up High High Delayed Treatment 0.3464
Ventilator
- Damaged power supply Medium Low Delayed Treatment 0.1600
- Going to standby during use High High Injury 0.4984
Water Bath Circulator
- Pump motor bearing failed High High Delayed Treatment 0.3464
- Relay no longer available Medium Medium Inappropriate therapy 0.2172
- Water bath not regulating temperature Medium Medium Injury 0.3160
The proposed model in this chapter incorporates all criteria suggested by clinical
engineers (Fennigkoh and Smith, 1989; Wang and Levenson, 2000; Ridgway, 2001; Hyman,
2003; Wang, 2006) to assess the criticality of medical equipment. Moreover, the model gives a
realistic estimate of the total risk of a device by taking into account its different failure modes
and assessing their frequency, detectability and consequences. The failure modes are extracted
from the device’s failure history available in the CMMS of the hospital and are analyzed with
respect to the risk’s sub-criteria. AHP enables the model to accommodate multiple criteria and
3.2. Model for Prioritization of Medical Devices 53
integrates scientific judgments with personal opinion in the evaluation of the alternatives (Herath
and Prato, 2006). Calculating the consistency ratio (Saaty, 1990) in pairwise comparison of the
criteria makes the model able to produce more precise and consistent criteria’s weights compared
to direct assignment of the weights. Furthermore, the intensities, which are obtained for the
criteria’s grades, are also more consistent due to applying the relative measurement method. One
of the limitations of the proposed model is that it requires experts to be involved in the process of
applying the model to medical devices and this process may be expert intensive. Moreover, the
prioritization results may not always be accepted by clinical engineers and requires reassigning
the criteria’s weights and/or grades’ intensities. To apply the proposed model to the CMMS of a
hospital and obtain accurate results, the CMMS should be up-to-date and should include the
As has been noted, the proposed model prioritizes devices according to their criticality.
The normalized score value indicates the relative criticality of a device compared to other
devices. Given such a model, hospitals could focus their maintenance efforts on more critical
devices. Devices with lower criticality scores could be discarded from a maintenance
management program, while devices with high score values could be monitored and supervised.
Normalized scores (e.g. the values in the last column of Table 3.2b) can be used for
prioritizing or ranking of devices. The normalized scores depend on the total number of devices
involved in the model. However, total scores of devices (e.g. the values in the 5th column of
Table 3.2b) can be used as absolute measurements for classification. The total score is a metric
that can be compared with predefined thresholds to decide to which category the device belongs.
3.2. Model for Prioritization of Medical Devices 54
In our proposed model, devices can have a total score between (0.1050, 1.0). Score 1.0 is for a
device which gets the highest intensity when assessed against every single criterion, and 0.1050
is obtained when the device gets the lowest intensity from all criteria (see Appendix B for this
calculation). Therefore, the score in the proposed model is always between 0.1050 and 1. The
total score can then be mapped to (0, 100%) using the following equation:
The thresholds in Table 3.6 are based on the score values obtained for the devices given
in the numerical example. In general, the thresholds can be adjusted after applying the model to
the inventory of a hospital and investigating the obtained transformed score values. The
thresholds can be adjusted, depending on the minimum and maximum of these values. Different
type of devices with different transformed score values should participate in the model to decide
how the budget should be allocated for management of devices and which maintenance strategies
should be applied. Participation of all devices is normally expected when the prioritization
module is integrated into the CMMS of a hospital. Otherwise, the output decisions might be
inaccurate. For example, if the prioritization model is just applied to a group of devices in a
3.2. Model for Prioritization of Medical Devices 55
hospital and not all devices, the managers may allocate the budget only to this group and decide
to apply predictive maintenance for all of them, which is not necessarily required.
of participant devices and their estimated transformed score values. Moreover, the number of
classes and established maintenance strategies depend on available resources (budget, personnel,
(cost, labor, and special equipment) and their impact on maintaining a piece of equipment
(Mobley, 2002). In corrective maintenance, a device is just run until it breaks. This strategy does
not require any staff or money to be applied; however, unplanned downtime and inefficient use
of the staff to repair the device can even be more costly. In time-based maintenance, the device is
periodically checked and preventive maintenance is performed if necessary. This strategy is easy
to implement and can reduce the failure rate of a device. However, it requires some resources
such as budget and labor to be implemented. Time-based maintenance may not always be
optimal and a device may be over or under maintained when the preventive maintenance is
carried out on a calendar-based scheme. In predictive maintenance, some variables such as the
device vibrations are measured to assess the device’s physical status. The measured variables are
used to predict failure of the device. This maintenance strategy requires some sensors or special
equipment to measure the prediction variables. However, it can increase the operational life of
Devices with low critically score can be excluded from the maintenance management
program of a hospital and just get fixed whenever they break down (corrective maintenance).
This group of medical devices more likely have a low score value with respect to the criteria
3.2. Model for Prioritization of Medical Devices 56
possessing higher weights such as function, risk, and recalls and hazard alerts; so the asset
management resources are better not to be spent on this group. Table 3.2b shows that corrective
maintenance can be used for the last eight devices on the list starting with Scale for patient case
and ending with Surgical and Exam lights. In other words, these devices are run to failure and
Proactive maintenance could potentially be applied to high critical devices. This type of
maintenance employs monitoring and correction of failing root causes (Swanson, 2001) or in
general, detection and elimination of root causes of such a high criticality score when it is
applicable and manageable. If physical conditions such as vibration, noise, etc. of a device with
applied here as well. Time-based maintenance can be considered for devices with average
with no need to have any special sensor equipment. In Table 3.2b, devices such as Infant
Incubator, Defibrillator, and Balloon Pump with transformed score of greater than 40% belong to
respect to their dynamic criteria, and their current maintenance policies should be adjusted
accordingly. For example, when the total risk score of a device is high, its root causes should be
investigated and appropriate actions should be taken to reduce the risk. If risk is high due to the
high frequency of failures, the root causes might be “use error” or imperfect design. In this case,
user training or redesigning the device if possible, could be effective risk reduction actions.
3.2. Model for Prioritization of Medical Devices 57
We also can use the score value obtained for a device with respect to each individual
criterion to determine appropriate maintenance strategies for the device. Here, the decision factor
is the score value of a device with respect to a criterion or a combination of criteria (antecedent
strategy. This multi-criteria methodology introduced here is a method that requires more research
and will not be further discussed, but some initial heuristic guidelines for selecting appropriate
maintenance strategies are suggested in Table 3.7. For example, Table 3.7 recommends “design-
out maintenance” when a device has a high score value for either “recalls” or “risk”, since the
device more likely has an engineering design problem. When the score value of the device for
both “recalls” and “risk” is low, and the device is function-critical with significant failure
Antecedent Consequent
(Recalls ↑) or (Risk ↑) Design-out maintenance
(Utilization ↑) Condition monitoring
(Function ↑) (Utilization ↓) Inspections
(Recalls ↓) and (Risk ↓) and Failure-finding interval
(Detectability ↑)
(Failure consequence ↑)
(Function ↓) (Mission ↓) Operate to failure
(Utilization ↑) Calendar-based maintenance
(Function ↕)
(Utilization ↓) Utilization-based maintenance
Symbols: ↑ high/critical ↓ low/non-critical ↕ medium/mid-critical
according to their criticality. The model uses AHP to identify and include the more critical
devices in the equipment management program of a hospital. The proposed hierarchy structure
contains six criteria to assess criticality of the devices, “Function”, “Mission criticality”, “Age”,
“Risk”, “Recalls and hazard alerts”, and “Maintenance requirements”. The model gives a
realistic estimate of the total risk of a device by taking into account its different failure modes
and assessing their frequency, detectability and consequences. The proposed model uses both
relative and absolute measurement in the application of AHP to determine weighting values for
the criteria and their grades’ intensities, and to evaluate the alternatives (devices). Calculating the
consistency ratio in pairwise comparison of the criteria makes the model able to produce more
precise and consistent criteria’s weights compared to direct assignment of the weights.
Furthermore, the intensities that are obtained for the criteria’s grades are also more consistent
This model can be integrated as a module into the CMMS of a hospital to prioritize
medical devices. Devices with lower criticality scores can be assigned a lower priority in a
maintenance management program. However, those with higher scores should be investigated in
detail to find the reasons of having such high criticality scores to take appropriate actions, such
as “preventive maintenance”, “user training”, “redesigning the device”, etc. when reducing the
criticality score is applicable and manageable. In addition, individual score values for each
device with respect to a criterion or a combination of several criteria can be used to establish
In our proposed model, the pairwise comparison matrices are constructed using Saaty’s
crisp 1-9 scales. Alternatively, fuzzy analytic hierarchy process (Kwong and Bai, 2002) can be
used; in this case, scales are fuzzy numbers instead of crisp values to handle uncertainty in
A common belief of a healthcare society is that medical devices fail independent of age,
following exponential distribution. This belief originates from the general conviction that
electronic equipment has a constant failure rate. Even the most widely known and used
proposes reliability models whose construction is based on the constant failure rate assumption.
This assumption has been criticized as inaccurate, however, and its use may result in erroneous
decisions (Choi and Seong, 2008; Puran, 1990; Mortin et al., 1995). Reliability and failure
environmental stress, expertise level of operators, etc. Therefore, to realistically determine their
reliability, devices should be studied in their operating context. To this end, learning about
existing maintenance procedures of equipment and statistical analysis of field data are essential
Trend analysis is a common statistical method used to investigate the operation and
changes in a repairable system over time. This method takes historical failure data of a system or
60
4.1. Literature Review 61
a group of similar systems and determines whether the recurrent failures exhibit an increasing or
decreasing trend. Both graphical methods and trend tests are used for trend analysis. The latter
are statistical tests for the null hypothesis that the failure process follows a homogeneous Poisson
Crow/AMSAA test (Crow, 1974; 1982) assumes that the failure process of a repairable
system has a Weibull intensity function and finds the maximum likelihood estimates of the
parameters. The shape parameter is then used as an indicator for a growth or deterioration in the
system reliability. The Laplace and the Military Handbook tests (Ascher and Feingold, 1984) are
used to test whether data follow a HPP. Kvaløy and Lindqvist (1998) propose the Anderson-
Darling trend test for a NHPP (Ross, 2007) with a bathtub shaped intensity function based on the
Anderson-Darling statistic test (Anderson and Darling, 1952). The Lewis-Robinson (Lewis and
Robinson, 1974) is a modification of the Laplace test and is used as a general test to detect trend
departures from a general renewal process (Lawless and Thiagarajah, 1996). The Mann test
(Mann, 1945) corresponds to a renewal process null hypothesis and the monotonic trend as the
alternative. Kvaløy et al. (2002) declare that the Mann test is a more powerful test against
decreasing trend. The MIL-HDBK-189 (1981; Crowder et al., 1991) is used for testing a NHPP
Most trend tests (Ascher and Feingold, 1984; Wang and Coit, 2005) assume that failure
times are known, so the failure data are complete. Currently in the literature, except for right
Poisson process (NHPP) which incorporates left and interval censored failure data if repairs are
not instantaneous or not performed immediately. However, it is expected that the data received
from industry include missing and incomplete information. Sometimes a particular type of data
4.1. Literature Review 62
of interest is not measured at all, or if measured may be incomplete or unreliable. For example,
hidden failures, which make up to 40% of all failure modes of a complex industrial system
(Moubray, 1997), are not evident to operators and remain dormant until they are rectified at
scheduled inspections. In this case, the times of hidden failures are either left or interval
censored, and the censoring interval is the interval between two consecutive inspections.
Weibull or power law (Crow, 1974) and log linear (Cox and Lewis, 1996) intensity
functions are two common models used to describe recurrent even data underlying NHPP. The
maximum likelihood estimates of the parameters are obtained for two intensity parameters
systems at the design and development stage over the last two decades (Fries, 2005; Dhillon,
2000), less literature deals with the reliability prediction and trend analysis of medical devices
while they are in use in hospitals. Ion et al. (2006) analyze the field data for medical imaging
systems during the warranty period. Roelfsema (2004) presents the results of early reliability
prediction of Philips medical systems based on field data. Baker (2001) analyses a large database
for failures of many types of repairable medical devices including ventilators, infusion pumps,
and ECG monitors. She assesses the validity of some well-known failure rate models such as
power law and log linear and linear Poisson processes. The results reveal an increasing rate of
failures with equipment age. Moreover, the power law process model is found to fit better to the
data. In addition, she considers the imperfect repair models, and finds out that for some
equipment, such as ventilator and ECG monitors, the repair effect models are better fits.
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 63
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data
In this study, we analyzed maintenance data (collected from 2000 to 2007) from a
Canadian General hospital. We conducted statistical analysis to learn about the failure types and
trends of a particular general infusion pump used to deliver liquids for therapeutic and/or
diagnostic purposes. However, since all medical devices undergo periodical safety and
performance inspections, the general policies derived from this study can be applied to other
medical devices. Although dealing with a specific pump, this chapter describes how available
For the purposes of this study, we selected a general infusion pump. General-purpose
infusion pumps (Figure 4.1) are used to accurately deliver liquids through intravenous (IV) or
epidural routes for therapeutic and/or diagnostic purposes (ECRI Institute, 2009). The reason for
The pilot hospital has 681 pieces of this model used in three departments. All scheduled
and non-scheduled work orders were extracted from the CMMS, and we performed a visual
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 64
examination of the data looking for possible errors. A typical work order is shown in Figure 2.4
of Chapter 2.
Device Contaminated 1 0
Component Failure 910 1182
User Input/Setup Error 3 1
Problem Not Found 275 316
Not Applicable 0 2
Device over used 1 0
NULL (the cause was not specified) 6 18
Total 1196 1519
Table 4.1 shows the total number of scheduled and non-scheduled work orders and the
associated findings at inspection. Minor defects are not usually reported by the operators and are
rectified only at SPIs. Moreover, clinical engineers perform opportunistic maintenance of some
components like batteries, which may be replaced before complete depletion. Therefore, as Table
4.1 makes clear, more failed components are identified in scheduled work orders than in non-
scheduled work orders. The average time between non-scheduled work orders is 452.8 days, with
a standard deviation of 423.1 days. However, 661.6 days is the average time between scheduled
work orders with relatively small variation (a standard deviation of 175.3 days).
We counted the percentage of times that a test resulted in “P” (pass), “F” (failed) and
“N/A” (not applicable) for different components/features of the devices (see Section 2.4.2 in
Chapter 2). Tables 4.2-4.4 summarize the results. Components/features with zero or few failures
The Pareto analysis of the qualitative test results of non-scheduled work orders identified
“Fittings/Connectors” are the most problematic components/features with the total number of
P F N/A P F N/A
Component Component
(%) (%) (%) (%) (%) (%)
AC Plug/Receptacles 98.92 1.08 0.00 AC Plug/Receptacles 90.75 0.54 8.71
Alarms 73.29 26.71 0.00 Alarms 76.20 15.08 8.71
Audible Signals 91.93 8.07 0.00 Audible Signals 75.12 16.17 8.71
Battery/Charger 97.92 2.08 0.00 Battery/Charger 79.39 11.90 8.71
Chassis/Housing 58.24 41.76 0.00 Chassis/Housing 37.26 54.03 8.71
Controls/Switches 91.26 8.74 0.00 Controls/Switches 86.12 5.17 8.71
Fittings/Connectors 80.03 19.97 0.00 Fittings/Connectors 64.48 26.80 8.71
Indicators/Displays 91.01 8.99 0.00 Indicators/Displays 83.11 8.17 8.71
Labeling 90.68 9.32 0.00 Labeling 63.88 27.34 8.77
Line Cord 88.77 11.23 0.00 Line Cord 67.13 24.16 8.71
Mount 83.94 15.89 0.17 Mount 79.27 12.02 8.71
Circuit Breaker/Fuse 3.99 1.08 94.93 Circuit Breaker/Fuse 1.98 0.18 97.84
Strain Reliefs 98.34 1.50 0.17 Strain Reliefs 88.52 2.46 9.01
Table 4.2. Qualitative tests results extracted from scheduled and non-scheduled work orders
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 66
P F N/A P F N/A
Component Component
(%) (%) (%) (%) (%) (%)
Flow Rate Accuracy Flow Rate Accuracy
12.90 0.00 87.10 90.50 0.00 9.50
(%) (%)
Grounding Resistance Grounding Resistance
99.75 0.17 0.08 91.17 0.12 8.71
(mohm) (mohm)
Maximum Leakage Maximum Leakage
100.00 0.00 0.00 91.29 0.00 8.71
Currents Currents
Occlusion Alarm 99.08 0.75 0.17 Occlusion Alarm 90.56 0.72 8.71
Table 4.3. Quantitative tests results extracted from scheduled and non-scheduled work orders
PM P F N/A PM P N/A
F (%)
Check (%) (%) (%) Check (%) (%)
Clean 99.75 0.25 0.00 Clean 91.29 0.00 8.71
Lubricate 99.67 0.33 0.00 Lubricate 91.11 0.18 8.71
Replace 29.37 70.63 0.00 Replace 13.04 78.19 8.77
Table 4.4. PM checks results extracted from scheduled and non-scheduled work orders
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 67
600 100
90
500
80
70
400
60
300 50
40
200
30
20
100
10
0 0
er
nt
g
d
fs
s
g
e
rs
s
or
lin
m
in
us
he
al
ay
cle
lie
ou
rg
to
us
gn
ar
be
/F
it c
ha
pl
Re
ec
ta
M
Al
ne
Ho
Si
s
er
La
Sw
ep
/C
nn
Di
in
ak
Li
e
s/
ry
ec
s/
Co
ra
ls/
bl
si
re
or
tte
St
/R
di
ro
as
s/
tB
at
Au
Ba
ug
nt
ng
Ch
ui
Co
di
Pl
tti
rc
In
Fi
AC
Ci
Figure 4.2. Pareto analysis for qualitative tests results
Failures can be classified into two broad categories: soft and hard failures. Soft failures
are the gradual loss of performance of a product, while hard failures cause it to stop working
(Meeker and Escobar, 1998). However, in this chapter, “hard” failures signify those failures
where the user is notified as soon as they occur. Hard failures either directly influence
problem occurs, users are notified that something is wrong with the device, and it is sent for
repair. We can assume that a hard failure occurs just before a SPI or corrective maintenance;
therefore, the work order’s starting time can be considered as the failure time (complete data).
Arguably, soft failures are identified at SPI when operators have postponed reporting them.
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 68
“Soft” failures are minor defects that have no or little influence on the device’s
functioning. These defects such as “chassis/housing” defects may occur a long time before a
scheduled or non-scheduled work order’s starting time, but the device still can function properly.
The problem is not reported by the user and is detected only at the next SPI or during corrective
maintenance triggered by a hard failure. Therefore, there is a time delay between the real
occurrence of a soft failure and its detection. Ignoring this time is misleading; it should be
While we can interpret soft failures as the result of system deterioration over time, hard
Interval censoring takes place when a minor failure or defect happens between a “no
failure found” inspection and a “failure detected” inspection. If a defect occurs between the
acceptance date and the first corrective maintenance or SPI, interval censoring is reduced to left
Although there is not always a clear distinction between soft and hard failures, we used
two groups depending on whether their associated failures were soft or hard failures and
analyzed the failure data separately for each group. We classified components/features according
to their failure percentage at scheduled and non-scheduled work orders. Classification of soft and
hard failures can be found in Appendix C. This classification has been approved from a practical
point of view by clinical engineers involved in this study. In addition, it should be noted that
Failure Mode and Effects Analysis (FMEA) could be also used to decide which components
should be included in each category of components with soft and hard failures.
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 69
In order to check the validity of our treatment with soft and hard failures, we calculated
the number of hard and soft failures detected at each single scheduled and non-scheduled work
order. Table 4.5 shows the results. The significant percentage of single hard failure detected in
non-scheduled work orders emphasizes that it is correct to assume all hard failures’ times as
complete data. The close values obtained for one and two detected soft failures at both scheduled
and non-scheduled work orders show that we can treat soft failures as left or interval censored
data.
1 79.72 83.72
2 18.88 14.88
Hard failures
3 1.22 1.24
4 0.17 0.16
Table 4.5. Soft and hard failure frequencies at scheduled and non-scheduled work orders
Because of the large number of censored data and combination of hard and soft failures,
we propose separate policies for analyzing failure data at different levels, including system level,
System level
All non-scheduled work orders’ starting times should be considered as a system’s exact
failure times. The time between SPIs having at least one failure identified should be considered
as either left or interval censored. The right censored data is specified according to the result of
the last inspection (failure or pass), and its distance to the end of the test. Figures 4.3a and 4.3b
illustrate the proposed policy, and the policy rules are explained in detail in Appendix D.
t1 t2 t3 t4 t5 Non-scheduled Scheduled
work order time work order time
TStart x1 x2 x3 x4 x5 x6 T with Failed status with Failed status
End
t1 t2 t3 t4 Non-scheduled Scheduled
work order time work order time
TStart x1 x2 x3 x4 x5 TEnd with Failed status with Failed status
Figure 4.3a. The policy for analyzing failures at the system level
TStart x TEnd
Event Times WO Type Data Type
Figure 4.3b. The policy for dealing with “no failure” devices at the system level
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 71
When the time between the last inspection or acceptance test (when there is no failure)
and the end of the test is long, particularly for soft failures, there are two possibilities: 1. a soft
failure may have occurred somewhere between the last inspection and the end of the test; 2. no
failure has occurred in the meantime. Since there is no information available regarding these two
possible scenarios, and as it is expected that a failure occurs sometime between the average (or
proportion of) times between failures and the end of the test, we consider a proportion of that
time as the right censoring time. In order to be consistent, we consider (0.5*mean time between
failures) as right censored data. We conducted different analyses with different values (0.5, 0.75
and 1) as the ratio to decide which proportion of mean time between failures should be
considered as the right censoring rule. Even though we may have several failures in that interval,
For the components corresponding to hard failures, work orders’ starting times are
considered as a component’s exact failure times if they report a failure of that component. The
time interval between the last failure of the component and the end of the test is considered as
Non-scheduled Scheduled
work order time work order time
TStart x1 x2 x3 x4 x5 TEnd
Figure 4.4. The policy for dealing with hard failures – component level
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 72
For soft failures, all rules are similar to the system level’s rules, except that we never
have complete data. If there is no inspection with “no problem found” between two consecutive
failures, the time interval between these two failures is considered as left censored data.
The “Laplace trend test”, the “Reverse Arrangement test” or the “Military Handbook
test” (NIST/SEMATECH, 2006) can be used to quantitatively determine whether the failure
The Laplace trend test can be used for multiple repairable systems to check between no
trend and non-homogeneous Poisson process (NHPP) exponential models. Assume that there are
m similar and independent systems. Let Tij represent the time of the i th failure of the j th system,
The trend test can then be conducted using the following formula:
m nj m
1
T 2n (T
j 1 i 1
ij
j 1
j j start T j end )
LA . (4.1)
m
1
n j (T j end T j start ) 2
12 j 1
The value of LA should be compared to high (for improvement) or low (for degradation)
percentiles of the standard normal distribution. The Laplace trend test compares the average of
failure arrival times to the midpoint of the observed time interval. There is a trend when the
average of the failure arrival times deviates from the midpoint of the observation interval. In the
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 73
Laplace trend test, when nj 3, the approximation with the Normal distribution generally
there is a trend, with direction of the trend (Ansell and Phillips, 1994) indicated by the following:
We analyze the failures data based on the proposed policy described in Section 4.2.3. We
use the Laplace trend test for trend analysis of failure data, taking the midpoint of censoring
intervals as an approximation of the actual failure times. Even though this is an approximation
and therefore not precise, we opt to use it because no available trend test analysis in the literature
incorporates interval censored data (except right censored). Trend analysis using the midpoints is
simple and does not require any complex calculation. Moreover, when the censoring intervals are
short, reliable results can be obtained. In Section 4.4 we propose using the likelihood ratio test to
check for trends in the failure data with censoring. We use the EM algorithm to find the
parameters, which maximize the data likelihood in the case of no trend and trend assumptions. A
recursive procedure is proposed to calculate the expected values in the E-step of EM.
The mean cumulative function (MCF) and confidence limits are plotted versus system
age in days using the RELIABILITY procedure in SAS. The MCF plot describes average
failure event by considering the number of systems at risk at that point in time. The number of
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 74
systems at risk depends on how many systems are contributing information at that particular
point in time. Information can be obscured by the presence of censoring and truncation (Trindade
and Nathan, 2008). The RELIABILITY procedure in SAS provides a nonparametric estimate and
plot of the MCF for the number or cost of repairs for a population of repairable systems.
The nonparametric estimate of the MCF, the variance of the MCF estimate, and
confidence limits for the MCF estimate are based on Nelson (Nelson, 1998). MCF, usually
denoted by M (t ) , is defined as the mean of the cumulative number of events up to time t . The
method does not assume any underlying structure for the repair process (Johnston,
As shown in Figure 4.5, the MCF plot at the system level is not a straight line. Its initial
concave appearance is followed by a stable period, and a convex curve appears at the end.
Therefore, the assumption of HPP for failure data is not valid. The end-of-history or right-
7. 5
S
a 5. 0
m
p
l
e
M
C
F 2. 5
- 2. 5
days
Figure 4.5. The MCF plot for the failures at the system level (0.5*MTBF)
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 75
Using equation 4.1, the Laplace trend test value for the failure data is LA=12.59. At the
Since the failure times are not independently and identically distributed, we first classify
the times between two consecutive events (n-1th, nth) of each system. An event can be either a
failure or a censoring event. The nth event of a system is censoring if at inspection no failure is
found; otherwise, it is a failure event. We then pool the times between n-1th and nth events of all
systems and conduct the Weibull analysis for each category separately. For example, for
calculating the times to the first event, all devices that have at least one failure during their life
are identified. For these devices, the time to the first failure is extracted with “failure” as an
event. For devices with no failure at all, the first and only event is a censoring event with the
censoring time from 0 to the last updating time of the CMMS database. Finally, the Weibull
analysis is performed for the times between n-1th and nth events using LIFEEG procedure in SAS.
Tables 4.6 and 4.7 summarize the Weibull parameters estimated for all times between events.
The results of the Weibull distribution fitted to the times between n-1th and nth event
(failure/censoring)
No of No of No of No of
No of
n non- right left interval 1
( 1)
observations
censored censors censors censors
1 674 237 28 348 61 1.5680 396.9852 356.6204
2 646 215 111 283 37 1.5175 318.9879 287.5696
3 536 177 100 222 37 1.3499 306.6525 281.2007
4 433 143 116 162 12 1.5622 311.9259 280.3152
5 318 101 115 96 6 1.3042 301.7416 278.5013
6 203 48 114 40 1 1.1024 381.6272 367.9733
7,8 129 34 58 35 2 0.9718 247.9506 251.0791
9 89 41 30 15 3 0.8472 156.9852 171.1464
Table 4.6. Estimated parameters of Weibull distribution for times between events – system level
4.2. A Model Proposed for Reliability Analysis of Medical Devices’ Failure Data 76
1
n ( 1)
95% confidence 95% confidence 95% confidence
estimate estimate estimate
limits limits limits
1 1.5680 1.4414 1.7058 396.9852 372.1665 423.4590 356.6204 333.3963 379.8445
2 1.5175 1.3867 1.6606 318.9879 297.9330 341.5308 287.5696 267.7068 307.4324
3 1.3499 1.2165 1.4868 306.6525 282.5753 332.7812 281.2007 257.6725 304.7289
4 1.5622 1.3977 1.7461 311.9259 288.1422 337.6727 280.3152 257.8563 302.7741
5 1.3042 1.1365 1.4967 301.7416 269.8376 337.4177 278.5013 246.4507 310.5519
6 1.1024 0.8865 1.3708 381.6272 307.6500 473.3927 367.9733 284.4419 451.5047
7,8 0.9718 0.7618 1.2395 247.9506 192.9140 318.6887 251.0791 182.2192 319.9390
9 0.8472 0.6710 1.0695 156.9852 115.1459 214.0272 171.1464 112.4674 229.8254
Table 4.7. Estimated 95% confidence limits for the Weibull distribution parameters
As can be seen in the tables, earlier failures show a degrading trend; however, later
failures appear more random, i.e. follow exponential distribution. The values are slightly
decreasing for the first five events, which have more non-censored data; thus, they constitute a
more reliable result. The results are also consistent with the result of the Laplace trend test.
Since SAS only gives the 95% confidence limits estimation for the estimated parameters
( and ), we calculated the limits for using the Delta method (Meeker and Escobar, 1998).
For details, see Appendix E. The estimated lower and upper limits for using the Delta method
Analysis of all hard failures, soft failures, and the failures of individual components
Appendix F describes the results obtained from analysis of all hard and soft failures
separately, and some of the system’s components individually. An overview of the analysis
Level *LA
System level 12.59 (↓) ↓ ↓ ↓
All hard failures 6.12 (↓) ↓ ↓ ↓
AC Plug/Receptacles 0.87 (--) -- -- --
Alarms 4.13 (↓) ↓ ↓ ↓
Circuit Breaker/Fuse 2.39 (↓) N/A N/A N/A
Controls/Switches 1.15 (--) -- -- --
Indicators/Displays -2.72 (↑) ↓ ↑ ↑
Mount 9.31 (↓) ↑ ↓ ↓
Occlusion Alarm -1.11 (--) -- -- --
All soft failures 11.08 (↓) ↕ ↕ ↕
Audible Signals -7.94 (↑) ↓ ↑ ↑
Battery/Charger 5.69 (↓) ↓ ↑ ↑
Chassis/Housing 11.40 (↓) ↕ ↓ ↓
Fittings/Connectors 8.15 (↓) ↕ ↓ ↓
Labeling 4.13 (↓) N/A N/A N/A
Strain Reliefs 3.03 (↓) N/A N/A N/A
Line Cord 10.61 (↓) ↕ ↕ ↕
Grounding Resistance -2.10 (↑) N/A N/A N/A
Symbols: ↑ improving/increasing ↓ degrading/decreasing ↕ no obvious trend -- no trend
*LA (the Laplace test value) N/A (not applicable)
Table 4.8. An overview of the results obtained from statistical analysis at different levels
The total number of failures and their completeness or censoring status determine the
accuracy of the results. With more complete failure data, the results are better and more
consistent. For example, due to the large number of exact hard failure data, the results for this
group exhibit a degrading trend consistent with the results of the Laplace trend test. However, in
the case of only a few failures, nothing can be learned because of the large number of censored
data. “N/A” and “↕” results in Table 4.8 stem from the scarceness of failure data. Nor were the
4.3. Basic Assumptions for the System Under Study 78
exact soft failure times available; thus, all data for this group were either left or interval
censored, and consequently, the results showed a large variation. Overall, the Laplace trend test
appears to yield results that are more reliable: analyzing the data marginally may show large
variations due to the censoring and scarceness of failure data; and when the results are
The results described in Section 4.2 show that components of a general infusion pump
(Figure 4.1) or any other medical device or more general complex repairable system could be
categorized into two groups. Failures of some components such as indicators, switches, and
occlusions make the system stop functioning as soon as they occur, and they are fixed
immediately. Then, these failure times are known. These components have so-called hard
failures. The second group consists of components with hidden or soft failures such as circuit
breakers or informative components such as audible signals. The pump can continue to operate if
one of these components, as for example, an audible signal, fails. Therefore, soft failures are
usually not self-announcing and are rectified only at inspections, and their failure times are either
left or interval censored. In other words, there is a time delay between the real occurrence of a
soft failure and its detection. Although the system still may be able to continue working, but the
presence of soft failures can reduce its performance and eventually require to be fixed.
Once the component experiences a soft failure it stays in the same condition until it is
fixed. Even if the component may deteriorate in some way after the failure the repair at
inspection will return it to the state just before failure (minimal repair), or will renew it.
4.3. Basic Assumptions for the System Under Study 79
Soft and hard failures can be also interpreted by their consequence on the system. Hard
failures have more significant influence on the system operation. Soft failures are less critical for
the system. Meeker and Escobar (1998, p. 327) also consider two types of failures: hard and soft.
They suggest that soft failures may be defined as when degradation (gradual loss of
performance) exceeds certain specified level, while hard failures cause the system to stop
working.
There are different categories of soft failure of components in complex devices. One
category includes a wide range of integrated protective components used to protect a system
from unwanted transient incidents such as current and voltage surges. For example, infusion
pumps and medical ultrasounds are equipped with circuit breakers to protect them against
overload and short circuits. If these protective components fail, the system can continue its main
function, although the risk of damage increases if an overload or short circuit takes place. The
failures of protective components are not self-announcing and can only be rectified at inspection.
This group of components does not age further when a failure occurs and until the
component is fixed and put back in service. Standby redundant components are another category
of components with soft failures. They are used in many critical systems to enhance system
safety and reliability. The system, depending on its design, can continue functioning if a required
number of these redundant components work. Uninterruptible Power Supplies (UPS), dual
redundant processors, and redundant fan trays are examples of redundant components used to
increase system safety and availability. A CAT scan (CT) may have several CPUs running
different algorithms for image processing. If one CPU fails, processing is routed to other CPUs
without interrupting or degrading the system’s performance. Again, components in this category
do not age after failure, and usually their failures are detected only at inspection.
4.3. Basic Assumptions for the System Under Study 80
In addition to protective and redundant components, there are other components whose
failure does not halt the system, even though failure can have serious consequences, even
catastrophic, if left unattended. With infusion pumps, for example, audible or visual signals are
used to communicate with operators and inform them of the status of the patient to whom the
device is attached. When the level of liquid delivered to a patient reduces to a certain level, the
component responsible for producing signals starts to send a warning alarm. If the component
fails, the pump can still function, but the patient’s health risk increases if the operator does not
take action. Devices in overcrowded hospitals are always in use, and even though some of their
informative features such as audible signals may fail, they remain in operation until the next
scheduled inspection.
The failures of components that do not carry out the main functions of a system can
sometimes be hidden. For example if an external hard drive backup of a computer fails, the
system can still function, so the operator may not be informed about its failure and the chance of
The followings are the basic assumptions, which we make for the system under study in
this and all future chapters. Additional assumptions if required will be made and stated at each
Basic Assumptions
At the beginning, a component may have any initial age, which can be zero if it starts as
good as new.
Failures of the units follow a NHPP with a power law intensity function.
Components are inspected periodically at times k (k 1,2,...,n) (Figure 4.6), over a given
Inspections are perfect and inspection and repair times are negligible.
HF HF HF
2 (k 1) k
n T
S S S S
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring
We consider a system with the assumptions stated in Section 4.3. Since soft failures are
subject to censoring, in this section we propose a method for trend analysis of this type of failure.
We assume that the failure process follows an NHPP with a power law intensity function. We
use the EM algorithm to estimate parameters of the power law process and propose a recursive
method to calculate the expected values in the EM expectation step. In the maximization step, we
use the Newton-Raphson method. We then apply the general likelihood ratio test to test HPP
against NHPP alternative (Crowder et al., 1991). In some practical situations, the midpoints of
the censoring intervals may be considered as the failure times and used to estimate the
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 82
parameters of the failure process and to perform trend test analysis. If the inspection intervals are
Additional Assumptions
1
x
It should be noted that a more common form to present a power law is ( x)
x x
or ( x) (t )dt ; however we use ( x) e x 1 for convenience of maximum
0
The censoring intervals are the intervals between two consecutive inspections (Figure 4.7).
X1 X2 Xr
Our first goal is to estimate parameters of the NHPP from censored data on
X 1 , X 2 ,... using the maximum likelihood method and to test for possible trend in the data using
the likelihood ratio test. Because the random variables X 1 , X 2 ,... are dependent, our problem is
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 83
complex. Note that in our approach, inspection times need not be periodic; nonequal inspection
Our main objective is to develop a method to perform a trend analysis of censored failure
We use the likelihood ratio test (Meeker and Escobar, 1998; Engelhardt et al., 1990),
Let L0 and L1 be the maximum likelihoods of the data when 1 , and without
restrictions, respectively. The likelihood ratio test in our case uses the statistic 2 2 ln( L0 / L1 )
with 1 degree of freedom and rejects the null hypothesis if 2 is greater than the appropriate
The likelihood function incorporates failure events and a possible last censoring event
u1 u2 ur ur 1
l1 l2 lr lr 1
x0 y0 T
Figure 4.8. Lower and upper bounds for failure times and the last censoring time
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 84
In Figure 4.8, li is the last known (inspection) time when the unit was still working before
the i th failure, and u i is the time that the unit was found failed, where i 1, 2,..., r , r 0 .
When (r 1)th event is a right censoring event (not a failure), lr 1 is the right censoring time, and
ur 1 . The origin for measuring li 1 and ui 1 is the previous failure detection time ( u i ).
Note that this case may be considered as a Type I censoring, with T being a calendar time
interval, but not the operating time interval, as common in other applications.
Let
i
yi x j , y0 x0 .
j 0
where x j are the actual survival times measured from u j 1 . Therefore, y i is the unobserved
actual age of the unit at the time of u i and yi 1 li yi yi 1 ui . In the case with right censoring,
H 1 r+1 { yi 1 li yi yi 1 ui , i 1,..., r 1} .
Let Ai { yi ( yi 1 li , yi 1 ui ]} . Thus,
H 1 r+1 A1 A2 ... Ar 1 .
Let ( , ) . Let n r 1 in the case with right censoring, and let n r in the case
without right censoring. The common case when a failure is observed as soon as it occurs and the
system is repaired instantly (that is, when all inspection intervals are of length zero, Crowder et
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 85
al., 1991)) reduces to our case when n r 1 . If all y i were known, the complete data likelihood
would be
n
L( y | ) fY ( y, ) fY i ( yi | yi 1 )
i 1
n n
(4.2)
( yi )e [ ( yi ) ( yi 1 )]
( ( yi ))e ( ( yn ) ( y0 ))
,
i 1 i 1
where ( yi ) ( yi , ) e yi .
n yn
ln L ln fY ( y, ) ( yi ) ( s)ds
i 1
n
y0 (4.3)
n n ln ( 1) ln( yi ) e ( yn y0 )
i 1
ln L n e ( yn y0 ) 0
(4.4)
ln L n ln( yi ) e ( yn ln yn y0 ln y0 ) 0
n
i 1
L(H 1 n | ) P( An , An 1 ,..., A1 | y0 , )
to find the maximum likelihood estimates (MLE) of the parameters ( , ) given the
observed data. The EM finds the MLE iteratively using the two following steps:
Expectation Step: In this step, the expected value of the log-likelihood function is calculated:
Q( , (t ) ) E ( t ) [ln fY (Y , ) | H1 n ]
Maximization Step: In this step, (t 1) is found as to maximize the expected log-likelihood:
n
Q( , ( t ) | H 1 n , y0 ) E ( t ) [ ln (Yi ) ( (Yn ) ( y0 )) | H 1 n , y0 ]
i 1
n
n n ln ( 1) E ( t ) [ ln Yi |H 1 n , y0 ] e ( E ( t ) [Yn | H 1 n , y0 ] y0 ) (4.6)
i 1
For simplicity, we will assume that y0 0 , i.e., the initial age of a unit is zero in the rest
of the section.
Q
n e E ( t ) [Yn | H 1 ]
n
, (4.7)
Q n E ( t ) [ ln Yi | H 1 n ] e E ( t ) [Yn ln Yn | H 1 n ]
n
i 1
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 87
2Q
2 e E ( t ) [Yn | H 1 ]
n
2Q 2Q
e E ( t ) [Yn ln Yn | H 1 n ] . (4.8)
2Q n
2 2 e E ( t ) [Yn (ln Yn ) 2 | H 1 n ]
gradient algorithm (Lange, 1995) which solves the M-step of the algorithm using one iteration of
2Q( , (t ) ) Q( , (t ) )
(t 1) ( t ) ( ) , (4.9)
2 ( t )
2Q ( , ( t ) )
where is the Hessian matrix of Q( , (t ) ) .
2
The parameters in the complete EM algorithm are obtained by applying the Newton-
Raphson method to find (t 1) . The (t 1) is the limit of the sequence { (t ,k ) } where
k 0
2Q( , (t ) ) Q( , ( t ) )
(t ,k 1) ( t ,k ) ( ) . (4.10)
2 ( t ,k )
n
E ( t ) [ ln Yi | H 1 n ] , E ( t ) [Yn | H1 n ] , E ( t ) [Yn ln Yn | H1 n ] , E ( t ) [Yn (ln Yn )2 | H1 n ] .
i 1
y
... n fY ( y1 ,..., yn )dy1...dyn
yi 1 li yi yi 1 ui
i 1,..., n
E ( t ) [Yn | H 1 n ] .
...
yi 1 li yi yi 1 ui fY ( y1 ,..., yn )dy1...dyn
i 1,..., n
Therefore, we will derive a general recursive procedure to calculate the required expected
values.
g ( yn ) yn ln yn . Then,
ui
Bi 1 ( yi 1 ) Bi ( yi 1 x) f X i ( x | yi 1 )dx , (4.11)
li
yi 1 x
i 1 i 1 ( s ) ds
where l
j 1
j yi u j and f X i ( x | yi 1 ) ( yi 1 x)e
j 1
yi 1
, and i n, n 1,...,1 .
Then,
For a proof, see Appendix G. A similar recursive equation can be derived to calculate the
n
expected value of ln Y , as follows:
i 1
i
Let Dn ( yn ) 0 , then
ui
i n, n 1,...,1 . Then
n
E ( t ) [ ln Yi | H 1 n ] D0 (0) / C0 (0) .
i 1
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 89
To obtain L0 , or the maximum likelihood of the data when 1 , first note that because
L0 (H 1 n | ) P(l1 X 1 u1 , l2 X 2 u2 ,..., ln X n un | )
n
(e e
li
e e ui
),
i 1
and
E ( t ) [Yn | H 1 n ] E ( t ) [ X 1 ... X n | H 1 n ]
n n
E[ X i | H 1 n ] E[ X i | li X i ui ]
i 1 i 1
ui
xe
e e xi dx
n
li
i 1 (e e li
e e ui )
n
(li e e li
ui e e ui ) e (e e li
e e ui )
i 1 (e e li
e e ui )
n
li e e li
ui e e ui
e li e ui
n e .
i 1 e e
In the case of right censoring, i.e. when n r 1 and ur 1 , the last element in the
sum is just lr 1 .
n
(t 1) ln( ). (4.13)
E ( t ) [Yn | H 1 n ]
When there is more than one unit, equation (4.6) should be changed appropriately.
Assume there are m similar independent units with n j histories ( j 1,..., m ), and that each unit
may have failures and a right censoring event. Then, equation (4.6) is revised as follows:
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 90
m m m nj
n j ln n j ( 1){E ( t ) [ ln Y ji | H 1 j , y j 0 ]}
n
j 1 j 1 j 1 i 1
m
e {E ( t ) [Y jn | H 1 j , y j 0 ] y j 0 }
n
(4.14)
j 1
In Section 4.2.5, for trend analysis of components of an infusion pump, we considered the
mid points of the censoring intervals as the failure times and applied the Laplace trend test.
We apply the method proposed in this section to subsets of failure data for several
components of the general infusion pump with hidden failures and compare the results with those
obtained by using the “mid-point” method. To obtain the maximum likelihood estimates of the
parameters we use both the EM algorithm (equation (4.10)) and the modification of the EM
(equation (4.9)) with one Newton-Raphson iteration, and compare their results and the
algorithms’ performances.
The first component considered is an audible signal. We use the failure data of 80
randomly selected units out of 674 units. The dataset contains 125 records or histories collected
over intervals of between one and seven years with inspection intervals of approximately 1.6
years. Each unit has up to 3 histories including failure censoring intervals and the last right
censoring interval. Table 4.9 shows sample data for three units. The third and fourth columns
show the lower and upper bounds of the censoring interval for a failure. The symbol in the
1 0.000 1.300
1 2 2.886 3.597
3 1.408
2 1 1.289
1 1.906 3.419
3
2 2.886
Table 4.9. Sample data for three units’ failure and censoring events
In total, we found 41 units with right censoring only; a summary of the data appears in
Table 4.10.
Total No of units 80
(0) ( (0) , (0) ) (0.3, 0.5) . We obtained functions Bi , C i and Di for i 0,1,..., n j
recursively, using equations (4.11) and (4.12), for every j , j 1,...,80 . We applied the
Composite Simpson’s rule for numerical integration, with step length h depending on the
required accuracy; after some experimentation, we used h 103 . We then used equation (4.14)
and the first and second derivatives and applied equations (4.9) and (4.10) to obtain
likelihood ratio test statistic 2 2 ln( L0 / L1 ) 0.128 , which is less than the critical value
1,0.05
2
3.841. The results are summarized in Table 4.11. The parameter e / is also
calculated for comparison. Accordingly, we do not reject the null hypothesis and conclude there
is no visible trend in the pattern of given failures. However, this may also be due to the small
number of histories per unit (up to three), not enough to show the trend clearly. Applying the
complete EM resulted in an estimate ˆ (ˆ , ˆ ) (1.672, 0.920) after 24 iteration for t and 52
iteration in total for both t and k . The parameters obtained are very close to the parameters
obtained by applying the modified EM. The required process time for the complete EM
algorithm was about 2 times longer than the time of the modified EM. The same conclusion
about trend as above is made when the parameters obtained from the complete EM are used in
ˆ (ˆ , ˆ ) (1.858,1.050) , slightly different from the above results using the complete EM and
its modified version. Although in this example, there is no significant difference between the
results, a larger dataset and more failure histories may yield different results, and the results
obtained when considering mid points might be misleading. On the other hand, using mid-points
is much simpler (as it does not need numerical integration), and the obtained values may be used
as initial values in the EM algorithm. This may greatly reduce the number of iterations.
Figure 4.9 shows an infusion pump with a cracked door hinge. A summary of the number of
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 93
failures for the units are given in Table 4.12. See Appendix P for complete dataset. All units have
a right censoring record. We applied the proposed method with the same initial parameters
(0) ( (0) , (0) ) (0.3, 0.5) . The results are summarized in the third column of Table 4.11.
The chi-square test for this case study also shows no trend in the failures. However, the value
obtained from the mid-points is slightly greater than 1 that would indicates the units are
degrading over time. Considering that the likelihood ratio test is more accurate, this would be
As the last case study, we consider the battery of a general infusion pump. An infusion
pump can be powered either from an AC line or from a rechargeable or disposable battery. The
infusion pump in our case study is equipped with a rechargeable battery. Battery depletion,
battery overcharged, battery low voltage, are common failure modes of a battery. Some battery
failures, which damage the battery, require it to be replaced. For example, Figure 4.10 shows a
damaged sealed lead acid battery, which is damaged by battery overcharging. A depleted battery
The battery dataset consists of 897 records describing failures of 674 batteries. Table 4.13
gives a summary of the number of failures. All units have a right censoring record. Due to the
The results of applying the proposed model for batteries are shown in the last column of
Table 4.11. High value of 108.835 for the Chi-square test indicates a low p-value. According to
the results, there is an obvious increasing trend in the failure of batteries. In other words,
We compared the proposed method of using likelihood test with the standard errors of the
estimation (Meeker and Escobar, 1998). We calculated the standard errors of the parameter
estimators using the modified EM from the observed information matrix. Calculation of the
observed information matrix requires additional effort and more complex numerical integration
and will not be discussed here. The standard errors of the parameter estimators for all three
matrix of the parameters and is obtained for the first case study (Audible Signals):
4.4. A Proposed Method for Trend Analysis of Failure Data Subject to Censoring 96
1
2 ln L 2 ln L
Var
ˆ (ˆ ) ˆ (ˆ , ˆ ) 2
Cov
0.0622 0.0281
2
0.0198
.
ˆ (ˆ , ˆ )
Cov Var ( ) ln L
ˆ ˆ ln L 0.0281
2
ˆ , ˆ
2
The estimated standard errors of ̂ and ˆ are the square roots of the diagonal elements
Standard Error
Case Study
̂ ˆ
Audible Signal 0.2494 0.1407
Housing/Chassis 0.2133 0.1176
Battery 0.1387 0.0851
The standard errors obtained are in accordance with the likelihood ratio test. For
well above 1. Or, for audible signals, the 95% confidence interval is [0.645,1.197], covering
1.
The total numbers of iterations required to obtain the parameters in the complete EM, the
modified EM and Mid-Point method are given in Table 4.15 for all three cases. In the complete
EM we have iterations over both t and k . In the modified EM, iterations are over t only. In the
midpoints method they are also over t (the number of iterations required for Newton-Raphson
method until convergence). In all methods, we considered a difference less than 0.001 as the
The total number of iterations in the complete EM is 2~2.5 times larger than in the
modified EM, and the required times for executing these two methods follow about the same
proportion.
The analysis of data from complex medical devices is not simple, largely because of the
amount of censored and missing information. In this chapter, we propose a solution to this
problem. Even with scarce failure data and a large number of censoring events, we are able to
conclude that the reliability of systems degrade over time. This result runs counter to the
common belief that failures of electronic equipment, including medical devices, are
exponentially distributed and the times between failures of the same device are independent.
Practicing statisticians make this erroneous assumption when they fit a distribution to the failure
data without checking whether the data exhibit a trend. When there is a trend, models such as
NHPP can be used to describe the failure process, and the time to each event can be analyzed
separately.
4.5. Concluding Remarks 98
Current trend analysis methods in the literature mention only right censoring and do not
address recurrent failure data with left or interval censoring and periodic or nonperiodic
In this chapter, we first propose a method to analyze statistically maintenance data for
complex medical devices with censoring and missing information. We present a classification of
different types of failures and establish policies for analyzing data at system and component
levels taking into account the failure types. We use the midpoints of the censoring intervals as
the failure times to apply the Laplace trend test at several levels of medical equipment (infusion
pump): the system level, hard failures, soft failures, and individual components. Trend analysis
using the midpoints is simple with no need of complex calculation. Reliable results are obtained
Obviously, precise and reliable results demand precise failure data. Since medical devices
tend to be highly reliable, scarce failure data is always a problem in statistical analysis. A
solution could be to aggregate the CMMS data from hospitals with similar equipment. This
aggregation should be performed cautiously, however, since the same device may exhibit
process with a power law intensity function. We propose using the likelihood ratio test to check
for trends in the failure data with censoring. We use the EM algorithm and a modification of the
EM to find the parameters which maximize the data likelihood in the case of null and alternative
hypotheses (no trend and trend assumptions). A recursive procedure is used to solve the main
technical problem of calculating the expected values in the E-step requiring numerical
integration. The proposed method is used to perform a trend analysis of some components of a
4.5. Concluding Remarks 99
general infusion pump such as audible signal, chassis/housing, and battery using a hospital’s
maintenance data. The results are compared to an ad-hoc method of replacing unknown failure
The results reveal that the parameters estimates which are obtained from the complete
EM and its modified version are very close. Moreover, although the parameters obtained from
the EM may be close to what is obtained from the mid-points method, the latter may not always
give a precise approximation of the parameters and the conclusion made for trend in failure data
may not be correct. At least, the values obtained from the mid-points method may be used as
It should be also noted that other methods, such as Monte-Carlo simulation can be also
used to estimate the expected values required in the E-step of the EM (Levine and Casella,
2001). In this case, no recursive procedure needs to be used, which is not discussed here.
model. To establish an optimal inspection policy at each level of the system, a model should take
into account the failure trend and life pattern at that level. This problem will be discussed in the
following chapters.
5. INSPECTION AND MAINTENANCE OPTIMIZATION MODELS
assumptions for a multi-component repairable system with soft (hidden) and hard failures.
Although the inspection optimization models are developed particularly for finding the optimal
periodic inspection interval for medical devices, they can still be applied more generally to any
kind of repairable system. The assumptions used to construct the optimization models are made
electronic instruments consist of a large number of interacting components, which perform the
system’s required functions. A system is called repairable, if after failure can be returned to
satisfactory functioning by any method other than replacement of the entire system (Ascher and
planned inspections during its life cycle. Periodic preventive maintenance (PM) due to its
relatively
100
5.1. Literature Review 101
simple implementation is perhaps the most common maintenance policy applied to repairable
systems in practice. Planned periodic inspections are performed to verify the system safety and
performance by detecting potential and hidden failures and taking appropriate actions. The
actions can be to fix the potential failure if the device is found defective or perform preventive
In many real-world situations where the safety and reliability of devices is vital, devices
must be inspected periodically. For example, according to regulations, all major components and
functions of clinical medical devices in a hospital must be checked at inspection to assure that
the devices are safe for use on patients. If a device fails between two consecutive scheduled
inspections and operators discover the failure, it is checked and repaired immediately regardless
the device’s failed component and to inspect all other major components/functions. In fact,
clinical engineers working in a hospital are supposed to go through a predesigned checklist when
inspections.
Numerous models have been developed for inspection and maintenance optimization of
repairable systems. These models can be classified according to the degree to which the
operating condition of the system is restored by maintenance (Nakagawa and Mizutani, 2009;
Wang and Pham, 2006; Wang, 2002; Pham, 2003, Pham and Wang, 1996; Kijima, 1989; Pham,
2008; Abdel-Hameed, 1995). Perfect PM models assume that the system is replaced or is as good
as new after PM (Barlow et al., 1965; Nakagawa, 2005; Dohi et al., 2003). In models with
assumption of minimal repair, the hazard rate of the system remains undisturbed by PM (Barlow
and Hunter, 1960; Dohi et al., 2003; Berg and Cleroux, 1982). Imperfect repair restores the
5.1. Literature Review 102
system to somewhere between as good as new and as bad as old. Worse and worst repairs make
the failure rate increase; however, the system does not break down in the worse repair, while the
Some models assume that at failure a system is minimally repaired with probability
p and replaced with probability 1 p (Brown and Proschan, 1983; Nakagawa and Yasui, 1987;
Murthy and Nguyen, 1981). This assumption is extended in other models in which the
probability of a repair type (replacement, minimal repair, or imperfect repair) is age dependent
Nakagawa and Mizutani (2009) in their recent paper give an overview of replacement
policies with minimal repair, block replacement and simple replacement for an operating unit in
a finite time horizon. In all three policies, they assume that the unit is replaced at periodic
replacement times and is as good as new. In the policy with minimal repair, the unit is minimally
repaired if it fails between periodic replacements. In the block policy, the unit is always replaced
at failures between replacements, and in the simple policy, there is an elapsed time between a
failure and its detection at the next inspection time after failure. Moreover, the paper presents
optimal periodic and sequential policies for an imperfect preventive maintenance and an
inspection model for a finite time span. In the inspection model, it is assumed that the unit is
replaced if a failure is detected at inspection; so, this model describes mainly a failure-finding
optimization problem.
Sometimes the probability of a repair type depends on both age and other time-dependent
variables. In the so-called “repair number counting” and “reference time” policies, the
actions in a cycle (Makis and Jardine, 1992; Sheu and Chang, 2009; Sheu et al., 2006; Sheu et
5.1. Literature Review 103
al., 2005). In some shock models the probability of replacement depends on system age and the
number of shocks that the system tolerates since the last replacement (Chien and Sheu, 2006;
Lam, 2009). Some shock models assume that the system is replaced depending on age, number
of imperfect repairs, and the cumulative damage exceeding a specific threshold (Qian et al.,
2003; Li and Pham, 2005). In repair limit policies, replacement of the system at failure is decided
according to the cost of repair (Chien et al., 2009; Lai, 2007) or the required repair time (Dohi et
al., 2001a). If the cost or repair time exceeds a certain limit, the system is replaced.
Further inspection and maintenance models for single-unit and multi-unit systems are
discussed in detail in Nakagawa (2003, 2005, 2008), Dohi et al. (2003),Wang and Pham (2006),
and Wang (2002) including age-dependent, periodic PM, failure limit, sequential PM, and repair
The majority of maintenance optimization models assume that the system is not operative after a
failure, so the failure is rectified immediately and the failure time is known. However, up to 40% of all
failure modes of a complex industrial system fall into the category of hidden failures (Moubray,
1997). Hidden failures are not evident to operators; they remain dormant, do not stop the
system’s operation, and are only rectified at inspections. Hidden or soft failures may not halt the
system, but can have serious consequences, even catastrophic, if left unattended. These types of
failures can reduce the system’s performance or cause production loss and require to be fixed
eventually.
There are some models developed for failures with delayed detection when the failure
can only be rectified at inspection. Optimization models developed for failure finding intervals
are among this category. The delay-time model (Christer, 2002; Wang and Christer, 2003)
regards the failure process as a two-stage process. First, a defect is initiated and if unattended, it
5.1. Literature Review 104
will lead to a failure. The delay time between the initiation of the defect and time of failure
represents a window of opportunity for preventing the failure. Wang in his recent paper (2009)
presents an inspection optimization model for minor and major inspections, which are carried out
for a production process subject to two types of deterioration. Minor and major defects are
identified and repaired respectively at routine and major inspections. Taghipour and Banjevic
(2011a) present two inspection optimization models over finite and infinite time horizon for a
A great number of models assume that the system is replaced when a failure is rectified at
inspection, so it is as good as new after failure (Cui, 2005; Lienhard et al., 2008; Okumura,
2006). Available models in the literature mainly consider a general cumulative distribution
function for the system’s time to failure (Okumura, 2006; Jiang and Jardine, 2005). Lienhard et
al. (2008) assume that the failure rate and occurrence rates of operational demand and corrective
maintenance actions are given, and propose a model to determine the optimal length of failure-
finding task interval. Taghipour et al. (2010a) consider a repairable system with components
subject to hard and soft failures; soft failures are only rectified at periodic inspections and are
fixed with minimal repairs. They propose a model to find the optimal periodic inspection interval
In the literature, the opportunistic inspection and maintenance is also discussed. When
maintenance policy is more effective. This policy is beneficial when the cost of joint
maintenance actions of several components is less than the cost of maintenance action for
individual components (Sherif and Smith, 1981). Nowakowski and Werbińka (2009) present an
for multi-unit systems. Gertsbakh (2000) formulates a policy for opportunistic replacement of a
to random failures. They assume that the components are configured in series, and propose a
model to find the replacement time at which a decision should be made regarding the preventive
should not be replaced until the next scheduled inspection. A general failure distribution for each
component is considered and the total cost is formulated for a cycle. The cycle is the minimal
multiple non-identical life-limited components with both economic and structural dependence.
Zhou et al. (2006) develop a dynamic opportunistic maintenance policy for a continuously
monitored multi-unit series system with integrating imperfect effect into maintenance activities.
They assume that a scheduled or unscheduled maintenance action has to be carried out for a unit,
whenever it fails or reaches a certain reliability threshold. At a maintenance action, the whole
system is shut down and this provides a maintenance opportunity for all other units. They
propose a dynamic decision process, which includes calculating the reliability threshold for each
unit, determining opportunistic maintenance cost saving, and maintenance activities grouping
Total expected cost and expected downtime per unit time are usually considered for the
inspection/replacement decision problems. Cui et al. (2004) present several inspection models
with the emphasis on meeting the availability requirement. Barlow and Hunter (1960) define an
5.1. Literature Review 106
optimal maintenance policy as a policy which maximizes the fractional amount of uptime over
long intervals. Boland and Proschan (1982) present two models, which minimize the expected
cost for a system under minimal repair-periodic replacement policy. Boland (1982) assumes that
the cost of minimal repair of a system is a nondecreasing function of its age and derives the
expected minimal repair cost in an interval in terms of the cost function and the failure rate of the
system. Hartman and Murphy (2006) present a dynamic programming approach to the finite-
horizon equipment replacement problem with stationary costs. Dohi et al. (2001b) propose a
model to find the optimal preventive maintenance schedule, which minimizes the relevant
repairable system subject to hard and soft failures. For practical purposes and simpler
implementation, most of organizations especially dealing with a large number of devices use a
periodic inspection policy. A non-periodic inspection policy can be also considered using the
same mathematical models but with more time consuming calculation and more demand on
practical implementation.
We consider two main cases with the following assumptions and construct the
The structure of the inspection optimization models developed in this research is shown
in Figure 5.1. The models are classified according to the assumptions introduced.
5.2. Case 1: A System Under Periodic Inspections 107
We consider a repairable system with the assumptions described in Section 4.3, and
develop a model to find the optimal periodic inspection interval for the system. In general, the
expected cost incurred from both soft and hard failures that include their inspection and failure
consider a non-opportunistic policy in which at hard failures only the failed component is
5.2. Case 1: A System Under Periodic Inspections 108
inspected and fixed and at periodic scheduled inspections only all components with soft failures
are checked and fixed if found failed. In this case, because of instantaneous repair or replacement
of hard components, they do not have impact on the inspection policy, and will be ignored in
calculation in this section. Therefore, the model reduces to a model of a system consisting of
We assume that the components with soft failures are repaired if found failed at
inspections to the same condition as just before failure (minimal repair), even if they age in
certain way while in the failed state. It should be noted that at inspection only the states of
components with soft failures are found, but not their age at failure (i.e., the times of soft failures
are censored), which makes the problem of estimating the failure rate complicated (see Section
4.2. in Chapter 4). A summary of additional assumptions in this section are given as follows:
Additional Assumptions
given. Our objective is to find the optimal scheduled inspection interval, which minimizes the
2 (k 1) k
n T
S S S S
S: Scheduled inspection
We assume that soft failures have influence on the system’s performance. For example,
any changes in the shape of a measuring head in a thickness gauge used in steel industry can
produce less accurate measurement of a steel strip, although the whole gauge is still working.
Therefore, the components with soft failures affect performance and maintenance cost, but not
Several methods can be employed to incorporate soft failures into the model. In fact, the
method depends on the actual application for which the optimization model is developed. The
importance and consequences of soft failures on the system’s performance can be used as criteria
by managers to decide about the form of including this type of failure in the model.
When the downtime of a component with soft failure is important, a method considering
downtime should be employed. At times, the longer a component is in a failed state, the more
significant is its influence on system performance or safety. For example, a defective head in a
thickness gauge can produce inaccurate measurements starting from the moment of its
deformation and lasting until the defect is rectified at inspection. Therefore, a penalty should be
imposed for the period of less accurate measurements. Another method of incorporating soft
failures in the cost model is discussed in Section 5.2.1.2. This mehtod considers the number of
soft failures which exceeds a pre-defined threshold. The method of considering downtime of soft
failures can be combined with the total number of failed components (Section 5.2.1.3) to
emphasize both total downtime and the number of failures exceeding the threshold.
5.2. Case 1: A System Under Periodic Inspections 110
Notation
E[CS(( k 1) ,k ] ] expected cost incurred from a scheduled inspection at k over time period
((k 1) , k ] .
For each inspection interval ((k 1) , k ] , k 1, 2,...,n , and for all components with soft
failure a penalty cost is incurred proportional to the elapsed time of the failure. The objective is
to find the optimal inspection interval, which minimizes the total penalty cost. We assume that
inspection and possible repairs are also done at the end of the cycle, T , that is, for k n . This
action may be considered as a sort of minimal preparation for the next cycle, but also makes the
- Cost of inspection ( c I ).
When a component fails, it remains failed until the next inspection, when it is minimally
repaired. So, the expected cost incurred from the k th scheduled inspection in a cycle is:
m
E[CS(( k 1) ,k ] ] cI c j P(component j fails in (( k 1) , k ])
j 1
m
(5.2)
cj E (downtime of component j in ((k 1) , k ])
j 1
From additivity of the cost function, it is clear that any structure of the system can be
ignored because only the marginal distributions of failure times are needed. Still the components
are economically dependent, by sharing the same inspection interval. Let X k be the survival
time of the component j in the k th interval, calculated from (k 1) (Figure 5.3), k 1, 2,..., n
(index j is suppressed for simplicity). Let X 0 is the initial age of the component at the
beginning of the cycle. Using initial age is convenient for recursive calculation, but in real
application, it is also possible that some components are not as good as new. In the numerical
We will derive a recursive equation for Pk j (t ) using conditioning over the first failure time X 1 .
X2 T
X1 X3 Xk
Let X 1 min{ X , } , where X is the time of the first failure. Therefore, X1 , and the CDF
of X 1 is
tx
j ( s ) ds
,0 x
F1 ( x | t ) P j ( X 1 x | X 0 t ) 1 e
j t
(5.3)
1 , x
tx
j ( s ) ds
f1 ( x | t )
F1 ( x | t ) j (t x)e , 0 x .
j j
x
t
(5.4)
t
j ( s ) ds
Notice that P1 (t ) P ( X1 | X 0 t ) F1 ( | t ) F1 ( 0 | t ) e
j j j j t
.
Pk j (t ) P j ( X k | X 0 t ) E[ P j ( X k | X 0 t , X 1 )]
P j ( X k | X 0 t , X 1 x) f1 ( x | t )dx P j ( X k | X 0 t , X 1 ) P j ( X 1 | X 0 t )
j
0
P (t x)dF1 ( x | t ) Pk j1 (t ) P1 j (t )
j j
k 1
0
Pk j1 (t x)dF1 ( x | t ) , for k 2,...,n .
j
j
t
j ( s ) ds
P1 (t ) e t ,
. (5.5)
Pk j (t ) Pk j1 (t x)dF1 j ( x | t ) , k 2,...,n
0
mentioned later in the numerical example. Direct calculation of Pk j (t ) would be tedious, except
P2 (t ) P (t x)dF1 ( x | t ) P1 j (t ) P1 j (t )
j j j
1
0
t x tx t 2 t
j ( s ) ds j ( s ) ds j ( s ) ds j ( s ) ds
e tx
j (t x)e t
dx e t
e t
0
t x t 2
j ( s ) ds j ( s ) ds
j (t x)e t
dx e t
,
0
and
P3 j (t ) P (t x)dF1 j ( x | t ) P2j (t ) P1 j (t )
j
2
0
t x y t x 2 tx
j ( s ) ds j ( s ) ds j ( s ) ds
[ j (t x y )e tx
dy e tx
] j (t x)e t
dx
0 0
t x 2 t 3 t
j ( s ) ds j ( s ) ds j ( s ) ds
[ j (t x )e t
dx e t
]e t
0
t x y t x 2
j ( s ) ds j ( s ) ds
j (t x) j (t x y )e t
dxdy j (t x)e t
dx
0 0 0
t x 2 t 3
j ( s ) ds j ( s ) ds
j (t x )e t
dx e t
.
0
Let
e (t ) E [ E [ X k | X 0 t, X 1 ]] E j [ X k | X 0 t, X 1 x ]dF1 ( x | t )
j j j j
k
0
E j [ X k 1 | X 0 t x ]dF1 ( x | t ) ekj1 (t x )dF1 ( x | t ) .
j j
0 0
j
So, the recursive equation for ek (t ) is
5.2. Case 1: A System Under Periodic Inspections 114
ek (t ) ek 1 (t x) f1 ( x | t )dx ek 1 (t ) P1 (t ) , k 2,...,n .
j j j j j
(5.6)
0
For k 1 ,
tx
j ( s ) ds
e (t ) E [ X 1 | X 0 t ] xdF1 ( x | t ) xf1 ( x | t )dx P1 (t ) R ( x | t )dx e
1
j j j j j
1
j t
dx.
0 0 0 0
If we consider a penalty cost of c j for the elapsed time of failure for component j , we will have
m n n m
E[CST ] ncI c j (n Pk j (t )) cj ( ekj (t ))
j 1 k 1 k 1 j 1
. (5.7)
m m m n
T cj n(cI c j ) (c j Pk (t ) c e (t )) j
j k
j
j 1 j 1 j 1 k 1
g kj (t ) c j Pk j (t ) cj ekj (t ) (c j Pk j1 (t x) cj ek 1 (t x))dF1 ( x | t ),
j j
0
g kj (t ) g kj1 (t x)dF1 ( x | t ).
j
n
Finally, if g nj (t ) g kj (t ) , the expected cost per cycle is given as follows:
k 1
m m m
E[CST ] T cj n(cI c j ) g nj (t ).
j 1 j 1 j 1
5.2. Case 1: A System Under Periodic Inspections 115
Our objective is to find the optimal inspection frequency n , which minimizes the
expected cost per cycle E[CST ] incurred from scheduled inspections. For that purpose, we
calculate E[CST ] for different values of n and find the value of n . The procedure is numerically
intensive. It should be noted that n is limited by an upper bound nU . This upper bound may
interval, e.g., one week, or one month. Then n L T , or n T / L nU . Note that E[CST ] ncI
for all n . Let E[CS(0,T ] ] be the expected cost incurred from a plan with only one inspection
Numerical example
Consider a complex repairable system with five components. We assume that the number
j 1
x
function j ( x) j . The parameters of the power law processes and the minimal repair
j j
costs of the components are given in Table 5.1. The system is currently inspected periodically
every 4 months. The number of components and the parameters are adapted from a case study on
medical equipment described in Section 4.2 in Chapter 4. We are interested in finding the optimal
inspection interval, which minimizes the total cost incurred from scheduled inspections over a
12-month cycle length. For simplicity, we will assume that the system starts as good as new,
which is not a limitation in our model, as we may assume any initial age, t , of the system (or
even different initial ages, t t j , of different components j ). We may assume that after 12
5.2. Case 1: A System Under Periodic Inspections 116
months a major overhaul of the system will be performed, e.g., some components are replaced
with new or used spare components, and the system will start again. Then, a new optimal
inspection frequency should be found. We also assume that the inspection interval cannot be
Table 5.1. Parameters of the power law intensity functions and costs
current policy = 4 months, or n 12/ 4 3 . The inspection cost is cI $200 . We use equation
We obtain Pk j (t ) , ekj (t ) , k 1,..., n , recursively using equations (5.5) and (5.6), and
numerical integration (Composite Simpson’s rule was applied for numerical integration with step
length h depending on n and required accuracy; after some experimentation, h 104 was
used). The value t 0 is used in the calculation, that is, we assume the equipment is as good as
The total cost E[CST ] for different values of n is shown in Figure 5.4. As it can be seen
on the Figure, the minimal cost is obtained for n 4 with the total cost E[CST ] = $3,932.88. So,
5.2. Case 1: A System Under Periodic Inspections 117
if the real costs were as given, the current inspection policy should be changed from inspecting
5.50
5.00
4.50
4.00
3.50
3.00
1 2 3 4 5 6 7 8 9 10 11 12
Cost 4.55 4.08 3.94 3.93 3.99 4.08 4.19 4.33 4.47 4.63 4.79 4.96
Figure 5.4. Total costs for different inspection frequencies (first model)
This model is sensitive to the penalty costs incurred for components’ downtimes. If the
downtime penalty costs c j are not small, the optimal frequency will likely not be n 1 , as it
may happen in the next proposed model (model considering the number of failures). In other
words, the longer interval between inspections the longer the elapsed time of failures, and
In this variation of the model, we introduce a threshold N trh for the acceptable number of
failures observed at a scheduled inspection. According to the total number of failures that
exceeds the threshold, a penalty cost is incurred. This method can be used when only the total
example of an application of this method can be a computer system with several external hard
5.2. Case 1: A System Under Periodic Inspections 118
drive backups; even if all hard drive backups fail, the system can function, but a penalty should
be incurred since the chance of losing data is increased by increasing the number of failed
backups. Thresholds can be set by managers responsible for the safety and reliability of the
system. A threshold may also be related to the number of redundant components in the system
when a minimum required number of components must be operational, or to the additional costs
of labor and downtime required to fix accumulated problems. We also assume that a penalty cost
Additional Notation
N trh acceptable threshold for the number of failures detected at a scheduled inspection.
c P penalty cost incurred for exceeding the acceptable number of failures at a scheduled
inspection.
c P penalty cost incurred per number of failures exceeding the threshold N trh .
- Cost of inspection ( c I ).
- Penalty cost which takes into account the number of failures exceeding N trh ( c P ).
The expected cost incurred from the k th scheduled inspection in a cycle is:
m
E[CS(( k 1) ,k ] ] cI c j (1 Pk j (t )) E[cP I{Sk Ntrh } ] E[cP max{0, S k N trh }] , (5.8)
j 1
5.2. Case 1: A System Under Periodic Inspections 119
where S k is the total number of failures in the k th interval, and I{Sk Ntrh } is the indicator function.
Let pk (l ) P( Sk l | initial age of all components t ) , l 0,1, 2,..., m ( n and t are suppressed
for simplicity). Then the expected penalty cost of exceeding the acceptable number of failures,
E[cP I{Sk Ntrh } ] E[cP max{0, Sk - N trh }] cP E[ I{Sk Ntrh } ] cP E[max{0, S k - N trh }]
cP
l Ntrh
pk (l ) cP
l - Ntrh 0
(l - N trh ) pk (l ) cP
l Ntrh
pk (l ) cP ( lpk (l )
l Ntrh
N
l Ntrh
trh pk (l ))
In our proposed model the more the number of failures exceeds N trh , the larger the
We need to find the probability of having N trh 1 or more of failures, which may be
failure of components need not be equal, the number of failed components in the system does not
follow the binomial distribution. So, in order to find pk (l ) , one can use the following equation
(Levitin, 2005):
m m l 1 qj1 m l 2 qj 2 m qj l
pk (l ) [ (1 q j )][ ... ] (5.9)
j 1 j 1 1 1 q j 1 j 2 j 1 1 1 q j2 j l j l 1 1 1 q j l
interval. In this calculation, we assume that components are independent. Otherwise, the
calculation would require knowledge of system structure and a joint distribution of failure times,
algorithm suggested by Barlow and Heidtmann (1984) based on generating function can be used
The probability mass function of each component can be represented by the generating
function:
ukj ( z ) E j ( z X k | X 0 t ) Pk j (t ) z 0 (1 Pk j (t )) z1 Pk j (t ) (1 Pk j (t )) z .
m m
U k ( z ) ukj ( z ) pk (l ) z l generates the distribution of S k . We are interested in
j 1 l 0
finding
l Ntrh
pk (l ) and lp (l ) .
l Ntrh
k
Define
m
U k ( z)
l Ntrh 1
pk (l ) z l (5.10)
Then
l Ntrh
pk (l ) U k (1) and lp (l ) U (1) , respectively. So, we have:
l Ntrh
k k
E[cP I{Sk Ntrh } ] E[cP max{0, Sk - Ntrh }] (cP NtrhcP )U k (1) cPU k (1) (5.11)
m
E[CS(( k 1) ,k ] ] cI c j (1 Pk j (t )) (cP N trh cP )U k (1) cPU k (1). (5.12)
j 1
Therefore, the expected cost incurred from all scheduled inspections over the cycle T n is:
5.2. Case 1: A System Under Periodic Inspections 121
n
E[CST ] E[CS(( k 1) ,k ] ]
k 1
n m n
ncI c j (1 Pk j (t )) [(cP N trh cP )U k (1) cPU k (1)]
k 1 j 1 k 1
m n n n
ncI c j (n Pk j (t )) (cP N trh cP )U k (1) cP U k (1)
j 1 k 1 k 1 k 1
m m n n
n(cI c j ) c j P n (t ) (cP N trh cP )U k (1) cP U k (1),
j
j 1 j 1 k 1 k 1
(5.13)
n
where P n (t ) Pk j (t ) .
j
k 1
This method is particularly useful when the components are similar or make almost the
same contribution to reducing the whole system’s performance. In other case, a weight can be
assigned to each component to describe its contribution to the system performance, system
production, etc. In the latter, different combinations of failures can cause different penalty costs
since both the occurrence of a failure and its individual influence on system performance or
production are incorporated into the model through the weight. The combined influence of
several failures can be considered as a weighted sum of the influence of all individual
section, we consider additive downtime costs incurred for individual but simultaneous soft
failures as their so-called joint failure consequence; but the model can be revised to intensify the
failure consequence or risk of simultaneous failures. In this case, if for example two soft failures
take place simultaneously, their joint downtime consequence should be greater than the sum of
their individual downtime consequences. For this variation of the model, possessing a good
5.2. Case 1: A System Under Periodic Inspections 122
knowledge of the system, its structure, components dependency, and their joint failure
consequences is required.
Moreover, we assume that soft failures cannot convert into hard failures, and hard and
soft failures have independent consequences on the system. In other words, we do not consider
joint risk or failure consequence of soft and hard failures even though they may be related
somehow. For example, if both protective component (component with soft failure) and
protected component (component with hard failure) fail simultaneously, our proposed model
considers only the consequence of the soft failure, since it assumes that a hard failure is fixed
instantaneously; however, the model can be extended to consider the joint failure consequence of
soft and hard failures when they have some sort of dependency. Such failure dependency and
joint risk and cost consequences have been extensively discussed in the risk assessment literature
Numerical example
We use the same data given in the numerical example in Section 5.2.1.1 to demonstrate
the model suggested in this section. The other parameters of the model are: threshold N trh 3 ,
We will use equation (5.13) in the following form for convenience of comparing the
m n n n
E[CST ] ncI c j (n P n (t )) cP U n ,k (1) cP ( U n ,k (1) N trh U n ,k (1)) .
j
j 1 k 1 k 1 k 1
5.2. Case 1: A System Under Periodic Inspections 123
n
To calculate P n (t ) Pnj,k (t ) for n 1,2,3,...,12 we use Pnj,k (t ) , k 1,..., n . Dependence
j
k 1
of Pnj,k (t ) and U n , k ( z ) on n is emphasized in the notation. Pnj,k (t ) values are also used to find
U n , k (1) and Un,k (1) . We first extract the terms with z4 and z5 from
5
U n,k ( z ) [(1 Pnj,k (t )) z Pnj,k (t )] and use it to create U n , k ( z ) . Then U n , k (1) and Un,k (1) can be
j 1
easily obtained. The value t 0 is used in the calculation, that is, we assume the equipment is as
The cost curve is shown in Figure 5.5. n 1 is the optimal inspection frequency, or to
5.00
4.50
4.00
3.50
3.00
1 2 3 4 5 6 7 8 9 10 11 12
Cost 3.76 4.93 4.92 4.73 4.53 4.40 4.32 4.31 4.35 4.42 4.52 4.64
Figure 5.5. Total costs of scheduled inspections for different inspection frequencies
(threshold model)
Obviously, the optimal n depends on the input parameters of the model. With the cost
parameters given in this example, n 1 is the optimal policy. However, we would like to find
conditions for costs under which some other value of n , such as n 9 is the optimal solution.
5.2. Case 1: A System Under Periodic Inspections 124
m n n n
Q1 (n) c j (n P n (t )), Q2 (n) U n ,k (1), Q3 (n) U n, k (1) N trh U n ,k (1) .
j
Let Then the
j 1 k 1 k 1 k 1
expected cost for n is E[CST ] ncI Q1 (n) cP Q2 (n) cP Q3 (n) , if we assume that c j costs are
1 1
fixed. Table 5.2 shows ncI , Q1 (n) , Q2 (n) , and Q3 ( n) calculated for n 1,2,3,...,12 .
For a particular n to be better than n the following relationship of c P and c P should
hold:
ncI Q1 (n) cP Q2 (n) cPQ3 (n) ncI Q1 (n) cPQ2 (n) cPQ3 (n) ,
inequalities can be obtained for other values of n . For cP $1500 , the previous inequality gives
5.2. Case 1: A System Under Periodic Inspections 125
the lower limit cP $1490 . It appears that for, e.g., cP $1700 , n 9 is also the optimal
inspection frequency. The cost function for that case is shown in Figure 5.6. Obviously, when the
cost curve is flat near the optimal point, i.e. when there is no a significant difference between the
total cost of the optimal frequency and the other frequencies close to the optimal, we can select
the frequency n close to the optimal that is easier to implement practically. For instance, the
costs for n 12 may be considered close enough to the minimal cost, so n 12 can be selected
implementation, then n 1 , or only one inspection at the end of the cycle may be the next
Figure 5.6. Total costs of scheduled inspections for different inspection frequencies (threshold
model); For penalty cost cP $1700 , n 9 is the optimal number of inspections when
cP $1500
Comparing the values of ncI , Q1 (n) , Q2 (n) , and Q3 ( n) given in Table 5.2 for different n
indicates that for some n the values are always greater than for some other n and consequently
no value for c P and c P can make them optimal. For instance, since all values for n 2 are
greater than the values for n 1 , n 2 can never be optimal. Therefore, some kind of “peculiar”
5.2. Case 1: A System Under Periodic Inspections 126
behavior of the cost function for n 1 in both Figure 5.5 and 5.6 should not be a surprise in this
example. In practical terms, it means that to decrease expected penalty costs it is better to leave
soft failures unattended ( n 1 ), than to inspect them after 6 months ( n 2 ) and minimally fix
them, because it will then only create another a not small chance for them to fail again and pay a
penalty due to exceeding the threshold. But, using n 2 may significantly reduce the chance of
exceeding the threshold, and then may reduce the expected penalty cost.
expected penalty cost. However, this calculation can be very complicated when the number of
components increases. Also, using a threshold may produce the situation discussed above that it
is sometimes better not to fix failures. The model discussed in Section 5.2.1.1 suggests another
version of the cost model whose optimization requires simpler calculation while it takes into
Equations (5.7) and (5.13) can also be combined to take into account both the exceeding
number and elapsed times of failures, thus making the cost model more flexible. It should be
noted that the “threshold” part of the model requires calculation of the distribution of the total
number of failures, whereas the “downtime” part of the model requires only the marginal
m n n n n m
E[CST ] ncI c j (n P n (t )) cP U n ,k (1) cP [ U n ,k (1) N trh U n ,k (1)] cj ( ekj (t )).
j
j 1 k 1 k 1 k 1 k 1 j 1
5.2. Case 1: A System Under Periodic Inspections 127
The results of applying the combined model on the data given in the examples in Sections
5.2.1.1 and 5.2.1.2 and the models considering threshold and elapsed times are shown in Figure
5.7. As it can be seen in Figure 5.7 n 10 with the total cost E[CST ] = $5,571.84 is the optimal
solution for the combined model. Moreover, obviously, there is a significant difference between
costs for smaller n and larger n . The cost for n 1 is still smaller than the cost for n 2 , but
not significantly, due to significant downtime cost for n 1 . For practical purposes, it may be
simpler to use n 12 , i.e. to inspect every month, which makes only a minor difference from the
optimal solution.
8.50
7.50
6.50
5.50
4.50
3.50
2.50
1 2 3 4 5 6 7 8 9 10 11 12
with threshold 3.76 4.93 4.92 4.73 4.53 4.40 4.32 4.31 4.35 4.42 4.52 4.64
with elapsed time 4.55 4.08 3.94 3.93 3.99 4.08 4.19 4.33 4.47 4.63 4.79 4.96
combined model 7.70 7.91 7.38 6.83 6.38 6.04 5.81 5.67 5.59 5.57 5.59 5.64
Figure 5.7. Total costs for different inspection frequencies (combined model)
In general, the methodology for considering the influence of different failures can be set
up by managers responsible for maintaining devices in a safe and reliable state. When
information about the influence of individual failures and different combinations of failures is
available, a threshold model with a weighted sum of the performance reduction or loss
production can be employed. However, in many real situations, such as hospitals dealing with
many different devices, the availability of this type of information is problematic. In this case,
5.2. Case 1: A System Under Periodic Inspections 128
simpler and more practical methods, such as considering the total number of failures or the sum
In this section, we have the same assumptions as Section 5.2.1, but in addition we assume
that the components with soft failures can be either minimally repaired or replaced if they fail
with some age dependent probabilities. So, in this section more general assumptions are made
compared to Section 5.2.1. We propose models to find an optimal periodic inspection interval for
the system over finite and infinite time horizons. We consider a penalty cost for the downtime of
the failed component. Additional assumptions of this section are summarized as follows:
Additional Assumptions
Failed components may be minimally repaired or replaced with age dependent probabilities.
A penalty cost is incurred for the elapsed time from a failure to its detection at the next
inspection. The longer the elapsed time, the larger penalty is incurred.
We assume that periodic inspections are taken at times k (k 1,2,...,n) . Our objective is
to find the optimal scheduled inspection interval, which minimizes the expected cost incurred
- Cost of inspection ( c I )
- Cost of minimal repair of a component if found failed and needing minimal repair ( c Mj )
all failed components, a penalty cost is incurred proportional to the elapsed time of the failure.
We assume that inspection and possible repairs are also done at the end of the cycle, T , that is,
for k n .
The expected cost incurred at the k th (k 1, 2,..., n) scheduled inspection in a cycle is:
m
E[CS(( k 1) ,k ] ] cI c Mj Pr{component j fails in ((k 1) , k ] and is minimally repaired}
j 1
m
c Rj Pr{component j fails in ((k 1) , k ] and is replaced} .
j 1
m
c Dj E[downtime of component j in ((k 1) , k ]] .
j 1
(5.15)
Let X 0j t j is the initial age of the component at the beginning of the cycle. The initial
Let X kj be the survival time of component j in the k th interval, calculated from (k 1)
Let
Pk j (t ) Pr{I kj 2 | X 0j t} P{ X kj | X 0j t} ,
M kj (t ) Pr{ X kj , I kj 0 | X 0j t} ,
Rkj (t ) Pr{ X kj , I kj 1| X 0j t} ,
ekj (t ) E[ X kj | X 0j t ] .
Then from (5.14) and (5.15) the total expected cost incurred over [0, T ] can be written as follows:
m n m n
E[CST ] ncI (c Mj M kj (t ) Rkj (t )c Rj ) c Dj ( ekj (t ))
j 1 k 1 j 1 k 1
(5.16)
m m n n n
T c ncI [c
D
j
M
j M k
j
(t ) c R
j R k
j
(t ) c D
j e k
j
(t )].
j 1 j 1 k 1 k 1 k 1
We can ignore any structure of the system since the cost function is additive and only
marginal distributions of failure times are required. However, the components still are
(k 1, 2,..., n) , which applies for every component j . For simplicity, we will suppress index j
in the following, unless it is needed for clarity. The recursive equations are identical aside from a
The CDF and pdf of X 1 are the same as what were introduced in Section 5.2.1.1
(equations 5.3 and 5.4 ) and P1 (t ) is with the same definition as follows:
t
( s ) ds
P1 (t ) Pr{X1 | X 0 t} F1 ( | t ) F1 ( 0 | t ) e t
.
Then from the assumptions of NHPP and minimal repair or replacement, we get the
recursive equations:
1 (t ) E[ ( X1 , I1 ) | X 0 t ] [ ( x, 0)r (t x) ( x,1)r (t x)] f1 ( x | t )dx ( , 2) P1 (t ), (5.17)
0
k (t ) E[ ( X k , I k ) | X 0 t ] k 1 (t x)r (t x) f1 ( x | t )dx k 1 (0) g1 (t ) k 1 (t ) P1 (t ), (5.18)
0
for k 2,...,n , where g1 (t ) r (t x) f1 ( x | t )dx and r ( x) 1 r( x) . For a proof, see Appendix
0
k (t ) r k 1 (t x) f1 ( x | t )dx r k 1 (0)(1 P1 (t )) k 1(t ) P1(t ) .
0
5.2. Case 1: A System Under Periodic Inspections 132
Using initial age is necessary for recursive calculation. In our numerical examples, initial
age t 0 will be used. In real applications, it is also possible that some components are not as
good as new, so that our method can be applied as well. We now can apply the general recursive
equation (5.17) for some special cases needed for calculating expected cost in equation (5.16).
The only differences between these cases are in their placeholder functions ( X k , I k ) and first
functions 1 (t ) .
( X k , I k ) ( X k , I k 2) ( X k ),
t
( s ) ds
P1 (t ) E[ ( X 1 ) | X 0 t ] Pr{ X k | X 0 t} e t
.
(b) For the probability of failure with minimal repair in the k th interval:
( X k , I k ) ( X k , I k 0),
M 1 (t ) E[ ( X 1 , I1 0) | X 0 t ] Pr{ X k , I k 0 | X 0 t} r (t x) f1 ( x | t )dx.
0
( X k , I k ) ( X k , I k 1),
R1 (t ) E[ ( X 1 , I1 1) | X 0 t ] Pr{ X k , I k 1| X 0 t} r (t x) f1 ( x | t )dx.
0
(d) For the expected survival time of the component in the k th interval:
( X k , Ik ) X k ,
e1 (t ) E ( X 1 | X 0 t ) xdF1 ( x | t ) F1 ( x | t )dx ,
0 0
where F1 ( x | t ) 1 F1 ( x | t ) .
5.2. Case 1: A System Under Periodic Inspections 133
In real applications, the last inspection interval does not need to be of the same length as
the planned periodic inspection interval . The device is checked periodically at every
interval, but the time between the two last inspections, i.e. T (n 1) , need not necessarily be
equal to , if is fixed in advance. We extend the equation (5.18) to include the case when the
then and the equation (5.18) can be used without any changes; however, in general we
have:
1 ( , ; t ) [ ( x,0)r(t x) ( x,1) r (t x)] f1( x | t )dx ( ,2) P1( ; t ),
0
k ( , ; t ) k 1 ( , ; t x)r(t x) f1 ( x | t )dx k 1 ( , ;0) g1(t ) k 1( , ; t ) P1( ; t ), (5.19)
0
tx
( s ) ds
for k 2,3,..., n , where P1 ( x; t ) e t
.
j 1 j 1 k 1 k 1 k 1
Our objective is to find the optimal inspection frequency n which minimizes the
expected cost per cycle E[CST ] if T / n . For that purpose, we calculate E[CST ] for different
values of n and find the optimal value n . If we use the model with a possibly different last
interval, we first choose , then calculate n and , and then calculate E[CST ] . We usually
Numerical example
We use the same data given in the numerical example in Section 5.2.1.1 to demonstrate
b j x
the model suggested in this section. Assume that rj ( x) a j e , x 0 . The parameters of the
power law processes, minimal repair and replacement costs, and parameters a j and b j of rj ( x)
are given in Table 5.3. The system is currently inspected periodically every 4 months. We are
interested in finding the optimal inspection interval, which minimizes the total cost incurred from
scheduled inspections over a 12-month cycle length. The value t 0 is used in the calculation
since we assume the equipment is as good as new at the beginning of the cycle. We assume that
the inspection interval cannot be shorter than one month, that is, n 12 , and that is a multiple
of one month time unit. The last inspection interval does not need to have the same length as the
Table 5.3. Parameters of the power law intensity functions, costs and rj ( x)
parameters for different components
Using equation (5.20) we calculate the total cost for different inspection intervals over
(5.17) and (5.18) and known functions M 1j (t ) , P1 j (t ) , and e1j (t ) . Rkj (t ) is calculated from
1 [ M kj (t ) Pk j (t )] . Composite Simpson’s rule was applied for numerical integration with step
length h 104 .
for small k .
11.00
10.50
10.00
9.50
9.00
8.50
1 2 3 4 5 6 7 8 9 10 11 12
Cost 10.33 9.37 9.20 9.20 9.11 9.36 9.25 9.16 9.15 9.24 9.51 9.80
Figure 5.9. Total costs for different inspection intervals ( ) – finite horizon
As shown, 5 is the optimal inspection interval. This means that the system should be
inspected three times in its life cycle of length T 12 months. The first inspection should be
performed after 5 months, the second inspection after 10 months, and the third two months later,
at the end of the 12 months interval. Some kind of peculiar behavior of the cost curve is due to
the combination of cost function for different components with different failure patterns.
5.2. Case 1: A System Under Periodic Inspections 136
The inspection optimization model over an infinite time horizon uses the renewal-reward
theorem (Tijms, 2003). With this theorem, the objective function can be formulated as the
expected cost per unit time in long run and calculated as the sum of expected costs per unit time
We first formulate the expected cost per cycle for component j . The cycle length for
cI
C j (u) C j (k ) Cfj (u) Ci j (u) Cfj (u) k ,
m
cI
where is the inspections cost shared by component j and Cfj (u ) expected cost due to
m
C j (u ) Cfj (k ) cI
Then , and
u k m
C j (u ) C j (k ) cI EC j cI
lim lim u , (5.21)
u u k k m ET j m
from the renewal reward theorem, where T j is the cycle length for component j , and C j is the
cycle cost for component j . T j can take values , 2 , … so that T j , because the first
replacement can be performed at earliest at the end of the first inspection interval.
We then have
T j (T j k ),
k 1
5.2. Case 1: A System Under Periodic Inspections 137
and E[T j | X 0j t ] Pr{T j k |X 0j t} where Pr{T j |X 0j t} 1 . We will consider later
k 1
when E[T j | X 0j t ] .
Equation (5.21) can be applied to all components j , j 1, 2,..., m , to find the expected
C j
(u ) m
C j (u ) m
C j (k ) m EC j cI
lim j 1
lim lim . (5.22)
u u l
j 1 u k
j 1 k j 1 ET
j
It should be noted that there is no renewal cycle for the system, but separate renewal
We give a general recursive equation for the infinite case similar to equation (5.18). For
1 (t ) E[ ( X 1, I1 ) | X 0 t ] [ ( x,0)r(t x) ( x,1) r (t x)] f1( x | t )dx ( ,2) P1(t ) , (5.23)
0
k (t ) E[ ( X k , I k ) (T k )|X 0 t ] k 1 (t x)r (t x) f1 ( x | t )dx k 1 (t ) P1 (t ), (5.24)
0
In the following special cases, we apply equations (5.23) and (5.24). The differences
( X k , I k ) 1 , V1 (t ) f1 ( x | t )dx P1 (t ) 1 .
0
1, I k 0
( X k , Ik ) , n1 (t ) r(t x ) f1 ( x | t )dx .
0, I k 1, 2 0
X k , I k 0,1
( X k , Ik ) ,
0, Ik 2
d1 (t ) [( x)r (t x) ( x)r (t x)] f1 ( x | t )dx ( x) f1 ( x | t )dx .
0 0
Let
G (t ) E[ ( X k , I k ) ( T k ) | X 0 t ] k (t ) converges for all t .
k 1 k 1
Then
G (t ) 1 (t ) [ i 1 (t x ) r (t x ) f1 ( x | t )dx i 1 (t ) P1 (t )]
i 2 0
1 (t ) r (t x ) f1 ( x | t ) i 1 (t x )dx i 1 (t )P1 (t )
0 i 2 i 2
1 (t ) r (t x ) f1 ( x | t )G (t x )dx G (t ) P1 (t ),
0
or
G(t ) 1 (t ) r (t x ) f1 ( x | t )G(t x )dx G(t ) P1 (t ) . (5.25)
0
Equation (5.25) implies that G (t ) k (t ) can also be calculated recursively from
k 1
G (t x ) .
5.2. Case 1: A System Under Periodic Inspections 139
To calculate n(t ) nk (t ) , d (t ) d k (t ) , and V (t ) Vk (t ) , we have to check
k 1 k 1 k 1
whether they are finite. The convergence of V (t ) is considered in Lemma 1 and Lemma 2 in
The results of Lemma 1, 2, and 3 can be used to calculate Vk (t ) ,
k 1
n (t ) ,
k 1
k and
d
k 1
k (t ) numerically. Details are given in Appendix K.
EC j c j c j n (t ) c j d (t )
R M j D j
. (5.26)
ET j V j (t )
From (5.22) and (5.26) the expected cost for the whole system is
m c R c M n j (t ) c D d j (t )
cI 1 m c j c j n (t ) c j d (t )
R M j D j
m
EC j cI
j 1 ET
j
j 1
j j
V j (t )
j
[
j 1 V j (t )
cI ] . (5.27)
Numerical example
We use the same data given in the numerical example of Section 5.2.2.1. The downtime
costs are given in the second column of Table 5.4. With restriction 12 , 12 months is the
$1,000.00
$950.00
$900.00
$850.00
$800.00
$750.00
$700.00
$650.00
1 2 3 4 5 6 7 8 9 10 11 12
Cost $948. $825. $775. $746. $726. $711. $699. $690. $683. $678. $674. $671.
Figure 5.10. Total costs for different inspection intervals ( ) – infinite horizon
Obviously, the optimum result depends on input parameters of the model. If we try the
calculation with higher penalty costs for downtime given in the last column of Table 5.4, the
$1,000.00
$980.00
$960.00
$940.00
$920.00
$900.00
$880.00
$860.00
$840.00
1 2 3 4 5 6 7 8 9 10 11 12
Cost $986 $894 $869 $861 $858 $859 $860 $863 $866 $870 $874 $878
Figure 5.11. 5 is obtained with downtime penalty costs from the last column of Table 5.4
In this section, we assume that at a hard failure, in addition to the failed component, all
other components are also inspected and fixed if found to have failed. This model is called
opportunistic maintenance since a hard failure provides an opportunity to inspect the other
inspections”. In real situations such as medical devices in a hospital, clinical engineers have to
go through a check list at both scheduled and non-scheduled inspections to test and verify all
major components and features of a device to assure their safety and reliability.
In this section, we want to find the optimal scheduled inspection interval in presence of
to either hard or soft failures with the assumptions stated in Section 4.3. The system is inspected
NS NS NS
HF
2 HF HF (k 1) k
n T
S S S S
It should be noted that some components may have several failure modes including soft
and hard, but in the analysis they may be considered separately. For simplicity, we will identify
components with their dominant failure modes and assume that each component has only one
Additional Assumption
At a non-scheduled inspection, all components are inspected and fixed if found to have failed.
Notation
m1 , m2 and m are the numbers of components with hard and soft failure, and the total
The cost incurred over the cycle T consists of the costs arising from occurrences of soft and
We consider a system with the assumptions stated in Section 4.3 and 5.3. Additional
Additional Assumptions
Additional Notation
Penalty cost for the elapsed time of soft failure for component j ( c Dj )
We assume that right after the occurrence of a hard failure, the failed component and all
components with soft failure are inspected (non-scheduled) with a cost greater than of a planned
inspection. The failed component with hard failure is minimally repaired. Due to the NHPP
assumption, the expected number of hard failures over the cycle T for component i is
T
0
i ( x )dx , i 1, 2,..., m1 . Therefore, the expected cost of unplanned inspections and minimal
repairs for m1 components with hard failure over cycle length T is:
m1
E[CHT ] [(cNI ciM ) i ( x)dx] .
T
(5.29)
0
i 1
Equation (5.29) implies that E[CHT ] does not depend on the periodic inspection interval .
Therefore, this cost does not affect the optimal inspection policy and can be excluded from the
optimization model. It should be noted that the cost of non-scheduled inspections is accounted in
E[CHT ] , and will not be included in E[CST ] . The optimal inspection interval is a value of
which minimizes the expected cost incurred from soft failures in a cycle, i.e. E[CST ] . We include
in E[CST ] the cost of periodic inspections since the components are periodically inspected to
If a component with soft failure fails before a hard failure, the hard failure time is when
the soft failure is rectified. If no hard failure takes place between the soft failure and the next
scheduled inspection, the scheduled inspection is when the soft failure is rectified and fixed. In
both cases, in addition to the cost of minimal repair, a penalty cost is incurred for the downtime
of the component, i.e. the time between the occurrence of a failure and when it is rectified.
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 145
Therefore, in the opportunistic case, hard failures contribute to the expected number of minimal
The total expected cost incurred from m2 components with soft failure over the cycle
We will derive recursive equations to calculate the expected number of minimal repairs
For the convenience of the presentation, we consider now a subsystem of the whole
system consisting of all components with hard failures combined in a serial configuration. Some
of these components may already have some initial ages when they are configured in the
subsystem. So, when the subsystem starts operating, not all its components are new. Let i ( x ) is
the intensity function of the i th component with hard failure, and i is its initial age,
i 1, 2,..., m1 . Then, following from the serial configuration, the intensity function of the “hard”
subsystem at time z is
m1
H ( z ) H ( z ) i (i z ) , (1 , 2 ,..., m ) . 1
i 1
H ( x ) dx
F ( z ) P( Z z ) 1 e
Z
0
,
s z
H ( x ) dx
F Z ( z | s) 1 e s
,
s z
H ( x ) dx
and f ( z | s ) H ( z ) e
Z s
.
To develop the model, we first assume that there is only one inspection interval [0, ] ,
i.e. one scheduled inspection is performed at the end of the interval (time ), and all other
Let X 0j t is the age of component j with soft failure, and Z 0 s is the age of the
“hard” subsystem at the beginning of the cycle. If the first hard failure takes place in the
inspection interval , and the first soft failure occurs before the hard failure, the soft failure is
rectified at the hard failure time and the time between the occurrence of the soft failure and hard
failure is considered as the downtime of the component with soft failure (Figure 5.13).
Figure 5.13. Hard and soft failures in the first inspection interval
Let for a given , Z1 min{ , Z } , where Z is the time of a hard failure. Therefore,
Z1 , and
CDF of Z1 is:
F Z ( z | s), z
F ( z | s)
Z
,
z
1
1,
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 147
s
H ( y ) dy
and P1Z ( | s ) P( Z1 | Z 0 s ) e s
Let X 1 min{ X , Z1} , where X is the time of the soft failure. Therefore, X 1 Z1 .
F X ( x), x z
F X1|Z1 ( x | z ) F X1 ( x | Z1 z ) ,
1, xz
tz
( y ) dy
P( X 1 Z1 | Z1 z , Z 0 s, X 0 t ) P ( X 1 Z1 | Z1 z , X 0 t ) P1 X ( z | t ) e t
.
random variables such as the total uptime of the component over the inspection interval, we
calculate the conditional expectation of the random variable given the times of the first hard
calculated/observed over an interval [0, ] . V can be the total uptime, number of minimal
repairs, etc. To calculate the expected value of V , we start with interval [0, ] , but after
conditioning on the first hard failure’ time , e.g. Z1 z , the interval for which the expected value
of V needs to be calculated is reduced to [ z , ] . Thus, the initial interval becomes shorter at the
The recursive equations for the total uptime, number of minimal repairs, etc. are identical
aside from a placeholder function ( X , Z ) , which differ according to the quantity we wish to
function ( X , Z ) given that the initial age of the component with soft failure is X 0 t , and the
Let, G1 ( , t , s) E[V | X 0 t , Z 0 s]
It can be shown, following equation (5.30), that (see Appendix L for the details)
z
G1 ( , t , s) { ( x, z ) G1 ( z, t x, s z)} f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z ) G1 ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz
0
. (5.31)
P1Z ( | s) ( x, ) f1X ( x | t )dx P1Z ( | s) P1X ( | t ) ( , )
0
When there are n inspection intervals of length over T , i.e. T n , we first calculate
G1 ( , t , s) using equation (5.31) for 0 t (n 1) , 0 s (n 1) which is the expected value
of variable V in the last inspection interval ( [(n 1) , T ] ) given t and s . In this case also
G1 ( , t , s) must be calculated for all . We then proceed to the second to the last inspection
interval (i.e. time (n 2) ), and we need to find G2 ( , t , s ) for 0 t (n 2) , 0 s (n 2)
and . G2 ( , t , s) gives the expected value of V over [(n 2) , T ] , so it includes also the
expected value calculated from G1 . In other words, the calculations are performed backward.
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 149
To generalize the equation for the interval [0, k ] , (k 1, 2,..., n) , we need to find
By taking similar steps as for G1 ( , t , s) , the following recursive equation can be derived
z
{ ( x, z ) Gk ( z, t x, s z )} f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z ) Gk ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz
0
P1Z ( | s) { ( x, ) Gk 1 ( , t x, s )} f1X ( x | t )dx
0
It should be noted that in practical applications, the last inspection interval may have a
0 t (n 1) , 0 s (n 1) and all 0 , which calculates E[V ] for the last inspection
interval , i.e. [(n 1) , (n 1) 0 ] , and eventually we calculate Gn ( , t , s ) for all , which
The general equation Gk ( , t , s ) can be considered for some special cases as follows. The
1, xz
( x, z ) ,
0, xz
z
M1 ( , t , s) {1 M1 ( z, t x, s z )} f1X ( x | t ) f1Z ( z | s )dxdz
0 0
M1 ( z, t z, s z ) f1Z ( z | s ) P1X ( z | t )dz P1Z ( | s ) f1X ( x | t )dx
0 0
( x, z ) x ,
z
e1 ( , t , s) {x e1 ( z, t x, s z )} f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{z e1 ( z, t z, s z )} f1Z ( z | s ) P1X ( z | t )dz
0
P ( | s) x f1X ( x | t )dx P1Z ( | s) P1X ( | t ).
1
Z
Thus,
m1 m2
E[CST ] ncI [c
j m1 1
M
j M nj ( , t , s) c Dj (T enj ( , t , s))] . (5.33)
Equation 5.33 should be calculated for different values of n to find the optimal inspection
interval, which gives the minimum expected cost (optimal) over the cycle T .
Numerical example
Consider a repairable system with eight components. Five components have soft failure,
and three components have hard failure. The parameters of the power law processes, and
minimal repair and downtime costs for components with soft and hard failures are given in Table
5.5. The system is currently inspected periodically every 4 months. We are interested in finding
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 151
the optimal inspection interval, which minimizes the total cost incurred from soft and hard
failures over a 12-month cycle length. For simplicity we assume what the equipment is as good
as new at the beginning of the cycle, so t 0 and s 0 are used in the calculation. The
Table 5.5. Parameters of the power law intensity functions, and costs for
components with soft and hard failure
k 1,..., n , recursively using equations (5.31) and (5.32). Direct calculation of M kj ( , t , s ) and
ekj ( , t , s ) would be tedious, except for small k . Therefore, Composite Simpson’s rule was
applied. To obtain an acceptable level of accuracy a very small step length h , such as h 105 or
less should be considered. This increases intensively the calculation time. So, instead of using
ekj ( , t , s ) for inspection intervals 1, 2,...,12 months. The expected values and their standard
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 152
errors were calculated for every value of using 1,000,000 simulated histories. The values of
0.0014, and 0.0011 were obtained as the average standard errors for the calculated expected
values M kj ( , t , s ) and ekj ( , t , s ) , respectively, for components with soft failure. The results of
The total expected costs for different inspection intervals over T 12 months were
obtained using equation (5.33) and are given in Table 5.6. Relatively similar estimated expected
costs obtained for 2,...,12 are due to the low minimal repair and downtime costs considered
Expected Cost
1 3681.05
2 3522.08
3 3509.65
4 3519.46
5 3526.75
6 3521.49
7 3521.60
8 3530.19
9 3526.87
10 3534.43
11 3543.85
12 3523.29
As shown, 3 is the optimal inspection interval with the expected cost of $3509.65,
without including the cost related to the hard failures ( E[CHT ] ). This means that the system
should be inspected four times in its life cycle of T 12 months, or that inspections should be
performed every 3 months. Increase of the cost curve at some points such as at 5 or 8
can be explained as follows: the system under the inspection policy of 5 is inspected
periodically 3 times during its life cycle, first time after 5 and the second time after 10 months,
and the last inspection after 2 months, i.e. at the end of the life cycle of 12 months. So we have
two intervals of length 5, and one of length 2 months. With 4 , we inspect 3 times, that is
every 4 months. In the longer inspection intervals of length 5 months compared to intervals of 4
months, it is more likely to have soft failures, and consequently penalty cost arising from their
downtimes. However, in the case of 5 , in the last inspection interval of 2 months, it is less
likely to have a failed soft component, but still the cost of an additional periodic inspection is
incurred. Thus, the total expected cost of 5 is greater than of 4 . A similar justification
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 154
can be given for the increase in the cost curve at 8 compared to 7 . A sharp drop in the
cost curve at 12 is due to the inspection cost, which is incurred only once, compared to
11 when it is incurred twice, but still with almost the same expected downtime and repair
costs.
We again consider a system with the assumptions stated in Section 4.3 and 5.3. In
addition, we assume that the components with hard failure is minimally repaired, and the
components with soft failures are either minimally repaired or replaced with new ones with some
age dependent probabilities. We assume that component j is minimally repaired at age x with
component’s age.
Additional Assumptions
Component with soft failures are either minimally repaired or replaced with some age
dependent probabilities.
Additional Notation
Penalty cost for the elapsed time of soft failure for component j ( c Dj )
Let I j denotes the status of component j at the current inspection (either periodic or
opportunistic):
placeholder function ( X , Z , I ) given the initial ages of the component with soft failure and the
G1 ( , t , s) E[V | X 0 t , Z 0 s] .
From the assumptions of NHPP and minimal repair or replacement, we get the recursive
G1 ( , t , s)
z
and
Gk ( , t , s )
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 156
z
{[ ( x, z,0) Gk ( z, t x, s z )]r (t x) [ ( x, z,1) Gk ( z,0, s z )]r (t x)} f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z, 2) Gk ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz
0
P1Z ( | s) {[ ( x, , 0) Gk 1 ( , t x, s )]r (t x) [ ( x, ,1) Gk 1 ( , 0, s )]r (t x)} f1X ( x | t )dx
0
P1Z ( | s ) P1 X ( | t ){ ( , , 2) Gk 1 ( , t , s )}
(5.35)
Then equations (5.34) and (5.35) can be applied to the special cases as follows:
1, x z and i 0
( x, z, i) ,
0, x z or i 0
z
M1 ( , t , s) {1 M1 ( z, t x, s z )}r (t x) f1X ( x | t ) f1Z ( z | s )dxdz
0 0
M1 ( z, t z , s z ) f1Z ( z | s) P1X ( z | t )dz P1Z ( | s ) r (t x) f1X ( x | t )dx
0 0
1, x z and i 1
( x, z, i) ,
0, x z or i 1
z
R1 ( , t , s) {1 R1 ( z, 0, s z )}r (t x) f1X ( x | t ) f1Z ( z | s)dxdz
0 0
R1 ( z, t z, s z ) f ( z | s) P1 ( z | t )dz P ( | s) r (t x) f1X ( x | t )dx
1
Z X
1
Z
0 0
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 157
( x, z, i) x ,
z
e1 ( , t , s ) {x r (t x)e1 ( z , t x, s z ) r (t x )e1 ( z , 0, s z )} f1 X ( x | t ) f1Z ( z | s )dxdz
0 0
{z e1 ( z, t z , s z )} f1Z ( z | s ) P1X ( z | t )dz
0
P1Z ( | s) x f1X ( x | t )dx P1Z ( | s) P1X ( | t ).
0
The total expected cost incurred over the cycle of length T with n scheduled inspections is
m1 m2
E[CST ] ncI [c
j m1 1
M
j M nj ( , t , s) c Rj Rnj ( , t , s) c Dj (T enj ( , t , s))] (5.36)
Numerical example
b j x
Assume that rj ( x) a j e , a j 0, b j 0, x 0 ; using the same parameters presented in
the numerical example of Section 5.3.1, with the replacement costs given in the second column
of Table 5.7, and the parameters a j and b j of rj ( x) , the expected cost curve is calculated using
simulation and is given in Figure 5.15 for 1, 2,...,12 . The estimated expected costs are
presented in Table 5.8. The expected number of minimal repairs and replacements, and the
expected uptimes and their standard errors for the components with soft failure for 1, 2,...,12
are given in Table M.2 of Appendix M. The values of 0.0010, 0.0007, and 0.0011 were obtained
as the average standard errors for the calculated expected values M kj ( , t , s ) , Rkj ( , t , s ) and
Table 5.7. Costs and parameters of the probability function for minimal repair of components
with soft failure
Table 5.8. Expected costs calculated for 1, 2,...,12 in Figures 5.15 and 5.16
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 159
As shown in Figure 5.15, 12 is the optimal inspection interval with the expected cost
of $7532.93. This means that the system should be inspected only once at the end of the life
cycle of length T 12 months. The optimal inspection interval varies by changing the
replacement costs. For example, by considering the replacement costs given in the third column
of Table 5.7, optimal 3 months is obtained with the expected cost of $3629.105 (Figure
5.16). The estimated expected costs are presented in the last column of Table 5.8.
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 160
Figure 5.16. Expected costs for different inspection intervals ( ), assuming soft
failures are either minimally repaired or replaced with different replacement costs
We consider a system with the assumptions stated in Section 4.3 and 5.3, additional
Additional Assumptions
Component with soft and hard failures are either minimally repaired or replaced with some
If we assume that a component with hard failure, depending on its age is either replaced
or minimally repaired at failure, then each hard failure component should be considered
individually in the model. When there is a hard failure in an interval and the component is
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 161
replaced, only the initial age of the failed component becomes zero and the age of the other
Let j ( z ) is the intensity function of the j th component with hard failure, and j is its
j ( z | j ) j ( j z ) .
Let J is the random variable which denotes the failed component with hard failure in
z Z z z is component j given initial ages of components with hard failure are 1 ,..., m . 1
So,
q j ( z ) lim P( J j | z Z z z, ) ,
z 0
Then
P( J j, z Z z z | ) j ( z | j )
q j ( z ) lim P( J j | z Z z z, ) lim ,
z 0 z 0 P( z Z z z | ) H ( z | )
and
z
j (z | j ) H ( x| ) dx
q j ( z )dF1Z ( z | ) H ( z | )e 0
dz
H ( z | ) .
z
H ( x| ) dx
j ( z | j )e 0
dz.
Let Z1 min{ , Z } , so Z1 denotes either the time to the first hard failure, if there is any
in [0, ] , or the inspection interval of length , if all components with hard failure survive. Let
I X and I Z be the indicators for survival or failure with minimal repair or replacement of the
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 162
component with soft and hard failures (component j ), respectively and r X ( x) and rjZ ( z) be their
Let Z01 1 , Z02 2 ,..., Z0m1 m1 , and G1 ( , t , 1 ,..., m1 ) E[V | X 0 t , (1 ,..., m1 ), U ] .
[ ( x, 0, z, j, 0) G1 ( z, t x, 1 z, 2 z,..., j z,..., m1 z )]r X (t x)
0 0 j 1
[ ( x,1, z, j,0) G1 ( z,0,1 z,2 z,..., j z,...,m1 z)]r X (t x) rjZ ( j z)
[ ( x,0, z, j,1) G1 ( z, t x,1 z,2 z,...,0,..., m1 z)]r X (t x)
[ ( x,1, j,1) G1 ( z, 0,1 z,2 z,..., 0,..., m1 z )]r X (t x) rjZ ( j z )
q j ( z) f1X ( x | t , z) f1Z ( z | (1,2 ,...,m1 ))dxdz
m1
{[ ( z, 2, z, j, 0) G1 ( z, t z,1 z, 2 z,..., j z,..., m1 z )]rjZ ( j z )
0 j 1
(5.37)
and for the general case, when the interval length is [0, (k 1) ] , (k 1, 2,..., n) :
z m1
[ ( x, 0, z, j, 0) Gk ( z, t x, 1 z, 2 z,..., j z,..., m1 z )]r X (t x)
0 0 j 1
[ ( x,1, z, j, 0) Gk ( z, 0, 1 z, 2 z,..., j z,..., m1 z )]r X (t x) rjZ ( j z )
[ ( x, 0, z , j ,1) Gk ( z , t x, 1 z , 2 z ,..., 0,..., m1 z )]r X (t x)
[ ( x,1, j ,1) Gk ( z , 0, 1 z , 2 z ,..., 0,..., m1 z )]r X (t x) rj Z ( j z )
q j ( z ) f1X ( x | t , z ) f1Z ( z | (1 , 2 ,..., m1 ))dxdz
m1
{[ ( z, 2, z, j, 0) Gk ( z, t z, 1 z, 2 z,..., j z,..., m1 z )]rjZ ( j z )
0 j 1
(5.38)
components with hard failure has failed. Equations (5.37) and (5.38) can be applied to the special
cases as follows:
1, iX 0
( x, i X , z , j , i Z ) ,
0, iX 0
M 1 ( , t , 1 , 2 ,..., m1 )
z m1
[1 M1 ( z, t x,1 z, 2 z,..., j z,..., m1 z )]r X (t x)
0 0 j 1
M1 ( z,0,1 z,2 z,..., j z,...,m1 z)r X (t x) rjZ ( j z)
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 164
[1 M1 ( z, t x,1 z,2 z,...,0,...,m1 z)]r X (t x)
M1 ( z, 0,1 z,2 z,..., 0,..., m1 z )r X (t x) rjZ ( j z )
q j ( z) f1X ( x | t , z) f1Z ( z | (1,2 ,...,m1 ))dxdz
m1
{M1 ( z, t z,1 z, 2 z,..., j z,..., m1 z )rjZ ( j z)
0 j 1
1, iX 1
( x, i X , z , j , i Z ) ,
0, iX 1
R1 ( , t , 1 , 2 ,..., m1 )
z m1
R1 ( z, t x,1 z, 2 z,..., j z,..., m1 z )r X (t x)
0 0 j 1
[1 R1 ( z, 0,1 z,2 z,..., j z,..., m1 z )]r X (t x) rjZ ( j z )
R1 ( z, t x,1 z, 2 z,..., 0,..., m1 z )r X (t x)
[1 R1 ( z, 0, 1 z, 2 z,..., 0,..., m1 z )]r X (t x) rj Z ( j z )
q j ( z) f1X ( x | t , z) f1Z ( z | (1,2 ,...,m1 ))dxdz
m1
{R1 ( z, t z,1 z,2 z,..., j z,...,m1 z )rjZ ( j z )
0 j 1
r X (t x) f1X ( x | t , )dxP1Z ( | (1 , 2 ,..., m1 ))
0
( x, i X , z , j , i Z ) x
E1 ( , t , 1 , 2 ,..., m1 )
z m1
[ x e1 ( z, t x,1 z,2 z,..., j z,..., m1 z )]r X (t x)
0 0 j 1
m1 m2
E[CST ] ncI [c
j m1 1
M
j M nj ( , t , 1 , 2 ,..., m1 ) c Rj Rnj ( , t, 1, 2 ,..., m1 ) c Dj (T enj ( , t, 1, 2 ,..., m1 ))]
(5.39)
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 166
Numerical example
Using the parameters given in Tables 5.9 and 5.10, and the inspection cost CI $70 , the
expected cost curve is calculated from equation (5.39) and is given in Figure 5.17 for
1, 2,...,12 .
Table 5.9. Parameters of the power law intensity functions, and costs for
components with soft and hard failure
Component aj bj
1 0.9 0.2317
2 0.9 0.1763
Soft Failure 3 0.9 0.1352
4 0.9 0.0811
5 0.9 0.2253
1 0.8 0.06
Hard Failure 2 0.5 0.02
3 0.9 0.28
As it is shown on Figure 5.17, the optimal value 12 is obtained with the total cost of
$7570.638. If replacement costs given in Table 5.11 are considered, the optimal value 2
months with the total cost of $4317.506 will be obtained (Figure 5.18).
1 233.33
2 150.00
Soft Failure 3 333.33
4 250.00
5 500.00
Table 5.11. Replacement costs considered in Figure 5.18
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 168
Figure 5.18. Expected costs for inspection intervals ( ), assuming soft and
hard failures are either minimally repaired or replaced with the replacement
costs given in Table 5.11
We consider again a system with the assumptions stated in Section 5.3.3, i.e. both
components with soft and hard failures are either minimally repaired or replaced with some age
dependent probabilities. In addition, we assume that the components with hard failures can be
preventively replaced at periodic inspections, so the most advanced case is considered (see the
structure shown in Figure 5.1). The preventive replacement of a component with hard failure still
depends on its age. In other words, the component may be replaced at a scheduled inspection or
Additional Assumptions
Component with soft and hard failures are either minimally repaired or replaced with some
Components with hard failure are preventively replaced at periodic inspections depending on
Additional Notation
Probability that the component with soft failure is minimally repaired at age x ( r X ( x) )
Probability that component j with hard failure is minimally repaired at age z ( rjZ ( z ) )
Components with hard failures are preventively replaced at scheduled inspections with
some age dependent probabilities with a lower cost, or they are repaired or replaced at failures
(non-scheduled) with a higher cost. Therefore, we need to calculate the expected number of
minimal repairs and replacements of hard components at both scheduled and non-scheduled
inspections.
m1 m2
E[C T ] ncI
j m1 1
[c Mj M nj ( , t , 1 , 2 ,..., m1 ) c Rj Rnj ( , t , 1 , 2 ,..., m1 ) c Dj (T enj ( , t , 1, 2 ,..., m1 ))]
m1
[c Mj M nHj ( , j ) c Rj RnHj ( , j )]
j 1
m1
+ c PR
j PRn ( , j ),
j
j 1
(5.40)
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 170
m1
where [c
j 1
M
j M nH j ( , j ) c Rj R nH j ( , j )]is the total costs of minimal repairs and replacements
of the components with hard failures inside inspection intervals (non-scheduled inspections), and
m1
c
j 1
PR
j PRnj ( , j ) is the replacement costs of hard failures at scheduled inspections.
To calculate the required variables used in E[C T ] , i.e. M nj , Rnj , enj , M nHj , RnHj , and PRnj ,
similarly to the methods described in the previous sections, we derive general recursive
case, function G1 ( , t , 1 , 2 ,..., m1 ) will be identical to equation (5.37). For the general case,
when the interval length is [0, (k 1) ] , (k 1, 2,..., n) a more complicated equation is
derived as follows:
Gk ( , t , 1 , 2 ,..., m1 )
z m1
[ ( x, 0, z, j, 0) Gk ( z, t x, 1 z, 2 z,..., j z,..., m1 z )]r X (t x)
0 0 j 1
[ ( x,1, z, j,0) Gk ( z,0,1 z,2 z,..., j z,...,m1 z)]r X (t x) rjZ ( j z)
[ ( x,0, z, j,1) Gk ( z, t x,1 z,2 z,...,0,...,m1 z)]r X (t x)
[ ( x,1, j,1) Gk ( z, 0, 1 z, 2 z,..., 0,..., m1 z)]r X (t x) rj Z ( j z)
q j ( z) f1X ( x | t , z) f1Z ( z | (1,2 ,...,m1 ))dxdz
m1
{[ ( z, 2, z, j, 0) Gk ( z, t z, 1 z, 2 z,..., j z,..., m1 z )]rjZ ( j z )
0 j 1
( x, 0, , 0, 2) Gk 1 ( , t x,1 (1 ),...,m ( m1 )) P(V1 1 ,..., Vm1 m1 ) r X (t x)
0
1
1 ,...,m
1
( x,1, , 0, 2) Gk 1 ( , 0,1 (1 ),...,m ( m1 )) P(V1 1 ,..., Vm1 m1 ) r X (t x)
1
1 ,...,m
1
f1X ( x | t, z)dxP1Z ( | (1,2 ,...,m1 ))
( , 2, , 0, 2) Gk 1 ( , t x,1 (1 ),...,m ( m1 )) P(V1 1,..., Vm1 m1 )
1
1 ,...,m
1
P1X ( | t )P1Z ( | (1,2 ,...,m1 )) ,
(5.41)
where
and
Z
ri ( zi ), minimal repair of component i
P(Vi i ) Z
,
ri ( zi ), replacement of component i
or
Also,
G1 and Gk can be simplified as G1H and GkH for calculating the expected values inside
inspection intervals, and at the inspections for component j with hard failure. Let r Z rjZ and
j , then
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 172
G1H ( , )
(5.42)
and
GkH ( , )
{[ ( z,0; ) GkH ( z, z )]r Z ( z ) [ ( z,1; ) GkH ( z,0)]r Z ( z )} f1Z ( z | )dz
0
(5.43)
Variables M1 , R1 , and e1 in equation (5.40) can be calculated from the equations given in
Section 5.3.3, the other required variables M 1H , R1H , and PR1 are the special cases of equation
(5.42):
(a) For the expected number of minimal repairs of the component in [0, ] :
1, z ,i 0
( z, i; ) ,
0, otherwise
M ( , s) {[1 M1H ( z, s z )]r Z ( s z ) M1H ( z, 0)r Z ( s z )} f1Z ( z | s)dz
1
H
1, z ,i 1
( z, i; ) ,
0, otherwise
R ( , s) {R1H ( z, s z )r Z (s z ) [1 R1H ( z, 0)]r Z (s z )} f1Z ( z | s)dz
1
H
0
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 173
1, z ,i 1
( z, i; ) ,
0, otherwise
PR1 ( , s)
P1Z ( | s)r Z ( s )
Numerical example
We use the power law parameters, the costs of components with soft failure, and
parameters a j and b j given in Tables 5.9 and Table 5.10. The costs of components with hard
failure are given in Table 5.12. Inspection cost CI $70 is considered. We developed a
simulation model to estimate the required expected values for inspection intervals 1, 2,...,12
months. The expected cost curve is calculated using equation (5.40) and is shown in Figure 5.19
Table 5.12. The costs for different components with hard failure
5.3. Case 2: A System Under Periodic and Opportunistic Inspections 174
With different replacement costs c Rj , given in the last column of Table 5.12 for the
components with hard failure, the optimal 3 is obtained with cost= $33959.92.
failure, hard and soft. So-called hard failures are revealed, i.e., either they are self-announcing or
the system stops when they take place. However, the system still can continue functioning when
soft failures occur. Both hard and soft failures should be incorporated in the general formulation
of an inspection optimization model for a repairable system. Soft and hard failures are important
in terms of the consequences which may result. Power plants, hospitals or any other organization
for which safety of instruments is a vital issue have to consider carefully all outcomes of their
systems’ failures. Unrevealed or soft failures may even result in a catastrophe if unattended.
This chapter proposes several inspection optimization models for a periodically inspected
repairable system under different assumptions. Two main cases of non-opportunistic and
opportunistic maintenance are considered. In all models, it is assumed that hard failures are
instantly rectified, and the failed component is inspected (non-scheduled) and fixed.
In the models proposed for the case of non-opportunistic maintenance, soft failures are
only rectified at periodic inspections. Therefore, occurrence of hard failures does not affect the
optimal periodic inspection interval, i.e. hard failures do not incorporate in the optimization
models. The optimal inspection interval minimizes the expected cost incurred only from the soft
failures.
We first assume that soft failures are minimally repaired, and develop two variants of the
model under this assumption. The first variant takes into account the elapsed times from soft
failures by considering downtime penalty costs for each component. The second variant sets up a
threshold for the total number of soft failures and considers penalty costs for exceeding the
threshold and for the number of failures, which exceed the threshold. A combined model is also
5.4. Concluding Remarks 176
proposed to incorporate both the threshold on number of soft failures and the elapsed times of
failures. Due to the assumption of minimal repair, a non-homogeneous Poisson process is used as
the model for failure processes. Since the times of soft failures are not known (they are interval
probabilities of failures and expected downtimes, which makes calculation of the optimum
inspection frequency computationally intensive. The procedure allows for considering initial age
of each component not necessarily equal to zero at the beginning of the cycle, and therefore,
possibility to perform incomplete repair at the beginning of the next cycle. In other words, some
components can be renewed, while the ages of others can be reduced. Then, the optimal
inspection frequency can be calculated for the next cycle. The recursive procedure that deals
with downtimes and minimal repairs in a NHPP is the main contribution of the proposed models.
Secondly, we extend the repair assumption of the soft failures to include both possibility
probabilities. Under this assumption, two models are proposed for finding the optimal inspection
interval, one over finite and the other over infinite time horizon. The model over a finite time
horizon assumes that at an inspection a failed component may be replaced or minimally repaired.
The probability of minimal repair is a function of the component’s age and is assumed to be
decreasing by age. A penalty cost is incurred for the elapsed time from a failure to its detection
at the next inspection. In this model, we find the optimal inspection interval with minimum
expected cost. The second model uses the renewal-reward theorem (Tijms, 2003) and formulates
the objective function as the expected cost per unit time for the system, which is equal to the sum
of the expected costs per unit time of all components. A recursive procedure is developed to
calculate probabilities of failures, expected number of minimal repairs, and expected downtimes.
5.4. Concluding Remarks 177
Then, the case of opportunistic maintenance is considered. It is assumed that hard failures
are detected and fixed instantaneously; however, soft failures are rectified at periodic or
opportunistic inspections. An opportunistic inspection is the moment of a hard failure, when all
In the opportunistic case, it is first assumed that hard failures are only minimally repaired.
Two models are proposed; the first model assumes the minimal repair of the components with
soft failures. The second model is an extension of the first model and considers the possibility of
minimal repair and replacement of components with soft failures, depending on some age
dependent probabilities. In the proposed models, the repair cost of a hard failure and the
opportunistic inspection cost do not affect the optimal periodic inspection interval. However, the
occurrences of hard failures influence the detection of soft failures and then their expected
Moreover, the proposed models are extended to the case when components with hard
failures are also either minimally repaired or replaced. In addition, preventive replacement of
components with hard failures at periodic inspections is also included in the model. In the latter,
both repair/replacement cost of a hard failure and opportunistic inspection cost influence the
The calculation of the excepted cost with delayed replacement or minimal repair of a
Due to involved numerical calculation, simulation is used instead to calculate the cost function
6.1. Conclusion
Hospitals’ equipment management program can be significantly improved by proper
fairly planned and executed, can be cost-effective, efficient and can assure reliability of
In spite of the fact that reliability engineering techniques and maintenance optimization
models are well-established and used in other industries, their application to the medical field is
fairly new. Most research mainly proposes methods to assess or improve the reliability of
devices in their operating context has not been much considered. In the research reported in this
industry. We looked into the challenges that clinical engineers are currently facing, in particular,
for maintenance management of medical equipment. We attempted to propose some models and
methods that can be practically used. We constructed the models after taking into
178
6.1. Conclusion 179
account the actual procedures used for maintaining medical devices in hospitals, and considering
model for prioritization of medical devices according to their criticality. The model can be
integrated into the CMMS of a hospital to prioritize medical devices. Devices with lower
criticality scores are excluded from the hospital’s MEMP. Devices with higher scores should be
investigated in detail to find the main causes of having such high criticality scores. Appropriate
actions, such as “preventive maintenance”, “user training”, “redesigning the device”, etc. should
Technically, Analytical Hierarchy Process was used to identify and decide on inclusion
of critical devices in the hospital’s equipment management program. The proposed model takes
six criteria, “Function”, “Mission criticality”, “Age”, “Risk”, “Recalls and hazard alerts”, and
“Maintenance requirements”, into account to evaluate the criticality of devices. The model
realistically estimates the total risk of a device by considering its individual failure modes. A risk
score is assigned to each failure mode after evaluating its frequency, detectability and
consequences. Pairwise comparison of the criteria was performed, and the consistency ratio was
calculated to produce more precise and consistent criteria’s weights. The relative measurement
method was applied to find more consistent values for the intensities of the criteria’s grades.
strategy for a device according to its individual score with respect to a criterion or a combination
of several criteria.
criteria to assess the criticality of a device. Moreover, it is more consistent in terms of criteria’
6.1. Conclusion 180
weights and the intensities of the criteria’s grades. Acceptable and consistent results were
On the other hand, the model requires involvement of experts in the process of applying
the model to medical devices, which might be expert intensive. In addition, the prioritization
results may not always be acceptable by clinical engineers and requires reassigning the criteria’s
(b) Field maintenance data- Then, we investigated the maintenance data from a hospital.
We categorized the failure types into hard and soft failures, and proposed policies to deal with
the censored and missing data at different levels of a device. Hard failures are self-announcing or
they make the system stop functioning. They are rectified and fixed immediately. However, the
system can still work in presence of soft failures, even though its performance may be reduced.
Soft failures are only rectified at inspections, so their failure times are not known. We applied the
proposed policies to general infusion pumps, and conducted reliability and trend analysis of the
data. For the trend analysis, midpoints of the censoring intervals for failures were calculated and
the Laplace trend analysis was applied. The results revealed that it is not uncommon to have
decreasing or increasing trend in the failure process of devices, mainly due to aging of the
mechanical parts of the device. This conclusion opposes the common belief of healthcare
societies that medical devices fail independent of age, following exponential distribution.
Obviously, more precise and reliable results can be obtained when more precise failure data is
available. Medical devices tend to be highly reliable; therefore, scarce failure data is always a
Our approach in using the midpoints of censoring intervals may not always provide
precise results, particularly, when censoring intervals are long. We opted to use it because no
6.1. Conclusion 181
available trend test analysis in the literature incorporates interval censored data if repairs are not
instantaneous.
(c) Trend test- We then proposed a method for a system whose failure process follows a
non-homogeneous Poisson process (NHPP) with a power law intensity function. The assumption
of NHPP is common for the failure process of complex systems such as medical equipment. We
used the likelihood ratio test to check for trends in the failure data with censoring. The EM
algorithm and one of its modifications were used to find the parameters, which maximize the
data likelihood in the case of no trend and trend hypotheses. We developed a recursive procedure
to calculate the expected values in the E-step requiring numerical integration. The proposed
method was applied for several components of general infusion pump. We compared the results
of the EM algorithm, its modification, and the ad-hoc method of replacing unknown failure times
with the mid points of censoring intervals. We concluded that the parameters estimates obtained
from the complete EM and its modified version are very close. Moreover, these parameters
might be close to what is obtained from the mid-points method, if the censoring intervals are
short, but not always. At least, the parameters obtained from the mid-points method can be used
(d) Inspection interval- Moreover, we developed several optimization models to find the
optimal periodic inspection interval for a system with hard and soft failures. We assumed that the
system undergoes periodic inspections due to its simple implementation, especially for a large
number of devices in a hospital. We first assumed that the soft failures can only be rectified at
periodic (scheduled) inspections. Several models were constructed under this assumption. Then,
we considered also the possibility of opportunistic inspection of components with soft failure, at
non-scheduled inspections arising from the occurrence of hard failures. In addition, we extended
6.1. Conclusion 182
the models to the case when the components with hard failure are preventively replaced at
In all models, the optimal inspection interval is found by minimizing the total expected
cost. We solved the main technical problem of calculating the excepted cost with delayed
models, we recursively calculated the contributed variables in the cost model, such as expected
downtime, expected number of minimal repairs, etc. However, for more complicated models,
such as opportunistic maintenance models, simulation was used due to involved numerical
calculation. We obtained reliable results from both recursive calculation and simulation,
although the former might be more time consuming when optimization is performed for a system
with a large number of components, or for a longer life cycle with more periodic inspections.
The optimal inspection interval depends significantly on the input cost parameters, such
as inspection cost, repair cost, downtime penalty cost, etc. Changing one or some of the cost
(e) Remarks- As the final remarks in this section, we emphasize that the optimizing
maintenance strategies for medical equipment has not been much discussed in the literature.
Hospitals must assure that they are in compliance with the risk and maintenance standards for
medical equipment. Therefore, some hospitals include all medical devices in the MEMP, and just
follow manufacturers’ maintenance recommendations to avoid liability. In other words, there has
not been much tendency to explore new methods to improve the maintenance procedures in
applied properly, not only can assure the safety of devices, but also can offer more efficient and
(f) devices, especially in their operating context is fairly new, which requires much
more work to be done in the future. In the next section, we discuss some new directions, which
are more realistic or taking into account new considerations. Moreover, other techniques and
methods can be investigated and employed to improve the precision and calculation time of the
that the score value assigned to a device with respect to a criterion, or a combination of several
criteria may be used to decide on appropriate maintenance strategy for the device. We presented
some initial heuristic guidelines in Table 3.7. However, more research is required to improve this
In addition, using Saaty’s crisp 1-9 scales, we constructed the pairwise comparison
matrices for the criteria and their intensities. Alternatively, fuzzy analytic hierarchy process
(Kwong and Bai, 2002) can be used, in which crisp values are substituted with fuzzy numbers for
scales. Thus, expert judgment is performed more confidently in comparing the criteria, especially
(b) Trend test- Furthermore, in the proposed model for trend analysis of the failure data
discussed in Section 4.4 of Chapter 4, other iterative methods such as Simplex (Lagarias et al.,
1998) can be employed instead of the EM algorithm. Simplex is a direct search method, which
does not require numerical or analytic gradients. In addition, Newton-Raphson (McLachlan and
6.2. Future Research 184
Krishnan, 1996) can be used in both E-step and M-step of the EM algorithm. In this case, the
convergence of the parameters is achieved with less number of iterations; on the other hand,
Other methods, such as Monte-Carlo simulation can be also used to estimate the
expected values required in the E-step of the EM (Levine and Casella, 2001). In this case, no
recursive procedure is required. Research can be conducted to compare all these numerical
algorithms, in terms of precision, number of iterations, and convergence time. In addition, failure
histories can be simulated to evaluate the performance of these algorithms for a larger number of
(c) Inspection interval- Finally, the inspection optimization models proposed in Chapter 5
inspections. Furthermore, the current models do not optimize any maintenance decision at
inspection for a component or the system. Maintenance decision can be also optimized in the
models to decide which components should be replaced at inspections. Another extension of the
current models is to assume that a soft failure or a combination of soft failures may eventually
convert to a hard failure, if left unattended. Moreover, different types of dependency can be also
considered for the system’s components which makes the optimization models much more
complicated, but more realistic. In this case, both maintained data and experts’ knowledge should
be used to incorporate the components’ structure. The CMMS of most hospitals keeps track of
the failures and maintenance actions at the component level, so this data can be used for
[3] Anderson T W and Darling D A (1952). Asymptotic theory of certain 'goodness of fit'
[4] Ansell J I and Phillips M J (1994). Practical Methods for Reliability Data Analysis.
[5] Ascher H E and Feingold H (1984). Repairable Systems Reliability: Modeling, Inference,
[6] Atles L R (2008). A Practicum for Biomedical Technology & Management Issues.
[7] Bagadia K (2006). Computerized Maintenance Management Systems Made Easy: How to
185
7. REFERENCES 186
[8] Baker R (2001). Data-Based Modeling of the Failure Rate of Repairable Equipment.
[9] Baretich M (2004). Equipment Control and Asset Management. In: Dyro J, editor.
[10] Barlow R E and Heidtmann K D (1984). Computing k-out-of-n system reliability. IEEE
[13] Berg M and Cleroux R (1982). The Block Replacement Problem with Minimal Repair
and Random Repair Costs. Journal of Statistical Computation and Simulation 15: 1-7.
[14] Bevilacquaa M and Braglia M (2000). The analytic hierarchy process applied to
maintenance strategy selection. Reliability Engineering and System Safety 70 (1): 71–83.
[15] Block H W, Borges W S and Savits T H (1985). Age dependent minimal repair. Journal
[16] Booth J G, Hobert J P and Jank W (2001). A survey of Monte Carlo algorithms for
333-349.
[17] Boland P J (1982). Periodic replacement when minimal repair costs vary with time.
[18] Boland P J and Proschan F (1982). Periodic replacement with increasing minimal repair
[19] Brown M and Proschan F (1983). Imperfect repair. Journal of Applied Probability 20:
851–859.
Countries. In: Dyro J, editor. Clinical Engineering Handbook, Elsevier: San Diego, p.
133-134.
[22] Cheng M (2004b). Single-Use Injection Devices. In: Dyro J, editor. Clinical Engineering
[23] Cheng M and Dyro J F (2004). Good Management Practice for Medical Equipment. In:
[24] Chien Y H and Sheu S H (2006). Extended optimal age-replacement policy with minimal
169–181.
[25] Chien Y H, Sheu S H and Chang C C (2009). Optimal age-replacement time with
minimal repair based on cumulative repair cost limit and random lead time. International
[26] Choi J G and Seong P H (2008). Reliability of Electronic Components. Reliability and
Risk Issues in Large Scale Safety-critical Digital Control Systems, Ch. 1, Seong P H
[27] Christer A H (2002). A Review of delay Time analysis for Modeling Plant Maintenance.
In: Osaki S, editor. Stochastic Models in Reliability and Maintenance, Springer: London,
p. 90-117.
7. REFERENCES 188
[28] Cohen T (2008). Computerized Maintenance Management Systems. In: Atles L R, editor.
[29] Cohen T and Cram N (2004). Computerized Maintenance Management Systems. In: Dyro
[30] Cox D R and Lewis P A W (1996). The Statistical Analysis of Series of Events. Chapman
[32] Crow L H (1974). Reliability analysis for complex repairable systems. In: Proschan F,
Serfing Rj. eds. Reliability and Biometry: Statistical Analysis of Life lengths. SIAM:
[33] Crow L H (1982). Confidence interval procedures for the Weibull process with
[35] Cui L, Xie M and Loh H (2004). Inspection schemes for general systems. IIE
[36] Cui L (2005). Availability of a periodically inspected system with random repair or
[37] Dekker R, Kleijn M J and De Rooij P J (1998). A spare parts stocking policy based on
[38] DeRosier J, Stalhandske E, Bagian J P and Nudell T (2002). Using health care failure
mode and effect analysis: the VA national center for patient safety’s prospective risk
analysis system. Joint Commission Journal on Quality and Patient Safety 28(5): 248-267.
[39] Dhillion B S (2000). Medical Device Reliability and Associated Areas. CRC Press: Boca
Raton, FL.
[40] Dohi T, Ashioka A and Osaki S (2001a). Optimizing the repair-time limit replacement
[41] Dohi T, Kaio N and Osaki S (2001b). Optimal periodic maintenance strategy under an
[42] Dohi T, Kaio N and Osaki S (2003). Preventive Maintenance Models: Replacement,
[43] Durkin J (1994). Expert Systems Design and Development. Macmillan: New York, NY.
[45] Engelhardt M E, Guffey J M and Wright F T (1990). Tests for Positive Jumps in the
356-360.
[46] Fennigkoh L and Smith B (1989). Clinical equipment management. JCAHO Plant,
[47] Fong S W and Choi S K Y (2000). Final contractor selection using the analytical
[48] Fries R C (2005). Reliable Design of Medical Devices. CRC: Boca Raton.
7. REFERENCES 190
[51] Gulati R and Smith R (2009). Maintenance and Reliability Best Practices, Industrial
[52] Harding G H and Epstein A L (2004a). Technology Evaluation. In: Dyro J, editor.
[53] Harding G H and Epstein A L (2004b). Technology Procurement. In: Dyro J, editor.
[54] Health Canada (1998). Therapeutic Products Directorate, Guidance for the risk-based
[55] Herath G and Prato T (eds.) (2006). Using Multi-Criteria Decision Analysis in Natural
[56] Hartman J C and Murphy A (2006). Finite-horizon equipment replacement analysis. IIE
[57] Ho W (2008). Integrated analytic hierarchy process and its applications - a literature
[58] Hyman W (2003). The Theory and Practice of Preventive Maintenance. Journal of
[59] Ion R A, Sonnemans P J M and Wensing T (2006). Reliability Prediction for Complex
677-680.
[62] Jiang R and Jardine A K S (2005). Two optimization models of the optimum inspection
[63] Johnston G (assessed on Feb 18, 2011). Reliability Data Analysis in the SAS System.
http://support.sas.com/rnd/app/papers/reliability.pdf.
Terrace, IL.
Failure mode and effects analysis in health care: Proactive risk reduction. 2nd ed. Joint
Terrace, IL.
[67] Joint Commission Accreditation Ambulatory Care (2010). Accreditation Handbook for
http://www.jointcommission.org/assets/1/6/2010IMGHdbk.pdf.
[69] Kelley T (2003). Solving Nonlinear Equations with Newton’s Method, Fundamental
[70] Kijima M (1989). Some Results for Repairable Systems with General Repair. Journal of
[71] Kvaløy J Y and Lindqvist B H (1998). TTT-based tests for trend in repairable systems
[72] Kvaloy J T, Lindqvist B and Malmedal H (2002). A statistical test for monotonic and
non-monotonic trend in repairable systems. In: Proc European Conference on Safety and
[73] Kwong C K and Bai H (2002). A fuzzy AHP approach to the determination of
[74] Labib AW, Williams G B and O’Conner R F (1998). An intelligent maintenance model
(1): 112-147.
[76] Laggoune R, Chateauneuf A and Aissani D (2009). Opportunistic Policy for Optimal
[77] Lai M T (2007). A periodical replacement model based on cumulative repair-cost limit.
[78] Lam Y (2009). A Geometric Process d-Shock Maintenance Model. IEEE Transactions on
[79] Lange K (1995). A gradient algorithm locally equivalent to the EM algorithm. Journal of
and time trend, with application to repairable systems. Technometrics 38: 131-138.
[81] Leung L C and Cao D (2001). On the efficacy of modeling multi-attribute decision
problems using AHP and Sinarchy. European Journal of Operational Research 132: 39-
49.
[82] Levine R A and Casella G (2001). Implementations of the Monte Carlo EM algorithm.
[84] Lewis P A W and Robinson D W (1974). Testing for a monotone trend in a modulated
renewal process. In: Proschan F, Serfing Rj. eds. Reliability and Biometry: Statistical
[85] Li W and Pham H (2005). An inspection-maintenance model for systems with multiple
[86] Lienhard B, Hugues E, Bes C and Noll D (2008). Failure-Finding Frequency for a
29–37.
[89] Makis V and Jardine A K S (1992). Optimal replacement policy for a general model with
[90] Mann H B (1945). Nonparametric tests against trend. Econometrica 13: 245-259.
[91] McCauley J (2004). Maintenance and Repair of Medical Devices. In: Dyro J, editor.
[92] Mckay N L (1986). Industry Effects of Medical Device Regulation: The Case of
Diagnostic Imaging Equipment. Journal of Policy Analysis and Management 6 (1): 35-
44.
[93] McLachlan G J and Krishnan T (1996). The EM Algorithm and Extensions. Wiley: New
York.
[94] Meados M (2006). The FDA and Product Recalls. FDA Consumer 40(3): 30-35.
[95] Meeker W and Escobar L (1998). Statistical Methodology for Reliability Data. Wiley:
New York.
Defense; 1981.
[97] Modarres M (2006). Risk analysis in engineering: techniques, tools, and trends. CRC
Heinemann: England.
7. REFERENCES 195
Symposium 54-59.
[100] Moubray J (1997). Reliability-centered Maintenance II. 2nd ed. Industrial Press: New
York, NY.
[101] Murthy D N P and Nguyen D G (1981). Optimal age-policy with imperfect preventive
[102] Nakagawa T (2008). Advanced Reliability Models and Maintenance Policies. Springer:
London.
[104] Nakagawa T (2003). Maintenance and optimum policy. In: Pham H, editor. Handbook of
[105] Nakagawa T and Mizutani S (2009). A Summary of maintenance policies for a finite
[106] Nakagawa T and Yasui K (1987). Optimal policies for a system with imperfect
378.
[108] Nelson W (1998). An application of graphical analysis of repair data. Quality and
http://www.itl.nist.gov/div898/handbook/.
[111] Partovi F Y, Burton J, and Banerjee A (1989). Application of Analytic Hierarchy Process
[112] Pham H (2003). Optimal Imperfect Maintenance Models. In: Pham H, editor. Handbook
[113] Pham H (2008). Recent Advances in Reliability and Quality in Design. Springer: London.
[114] Pham H and Wang H (1996). Imperfect maintenance. European Journal of Operational
[115] Puran L (1990). Mil-Hdbk-217: What is Wrong with It. IEEE transactions on
[116] Qian C, Nakamura S and Nakagawa T (2003). Replacement and minimal repair policies
for a cumulative damage model with maintenance. Computers and Mathematics with
[118] Rice W (2007). Medical Device Risk Based Evaluation and Maintenance Using Fault
http://alexandria.tue.nl/extra2/afstversl/tm/roelfsema2004.pdf.
[122] Ross S M (2007). Introduction to Probability Models. 9th edition. Academic Press: New
York.
Heinemann: England.
http://www.wbdg.org/om/cmms.php.
[125] Saaty T L (1980). The Analytic Hierarchy Process. McGraw Hill: New York, NY.
[126] Saaty T L (1986). Absolute and relative measurement with the AHP. The most livable
[128] Saaty T L (1990), How to make a decision: The analytic Hierarchy Process. European
[129] Saaty T L (2008). Decision making with the Analytic Hierarchy Process. International
[130] Samuel F E Jr (1991). Safe Medical Devices Act of 1990. Health Affairs 10 (1): 192-195.
7. REFERENCES 198
[131] Sherif Y S and Smith M L (1981). Optimal Maintenance Models for Systems Subject to
(2): 397–405.
[133] Sheu S H and Griffith W S (2001). Optimal age-replacement policy with age-dependent
[134] Sheu S H, Lin Y B and Liao G L (2005). Optimal policies with decreasing probability of
[135] Sheu S H, Lin Y B and Liao G L (2006). Optimum policies for a system with general
[137] Stiefel R H (2009). Medical Equipment Management Manual. 7th ed. AAMI: Annapolis
Junction, MD.
[139] Taghipour S and Banjevic D (2010a). Trend Analysis of the Power Law Process with
Censored Data Using EM Algorithm. Reliability Engineering and System Safety (under
[140] Taghipour S and Banjevic D (2011a). Trend Analysis of the Power Law Process with
2011.
7. REFERENCES 199
[141] Taghipour S and Banjevic D (2011b). Periodic Inspection Optimization Models for a
[142] Taghipour S and Banjevic D (2011c). Optimum Inspection Interval for a System Under
Model for a Complex Repairable System. Reliability Engineering and System Safety
95(9): 944-952.
for a System subject to Hidden Failures. Proceedings of the 2010 ICOMS Asset
Management Conference.
Data for Complex Medical Devices. Quality and Reliability Engineering International
27(1): 71-84.
[148] Tijms H C (2003). A first Course in Stochastic Models, 2d ed. John Wiley & Sons:
Chichester.
7. REFERENCES 200
[149] Taylor K (2005). A Medical Equipment Replacement Score System. Journal of Clinical
[151] Trindade D and Nathan S (2008). Field Data Analysis for Repairable Systems: Status and
http://www.aami.org/tmcconnect/exchange/department/dept.eqtplan.doc.
http://www.fda.gov/cdrh/prodcode.html.
[156] Vaurio J K (1995). Optimization of test and maintenance intervals based on risk and cost,
[157] Vaurio J K (2003). Common cause failure probabilities in standby safety system fault tree
[158] Vaurio J K (2005). Uncertainties and quantifications of common cause failure rates and
probabilities for system analyses. Reliability Engineering and System Safety 90: 186–195.
7. REFERENCES 201
[159] Vellani K H (2006). Strategic Security Management: A Risk Assessment Guide for
[160] Wang B (2006). Fennigkoh and Smith Model for Inclusion Criteria: 15-Year
[161] Wang B, Furst E, Cohen T, Keil O R, Ridgway M and Stiefel R (2006). Medical
233-237.
[162] Wang B and Levenson A (2000). Equipment Inclusion Criteria- a New Interpretation of
25(1): 26-35.
[163] Wang B and Rice W (2003). JCAHO’s Equipment Inclusion Criteria Revisited-
37-48.
[165] Wang H and Pham H (2006). Reliability and optimal maintenance. Springer: London.
[166] Wang P and Coit D W (2005). Repairable Systems Reliability Trend Tests and
421.
[167] Wang W (2009). An inspection model for a process with two types of inspections and
[168] Wang W and Christer A H (2003). Solution algorithms for a nonhomogeneous multi-
[169] World Health Organization (1998). District Health Facilities, Guidelines for
Development & Operations. WHO Regional Publications, Western Pacific Series, No.
22.
[170] Yoon K P and Hwang C L (1981). Multi Attribute Decision Making. Springer: New
York, NY.
system with failure dependencies. Reliability Engineering and System Safety 92: 1627–
1634.
[172] Zhou X, Xi L and Lee J (2006). A dynamic opportunistic maintenance policy for
294–305.
8. APPENDICES
203
APPENDIX A 204
APPENDIX A
We explain how the intensities for grades of criterion “Function” can be obtained; the
intensities of other criteria’s grades are obtained using the same method.
Step 1. Pairwise comparison matrix of the grades is constructed using expert opinion ( aij for
i 1,...,5 , j 1,...,5 ).
Table A.1. Pairwise comparison matrix for the grades of criterion “Function”
5 1
( aij ) 5
j 1
vi 5 1
, i 1,..., 5, j 1,..., 5.
5
( a
i 1
j 1
ij ) 5
vi
Intensity , i 1,...,5.
max(vi )
APPENDIX A 205
5 1 5
1 Intensity
5
aij ( aij ) 5
j 1 ( aij ) 5 j 1 vi
j 1 vi
5 1 max(vi )
5
( aij ) 5
j 1
i 1
i1
= 7.45
Therefore,
Function
Grade Intensity
Life support 1.00
Therapeutic 0.21
Patient Diagnostic 0.16
Analytical 0.13
Miscellaneous 0.11
The same method is employed for calculating the intensities of other criteria.
APPENDIX A 206
Utilization
Grade Description Intensity
High Usage hours per week 24 1.00
Medium 12 Usage hours per week < 24 0.34
Low 0 Usage hours per week < 12 0.15
Age
Grade Description Intensity
Old Actual life/Life span > 1 1.00
Almost old 0.75 Actual life/Life span 1 0.67
Average 0.5 Actual life/Life span 0.75 0.43
Almost new 0.25 Actual life/Life span 0.5 0.17
New 0 Actual life/Life span 0.25 0.12
Failure frequency
Grade Description Intensity
Frequent Likely to occur (several occurrences in 1 year) 1.00
Occasional Probably will occur (several occurrences in 1-2 years) 0.33
Uncommon Possible to occur (one occurrence in 2-5 years) 0.20
Remote Unlikely to occur (one occurrence in 5-30 year) 0.15
Failure detectability
Grade Description Intensity
Very low Not detected by regular inspection 1.00
Low Detected by inspection 0.33
Moderate Visible by naked eye 0.20
High Self-announcing 0.13
Downtime
Grade Description Intensity
High Total waiting time 72 hrs per day 1.00
Medium 24 Total waiting time < 72 hrs per day 0.25
Low Total waiting time < 24 hrs per day 0.14
Cost of repair
Grade Description Intensity
High Total repair cost 2500 $ 1.00
Medium 500 $ Total repair cost < 2500 $ 0.22
Low 0 $ Total repair cost < 500 $ 0.17
Maintenance requirements
Grade Description Intensity
Equipment that is predominantly mechanical, pneumatic or fluidics
High 1.00
in nature often requires the most extensive maintenance.
A device is considered to have average maintenance requirements
Medium 0.50
if it needs only performance verification and safety testing.
Equipment that receives only visual inspection, a basic
Low performance check, and safety testing is classified as having 0.17
minimal maintenance requirements
APPENDIX B
Table B.1 shows how a score value of 0.1050 is obtained when a device gets the lowest
Sub-Criteria Sub-Criteria
Sub-Criteria (Weight)
Main Criteria (Weight) (Weight) (Weight)
Assigned/calculated
Assigned/calculated score Assigned/calculated Assigned/calculated
score
score score
c1-Function
(0.45)
0.11
c21-Utilization
c2-Mission criticality (0.70)
(0.10) 0.15
0.165 c22 -Availability of
= 0.7*0.15 alternatives devices
+0.3*0.2 (0.30)
0.2
c3-Age
(0.06)
0.12
c4-Total risk c41-Failure frequency
(0.16) (0.30)
0.15
0.124224 c42-Detectability
=0.30*0.15 (0.24)
+0.24*0.13 0.13
+0.46*0.1044 c43-Failure c431-Operational c4311- Downtime (1.00)
APPENDIX B 210
(0.45*0.11)+(0.1*0.165)+(0.06*0.12)+(0.16*0.124224)+(0.16*0)+(0.07*0.17)=0.1050
APPENDIX C 211
APPENDIX C
Table C.1. Components/features that can be ignored since they have no failure reported
% % % %
AC 1189 13 0 1202 98.92 1.08 1510 9 145 1664 99.41 0.59 1.83
Plug/Receptacles
Alarms 881 321 0 1202 73.29 26.71 1268 251 145 1664 83.48 16.52 1.62
Circuit 48 13 1141 1202 78.69 21.31 33 3 1628 1664 91.67 8.33 2.56
Breaker/Fuse
Controls/Switches 1097 105 0 1202 91.26 8.74 1433 86 145 1664 94.34 5.66 1.54
Indicators/Displays 1094 108 0 1202 91.01 8.99 1383 136 145 1664 91.05 8.95 1
Mount 1009 191 2 1202 84.08 15.92 1319 200 145 1664 86.83 13.17 1.21
Occlusion Alarm 1191 9 2 1202 99.25 0.75 1507 12 145 1664 99.21 0.79 0.95
Table C.2. Components/features that have more failure at non-scheduled work orders
(hard failures)
APPENDIX C 212
Table C.3. Components/features that have more failures at scheduled work orders (soft failures)
We included “Occlusion Alarm” in the hard group because of its similarity to the
“Alarms” component in the qualitative test. “Grounding Resistance” has only a few failures in
both scheduled and non-scheduled work orders, so we included it in the soft group.
APPENDIX D 213
APPENDIX D
We consider all non-scheduled work order starting times as system exact failure times
(complete data) since we assume that a non-scheduled work order is requested upon a hard
failure.
The time interval between two consecutive failures ending in a SPI with at least one failure
(scheduled work order) is treated as left censored data since we consider this detected failure
as a soft failure.
The time interval between an inspection time with “no problem found” and a SPI with at
If the last available history ends in a failure and the time interval between this failure and
the end of the test is short (<=0.5*mean time between failures), we consider the time
If the last available history is a failure and the time interval between this failure and the end
of the test is long (>0.5*mean time between failures), we consider 0.5*mean time between
If the last available history is an inspection with “no failure found” and the time interval
between this inspection and the end of the test is short (<= 0.5*mean time between failures),
we consider the time interval between the last failure event and the end of the test as right
censored data.
APPENDIX D 214
If the last available history is an inspection with “no failure found” and the time interval
between this inspection and the end of the test is long (> 0.5*mean time between failures),
we consider the time interval between the last failure event + 0.5*mean time between
APPENDIX E
The “Delta method” (Meeker and Escobar, 1998) can be applied to estimate Var ( g ( )) ,
if g ( ) is a smooth nonlinear function of i and can be approximated by a linear function of the
i values in the region with no negligible likelihood. The Delta method is applied based on a
first order Taylor series expansion of g ( ) about γ [ E (1 ),..., E ( r )] .
If we apply the Delta method to the as a function of ̂ and ˆ , and assuming low
2
2 1 1
Var[ (ˆ , ˆ )] ( ) 2 Var (ˆ ) ( ) 2 Var ( ˆ ) (1 1/ ˆ ) Var (ˆ )
ˆ (1 )(1 ) / ˆ 2 ,
ˆ ˆ
d(log((x)) d((x))/dx
where ( x) digamma(x) .
dx (x)
The Var ( ˆ ) and Var (ˆ ) can be obtained from standard error estimates of SAS output.
APPENDIX F 216
APPENDIX F
The results of the Weibull distribution fitted to the times between n-1th and nth event (failure/censoring)
No of No of No of No of
No of 1
n non- right left interval ˆ ̂ ˆ ˆ( 1)
observations ˆ
censored censors censors censors
1 674 502 172 0 0 1.2982 1215.687 1123.1
2 502 317 185 0 0 1.1256 1077.554 1032.1
3 317 185 132 0 0 1.1007 924.7741 892.1
4 185 91 94 0 0 0.9864 768.3577 772.9
5 91 46 45 0 0 0.8626 701.4417 756.3
6 121 76 45 0 0 0.7785 310.3867 358.8
Table F.1. Estimated parameters of Weibull distribution for times between events – hard failures
LA=6.12 (systems are degrading)
3
No. of Uni t s 674
No. of Event s 1217
Conf . Coef f . 95%
S
a
m
p
l
e 1
M
C
F
-1
days
The results of the Weibull distribution fitted to the times between n-1th and nth event (failure/censoring)
No of No of No of No of
No of 1
n non- right left interval ˆ ̂ ˆ ˆ( 1)
observations ˆ
censored censors censors censors
1 674 0 30 463 181 1.1162 350.7222 336.8
2 644 0 122 418 104 1.2525 254.8389 237.2
3 523 0 116 335 72 1.0424 247.6217 243.5
4 404 0 158 217 29 1.0465 303.4680 298
5 247 0 124 108 15 0.8027 409.1623 462.5
6 123 0 76 40 7 0.5665 701.3626 1142.3
7 82 0 46 26 10 0.7959 351.3244 399.5
Table F.2. Estimated parameters of Weibull distribution for times between events – soft
failures LA=11.08 (systems are degrading)
8
No. of Uni t s 674
No. of Event s 2027
Conf . Coef f . 95%
S
a 4
m
p
l
e
M
C
F 2
-2
days
The results of the Weibull distribution fitted to the times between n-1th and nth event (failure/censoring)
No of No of No of No of
No of 1
n non- right left interval ˆ ̂ ˆ ˆ( 1)
observations ˆ
censored censors censors censors
1 674 306 368 0 0 1.1876 2564.701 2419.1
Table F.3. Estimated parameters of Weibull distribution for times between events – Alarms
failures (hard) LA=4.13 (systems are degrading)
The results of the Weibull distribution fitted to the times between n-1th and nth event (failure/censoring)
n No of No of No of No of No of
1
observations non- right left interval ˆ ̂ ˆ ˆ( 1)
ˆ
censored censors censors censors
1 674 0 101 317 256 1.2662 645.6414 599.6
2 573 0 146 251 176 1.1804 497.9293 470.4
3 428 0 173 178 77 1.2267 477.8781 447.1
4 252 0 146 79 27 1.5471 473.8461 426.3
5 107 0 75 24 8 1.3434 597.5264 548.4
6 40 0 32 7 1 0.7905 1395.492 1594.9
Table F.4. Estimated parameters of Weibull distribution for times between events –
Chassis/Housing failures (soft) LA=11.40 (systems are degrading)
APPENDIX G 219
APPENDIX G
For the general case and the special case of Cn ( yn ) 1 y0 0 is considered. We first
want to calculate
Let
Therefore, P( An | yn 1 ) Cn 1 ( yn 1 ) , and
An1
fYn1 ( yn 1 | yn 2 )[ fYn ( yn | yn 1 )dyn ]dyn 1
An
An1
fYn1 ( yn1 | yn2 )Cn1 ( yn1 )dyn1 Cn2 ( yn2 ) ,
or
ui
Therefore,
E ( t ) [ g (Yn ) I ( An ) | yn 1 ] g( y ) f
An
n Yn ( yn | yn 1 )dyn
B (y ) f
An
n n Yn ( yn | yn 1 )dyn Bn 1 ( yn 1 ) ,
E ( t ) [ g (Yn ) I ( An , An 1 ) | yn 2 ]
g( y ) f
An1 An
n Yn ( yn | yn 1 ) fYn1 ( yn 1 | yn 2 )dyn dyn 1
An1
fYn1 ( yn 1 | yn 2 )[ g ( yn ) fYn ( yn | yn 1 )dyn ]dyn 1
An
An1
fYn1 ( yn 1 | yn 2 ) Bn 1 ( yn 1 )dyn 1 Bn 2 ( yn 2 )
Thus, in general,
ui
Bi 1 ( yi 1 ) Bi ( yi 1 x) f X i ( x | yi 1 )dx
li
Therefore,
APPENDIX H
n n
E (t ) [( ln Yi )I ( An , An 1 ,..., A1 )] ... ( ln yi ) fYn ( yn | yn1 )... fY1 ( y1 )dyn ...dy1.
i 1 A1... An i 1
[ln y f
An1
n 1
An
Yn ( yn | yn1 )dyn ln yn fYn ( yn | yn 1 )dyn ] fYn1 ( yn 1 | yn 2 )dyn 1
An
[ln y
An1
Cn1 ( yn1 ) Dn1 ( yn1 )] fYn1 ( yn1 | yn2 )dyn1
n 1
Thus, in general,
ui
Therefore,
n
E ( t ) [ ln Yi | An , An 1 ,..., A0 ] D0 (0) / C0 (0) .
i 1
APPENDIX I 222
APPENDIX I
1 (t ) E[ ( X 1 , I1 ) | X 0 t ] [ ( x, 0) Pr{I1 0 | X 1 x, X 0 t}
0
k (t ) E[ ( X k , I k ) | X 0 t ] E{E[ ( X k , I k ) | X 1 ] | X 0 t}
E[ ( X k , I k ) | X 1 x, X 0 t ]dF1 ( x | t )
0
E[ ( X k , I k ) | X 1 x, X 0 t ] f1 ( x | t )dx E[ ( X k , I k ) | X 1 , X 0 t ]Pr{ X 1 | X 0 t}
0
E{E[ ( X k , I k ) | I1 ] | X 1 x, X 0 t} f1 ( x | t )dx k 1 (t ) P1 (t )
0
{E[ ( X k , I k ) | I1 0, X 1 x, X 0 t ]Pr{I1 0 | X 1 x, X 0 t}
0
APPENDIX J
E[ ( X 1 , I1 ) (T ) | I1 1, X 1 x, X 0 t ] ( x,1) ,
we find that
1 (t ) E[ ( X 1, I1 ) | X 0 t ] [ ( x,0)r(t x) ( x,1)r (t x)] f1( x | t )dx ( ,2) P1(t ) .
0
because E[ ( X k , I k ) I (T k ) | I1 1, X 1 x, X 0 t ] 0 .
APPENDIX K 224
APPENDIX K
Upper bound for V (t ) Vk (t )
k 1
Lemma 1: Let Vk (t ) Pr{T k |t} and V2 (t ) sup V2 (t x) . If V2 (t ) 1 , t 0 then also
x 0
1 1
Vk (t ) 1 , k 3, 4,... , and V (t ) .
1 V2 (t ) 1 V2 (0)
Proof:
We will show that Vk (t ) V2 (t )Vk 1 (t ) [V2 (t )]k 1 for k 2 . We have V1 (t ) Pr{T |t} 1
and obviously for k 1, 2 we have Vk (t ) [V2 (t )]k 1 and for k 2 , V2 (t ) V2 (t )V1 (t ) V2 (t ) .
k
Let X i be the survival time on ((i 1) , i ], i 1, 2,... , and Sk X i , Fk ( x | t ) Pr{S k x | t} .
i 1
( k 2)
Vk (t )
0
[Pr{T (k 2) | S k 2 x, t}V2 (t x)dFk 2 ( x | t )
( k 2)
Proof:
APPENDIX K 225
We will prove that for V2 (t ) 1 V2 (t ) r (t x) f1 ( x | t )dx , V2 (t ) inf V2 (t x) 1 V2 (t ) 0 .
x 0
0
u
From our assumptions, r (u) 1 r (u) and P1 (u ) 1 exp[ ( x)dx] 0
u
are nondecreasing
V2 (t x ) r (t x y ) f1 ( y | t x )dy r (t ) f1 ( y | t x )dy r (t ) P1(t x ) r (t ) P1(t ) 0 .
0 0
Then V2 (t ) inf V2 (t x) r (t ) P1 (t ) 0 for t 0 , but it is not clear whether V2 (0) 0 because
x 0
r (0) may not be positive. On the other hand, for any fixed x0 0 and 0 x x0 , and t 0 ,
t x y t x0 y
( s ) ds ( s ) ds
V2 (t x ) r (t x y ) (t x y ) e tx
dy r (t y ) (t y ) e t x0
dy g (t, x0 ) 0 .
0 0
Then
for t 0 .
n n n
V2 (t )
Vk (t ) Vk (t ) Vk (t ) Vn (t ) [V2 (t )]k n Vk (t ) Vn (t )
k 1 k 1 k 1 k n 1 k 1 1 V2 (t )
.
n
V2 (t )
n 1 V2 (t ) [V2 (t )]n
So, 0 Vk (t ) Vk (t ) Vn (t ) [V
(t )] .
1 V2 (t ) 1 V2 (t ) 1 V2 (t )
2
k 1 k 1
APPENDIX K 226
V2 (t )
Therefore, criterion Vn (t ) can be used to decide about terminating calculation of
1 V2 (t )
V (t ) .
k 1
k
t
( s ) ds
Note: If ( x) , then P1 (t ) e t
P1 exp{- lim ( x)} 0, t
x
Note also that V2 (t ) V2 (t; ) r (t x) f1 ( x | t )dx is a nondecreasing function in , as a function
0
lim V2 (t; ) r (t x) f1 ( x | t )dx , limV2 (t; ) 0 .
0
0
Lemma 3: Let the placeholder function ( X k , I k ) be bounded on finite intervals, that is,
G (t ) k (t ) converges.
k 1
Proof:
Obviously, from G (t ) E[ ( X k , I k ) ( T k ) | X 0 t ] ,
k 1
| G (t ) | E{[| ( X k , I k ) | ( T k )] | X 0 t} A E[ ( T k ) | X 0 t ] AV (t ) .
k 1 k 1
so that
APPENDIX K 227
n
V2 (t ) [V2 (t )]n
| k (t ) k (t ) | A Vk (t ) AV
n (t ) A .
k 1 k 1 k n 1 1 V2 (t ) 1 V2 (t )
A somewhat better approximation can be obtained following the proof of Lemma 2. Let
Then
and
n
V2 (t ) [V2 (t )]n
| k (t ) k (t ) | 1 (t ) Vk (t ) 1 (t )Vn (t )
( t ) .
1 V2 (t ) 1 V2 (t )
1
k 1 k 1 k n 1
and then
n
1 [V2 (t )]n1
| k (t ) k (t ) | 2 (t ) Vk 1(t ) 2 (t )Vn (t )
( t ) ,
1 V2 (t ) 1 V2 (t )
2
k 1 k 1 k n 1
APPENDIX L
E[V | Z1 z, X 0 t , Z 0 s,U ] f1Z ( z | s)dz
0
E[V | X1 x, Z1 , X 0 t , Z 0 s,U ]dF1X ( x | s, z ) P1Z ( | s)
0
z
{ ( x, z ) E V | X 0 t x, Z 0 s z,U z } f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z ) E[V | X 0 t z, Z 0 s z,U z ]}P1X ( z | t ) f1Z ( z | s)dz
0
P1Z ( | s) ( x1 , ) f1X ( x | t )dx P1Z ( | s) P1X ( | t ) ( , )
0
z
{ ( x, z ) G1 ( z, t x, s z )} f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z ) G1 ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz
0
P ( | s) ( x, ) f1X ( x | t )dx P1Z ( | s) P1X ( | t ) ( , ) .
1
Z
So,
APPENDIX L 229
z
G1 ( , t , s) { ( x, z ) G1 ( z, t x, s z )} f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z ) G1 ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz .
0
P ( | s) ( x, ) f1X ( x | t )dx P1Z ( | s) P1X ( | t ) ( , )
1
Z
E[V | Z1 z, X 0 t , Z 0 s,U (k 1) ] f1Z ( z | s)dz
0
E[V | X1 x, Z1 , X 0 t , Z 0 s,U (k 1) ]dF1X ( x | s, z ) P1Z ( | s)
0
z
{ ( x, z ) E V | X 0 t x, Z 0 s z,U z (k 1) } f1X ( x | t ) f1Z ( z | s)dxdz
0 0
{ ( z, z ) E[V | X 0 t z, Z 0 s z,U z (k 1) ]}P1X ( z | t ) f1Z ( z | s)dz
0
P ( | s) { ( x1 , ) E[V | X 0 t x, Z 0 s ,U (k 1) ]} f1X ( x | t )dx
1
Z
{ ( z, z ) Gk ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz
0
P1Z ( | s) { ( x, ) Gk 1 ( , t x, s )} f1X ( x | t )dx
0
P1Z ( | s ) P1 X ( | t ){ ( , ) Gk 1 ( , t , s )} .
APPENDIX M 231
APPENDIX M
Table M.1. Estimated expected number of minimal repairs and uptime and their standard
errors calculated for the components with soft failure, for 1, 2,...,12 (used in Figure 5.14)
APPENDIX M 233
Component
M nj STDE Rnj STDE enj STDE
Table M.2. Expected number of minimal repairs and replacements, and the uptimes and
their standard errors calculated for the components with soft failure, for 1, 2,...,12
(used in Figures 5.15 and 5.16)
APPENDIX N 236
APPENDIX N
G1 ( , t , s) G1 ( , t , s) E[V | X 0 t , Z 0 s, U ] E{E[V | Z1 ] | X 0 t , Z 0 s, U }
z
( x, z, 0) E V | X 0 t x, Z 0 s z ,U z (k 1) Pr{I1 0 | X 1 x, X 0 t}
0 0
( x, z,1) E V | X 0 0, Z 0 0,U z (k 1) Pr{I1 1| X 1 x, X 0 t} f1X ( x | t ) f1Z ( z | s)dxdz
{ ( z, z, 2) E[V | X 0 t z, Z 0 s z,U z (k 1) ]}P1X ( z | t ) f1Z ( z | s)dz
0
P1Z ( | s) ( x, , 0) E[V | X 0 t x, Z 0 s ,U (k 1) ] Pr{I1 0 | X 1 x, X 0 t}
0
( x, z, 0) E V | X 0 t x, Z 0 s z,U z (k 1) Pr{I1 0 | X 1 x, X 0 t}
0 0
( x, z,1) E V | X 0 0, Z 0 s z ,U z (k 1) Pr{I1 1| X 1 x, X 0 t} f1X ( x | t ) f1Z ( z | s )dxdz
{ ( z, z ) E[V | X 0 t z, Z 0 s z,U z (k 1) ]}P1X ( z | t ) f1Z ( z | s)dz
0
APPENDIX N 237
P ( | s) ( x, , 0) E[V | X 0 t x, Z 0 s ,U (k 1) ] Pr{I1 0 | X 1 x, X 0 t}
1
Z
{ ( z, z, 2) Gk ( z, t z, s z )} f1Z ( z | s) P1X ( z | t )dz
0
P1Z ( | s) {[ ( x, , 0) Gk 1 ( , t x, s )]r (t x) [ ( x, ,1) Gk 1 ( , 0, s )]r (t x)} f1X ( x | t )dx
0
P1Z ( | s ) P1 X ( | t ){ ( , , 2) Gk 1 ( , t , s )} .
APPENDIX O 238
APPENDIX O
E[ E[V | J ] | Z1 z, X 0 t , (1 , 2 ,..., m1 ),U ]dF1Z ( z | (1, 2 ,..., m1 ))
0
m1
E[V | Z1 z, J j, X 0 t , (1 ,2 ,..., m1 ),U ] q j ( z )dF1Z ( z | (1 , 2 ,..., m1 ))
0 j 1
m1
E[V | Z1 z, J j, X 0 t ,(1 , 2 ,..., m1 ),U ] q j ( z ) f1Z ( z | (1 , 2 ,..., m1 ))dz
0 j 1
z m1
{E[V | Z1 z , J j , X 1 x, X 0 t , I Zj 0, (1 , 2 ,..., m1 ), U ] rjZ ( j z )
0 0 j 1
E[V | Z1 z , J j , X 1 x, X 0 t , I Zj 0, I X 0, (1 , 2 ,..., m1 ),U ] r X (t x)
0 0 j 1
E[V | Z1 z , J j , X 1 x, X 0 t , I Zj 0, I X 1, (1 , 2 ,..., m1 ),U ]r X (t x) rjZ ( j z )
E[V | Z1 z , J j , X 1 x, X 0 t , I Zj 1, I X 0, (1 , 2 ,..., m1 ),U ] r X (t x)
E[V | Z1 z , J j , X 1 x, X 0 t , I Zj 1, I X 1, (1 , 2 ,..., m1 ),U ] r X (t x) rj Z ( j z )
q j ( z ) f1X ( x | t , z ) f1Z ( z | (1 , 2 ,..., m1 )) dxdz
m1
{E[V | Z1 z , J j , X 1 z , X 0 t , I Zj 0, (1 , 2 ,..., m1 ), U ] rjZ ( j z )
0 j 1
z m1
[ ( x, 0, z, j, 0) G1 ( z, t x, 1 z, 2 z,..., j z,..., m1 z )] r X (t x)
0 0 j 1
[ ( x,1, z, j, 0) G1 ( z, 0, 1 z, 2 z,..., j z,..., m1 z )] r X (t x) rjZ ( j z )
[ ( x, 0, z, j ,1) G1 ( z , t x, 1 z , 2 z ,..., 0,..., m1 z )] r X (t x)
[ ( x,1, j ,1) G1 ( z , 0, 1 z, 2 z,..., 0,..., m1 z )] r X (t x) rj Z ( j z )
q j ( z ) f1X ( x | t , z ) f1Z ( z | (1 , 2 ,..., m1 ))dxdz
m1
{[ ( z, 2, z, j, 0) G1 ( z, t z, 1 z, 2 z,..., j z,..., m1 z )] rjZ ( j z )
0 j 1
APPENDIX P
Tables P.1 and P.2 present the data history for audible signal and housing/chassis units,
respectively. The data include unit no, and the lower and upper bounds for the censoring
intervals. The number in the brackets indicates the history number of the unit. The origin for
measuring l j ,i 1 and u j ,i 1 is the previous failure detection time ( u j ,i ). u ji indicates the last
(right) censoring.
1 1.289 24(1) 1.211 2.389 49(1) 0.000 1.522 68(1) 1.439 2.583
9 (1) 2.331 2.369 (2) 4.806 55(1) 0.000 1.433 (3) 2.917
14(1) 0.000 1.300 35 1.289 58(1) 0.519 1.983 (2) 0.000 1.689
16(1) 1.831 3.853 39 1.289 60(1) 1.842 3.814 78(1) 1.733 3.908
17(1) 1.436 1.961 41(1) 0.000 1.942 61(1) 1.531 1.828 79(1) 0.000 1.947
(3) 2.436
Table P.1. Audible signal dataset for 80 units’ failure and censoring events
APPENDIX P 243
1(1) 2.183 3.739 12(1) 1.206 2.128 21(1) 0.001 2.258 29(1) 0.001 1.089
(2) 0.833 1.925 (2) 1.031 1.978 (2) 0.000 1.308 (2) 0.000 0.328
(3) 0.000 0.697 (3) 0.000 0.542 (3) 0.439 2.594 (3) 0.506 1.967
(4) 1.406 (4) 0.000 1.461 (4) 0.000 1.078 (4) 0.000 0.833
2(1) 0.544 1.092 (5) 0.956 (5) 0.528 (5) 0.086 0.317
(2) 0.000 1.092 13(1) 0.001 1.853 22(1) 0.001 2.294 (6) 1.772
(3) 3.436 3.867 (2) 0.283 1.400 (2) 0.000 1.364 30(1) 0.001 2.047
(4) 0.956 (3) 0.000 0.794 (3) 1.261 1.717 (2) 2.622 4.125
3(1) 0.000 2.214 (4) 0.000 2.278 (4) 0.000 1.100 (3) 0.956
(2) 1.431 1.931 (5) 0.000 0.542 (5) 0.000 0.953 31(1) 0.000 2.028
(3) 0.000 1.244 (6) 0.900 (6) 0.339 (2) 0.000 1.803
(4) 0.000 1.008 14(1) 0.000 0.544 23(1) 0.000 2.136 (3) 0.117 1.378
(5) 0.000 1.017 (2) 0.000 1.075 (2) 0.000 0.194 (4) 0.000 1.992
(6) 0.353 (3) 0.000 0.867 (3) 0.000 0.039 (5) 0.567
4(1) 0.000 0.544 (4) 1.256 1.933 (4) 2.322 3.017 32(1) 0.000 1.736
(2) 0.756 1.497 (5) 0.953 1.336 (5) 0.000 1.319 (2) 0.000 0.872
(3) 2.031 4.072 (6) 1.167 1.711 (6) 0.000 1.033 (3) 0.000 1.050
(4) 0.956 (7) 0.000 0.036 (7) 0.028 (4) 0.000 2.408
5(1) 0.000 5.239 (8) 0.264 24(1) 0.000 2.178 (5) 0.239
(2) 0.000 1.156 15(1) 0.000 2.294 (2) 1.894 2.725 33(1) 0.000 1.942
(3) 1.372 (2) 0.000 1.519 (3) 0.000 1.228 (2) 1.961 2.981
APPENDIX P 244
6(1) 0.000 2.119 (3) 0.211 2.347 (4) 0.000 0.761 (3) 1.383
(2) 0.000 1.286 (4) 0.000 0.742 (5) 0.172 0.731 34(1) 0.000 1.478
(3) 0.636 2.778 (5) 0.864 (6) 0.144 (2) 0.083 1.783
(4) 0.956 16(1) 0.000 2.203 25(1) 0.000 2.214 (3) 0.181 2.186
7(1) 2.058 3.819 (2) 0.000 1.617 (2) 0.000 1.431 (4) 0.858
(2) 0.292 1.381 (3) 1.644 3.247 (3) 0.369 2.508 35(1) 0.000 1.386
(3) 0.000 1.997 (4) 0.700 (4) 0.956 (2) 0.425 2.469
(4) 0.569 17(1) 0.000 2.047 26(1) 0.000 2.042 (3) 0.000 2.150
8(1) 1.206 2.136 (2) 3.600 5.108 (2) 0.000 1.786 (4) 0.300
(2) 3.139 4.578 (3) 0.611 (3) 1.819 3.044 36(1) 0.000 1.300
(3) 1.053 18(1) 0.000 1.444 (4) 0.894 (2) 0.211 0.975
9(1) 2.008 3.197 (2) 2.014 3.317 27(1) 0.000 1.203 (3) 1.911 2.622
(2) 3.658 (3) 1.544 (2) 0.233 0.758 (4) 0.000 1.264
10(1) 0.000 1.572 19(1) 1.825 3.408 (3) 0.000 1.311 (5) 0.144
(2) 0.000 2.003 (2) 0.000 1.106 (4) 0.117 0.656 37(1) 3.669 5.697
(3) 0.000 1.961 (3) 0.000 0.992 (5) 0.000 1.497 (2) 0.000 0.542
11(1) 3.419 4.567 20(1) 1.031 1.831 28(1) 0.000 1.386 38(1) 3.869 5.831
(2) 0.000 0.219 (2) 2.022 2.889 (2) 0.000 1.100 (2) 0.475
(4) 0.708
Table P.2. Housing/Chassis dataset for 38 units’ failure and censoring events
APPENDIX Q 245
APPENDIX Q
Source Codes
In the optimization model over a finite time horizon we need to calculate the following
variables. The name and arguments of the MATLAB function which calculates the required
The probability that a component survives in the ith inspection interval, where i=1,2,…,k.
The probability that a component fails in the ith inspection interval, and is minimally repaired,
where i=1,2,…,k.
function intout=Simp_int_integral_GeneralM(k,tow,sigma,beta,eta,h,a,b)
The probability that a component fails in the ith inspection interval, and is replaced, where
i=1,2,…,k.
function intout=Simp_int_integral_GeneralR(k,tow,sigma,beta,eta,h,a,b)
APPENDIX Q 246
The expected downtime of a component in the ith inspection interval, where i=1,2,…,k.
function intout=Simp_int_integral_GeneralR(k,tow,sigma,beta,eta,h,a,b)
In addition, other functions are developed to call the required functions for calculating the
probabilities and expected values for all components. For example function
In the optimization model over infinite time horizon we need to calculate the following
variables. The name and arguments of the MATLAB function, which calculates the required
The expected number of inspection intervals in a cycle for a component, i.e., number of
In addition, other functions are developed to call the required functions for calculating the
probabilities and expected values for all components. For example function
given tow.
2. Minimal repair of hard failures, and minimal repair or replacement of components with soft
failure.
3. Minimal repair or replacement of both components with soft and hard failure.
inspections.
each case. However, calculating accurately the multi-dimensional integrals requires to consider a
very small value for step h in applying Simpson’s rule. This increases intensively the
APPENDIX Q 248
calculatation time of the procedure. We even improved the performance of the code, but still the
running time is too long. Therefore, eventually we developed a simulation model for each
In the first case, we need to calculate the expected number of minimal repairs and the
expected downtime of each component over its life cycle, in presence of opportunistic
inspections. We present the function developed for recursively calculating the expected number
function GCalculator_Minimal_Hard_improve3(totaltime,tau,t,beta1,eta1,s1,
beta2,eta2,s2,beta3,eta3,h)
For each opportunistic case, a simulation model was developed in MATLAB to obtain
Case 1. The expected number of minimal repairs and the expected downtime of components with
function SO = SimulationOpportunisticFile(TotalTime,Tau)
Case 2. The expected number of minimal repairs and replacements, and the expected downtime
function SO = SimulationOpportunisticMRSoftFile(TotalTime,Tau)
Case 3. The expected number of minimal repairs and replacements, and the expected downtime
of components with soft failure should be calculated. In the simulation model developed for this
case, the expected number of minimal repairs and replacements of components with hard failure
are also calculated, but they do not contribute in the cost model.
function SO = SimulationOpportunisticMRSoftHardFile(TotalTime,Tau)
Case 4. The expected number of minimal repairs and replacements, and the expected downtime
of components with soft failure should be calculated. In addition, the expected number of
minimal repairs and replacements of components with hard failures within inspection intervals,
function SO = SimulationOpportunisticMRSoftHardPMFile(TotalTime,Tau)
functions C , B1 , B2 , B3 , and B4 (shown below) at each iteration. The iterations continue until
B1 ... y
n fY ( y1 ,..., yn )dy1...dyn
yi1 li yi yi 1 ui
i 1,..., n
APPENDIX Q 250
B2 ... y
n log yn fY ( y1 ,..., yn )dy1...dyn
yi1 li yi yi 1 ui
i 1,..., n
B3 ... y
n (log yn )2 fY ( y1 ,..., yn )dy1...dyn
yi1 li yi yi1 ui
i 1,..., n
n
B4 ...
yi1 li yi yi1 ui y f i 1
i Y ( y1 ,..., yn )dy1...dyn
i 1,..., n
A function is required to estimate the parameters using the modified version of EM.
function GN = generalLike()
algorithm.
B1 , B2 , B3 , and B4 at each iteration. The iterations continue until the convergence of the
parameters.
function GN = generalLikeEM()
function [C,B1,B2,B3,B4]=
APPENDIX Q 251
Simp_int_integral_GeneralLikelihoodEMX(k,beta0,alpha0,beta1,alpha1,h,L,U)
To calculate the standard errors of the parameters, which are calculated using any
methods (modified EM and complete EM), we developed a function to calculate the Hessian
matrix, or the second order derivatives of the likelihood, and the standard errors.
function GN = InformationMatrix()
In order to compare the parameters estimated from different methods (modified EM and
complete EM) with the case that the midpoint of each failure interval is considered as an exact
failure time, we developed a function to estimate the parameters under the assumption of having
complete data.
function GN = generalLike_Mid()
To perform the likelihood ratio test we need to find parameter iteratively assuming that
1 .
GN = generalLike_Beta1_2()
function GN = generalprobs(n,tow,sigma)
beta=[1.3,1.1,2.1,1.8,1.7];
eta=[3.5,4.6,6,10,3.6];
% prob of minimal repair at etha=0.6
a=[0.9,0.9,0.9,0.9,0.9];
b=[0.2317,0.1763,0.1352,0.0811,0.2253];
for i=1:numel(beta);
N(i,1)=i;
intout
=Simp_int_integral_GeneralP(n,tow,sigma,beta(i),eta(i),0.0001,a(i),b(i))
N(i,2:numel(intout)+1)=intout(1:n);
end;
GN=N;
xlswrite('C:\Project\paper 4\excel\probs\Generalprobs5.xls',N);
end
end;
else
TmpPArray=PArray(fix(x/h+1):fix((x+tow)/h+1));
if sigma ~= 0
TmpPSigmaArray= PSigmaArray(fix(x/h+1):fix((x+tow)/h+1));
end;
TmpIndArray=IndArray(fix(x/h+1):fix((x+tow)/h+1));
if x==0
PArrayZero=PArray(1);
if sigma ~= 0
PSigmaArrayZero=PSigmaArray(1);
end;
end;
PArray(i)=TmpPArray(numel(TmpPArray))*exp(-
1*(((x+tow)/eta)^beta-(x/eta)^beta));
if sigma ~= 0
PSigmaArray(i)=TmpPSigmaArray(numel(TmpPSigmaArray))*exp(-
1*(((x+tow)/eta)^beta-(x/eta)^beta));
end;
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
TmpPArray(r)=(TmpPArray(r)*rf+PArrayZero*(1-
rf))*(beta/eta)*((TmpIndArray(r)/eta)^(beta-1))*exp(-
1*((TmpIndArray(r)/eta)^beta-(x/eta)^beta));
if sigma ~= 0
TmpPSigmaArray(r)=(TmpPSigmaArray(r)*rf+PSigmaArrayZero*(1-
rf))*(beta/eta)*((TmpIndArray(r)/eta)^(beta-1))*exp(-
1*((TmpIndArray(r)/eta)^beta-(x/eta)^beta));
end;
end;
PArray(i)=PArray(i)+h*(TmpPArray(1)+4*sum(TmpPArray(2:2:(numel(TmpPArray)-
1)))+2*sum(TmpPArray(3:2:(numel(TmpPArray)-
2)))+TmpPArray(numel(TmpPArray)))/3;
if sigma ~= 0
PSigmaArray(i)=PSigmaArray(i)+h*(TmpPSigmaArray(1)+4*sum(TmpPSigmaArray(2:2:(
numel(TmpPSigmaArray)-1)))+2*sum(TmpPSigmaArray(3:2:(numel(TmpPSigmaArray)-
2)))+TmpPSigmaArray(numel(TmpPSigmaArray)))/3;
end;
IndArray(i)=x;
if x==0
intout(j)=PArray(i);
end;
APPENDIX Q 254
end;
end;
end;
if sigma == 0
TmpPArray=PArray(1:fix(tow/h+1));
TmpIndArray=IndArray(1:fix(tow/h+1));
intout(k)=TmpPArray(numel(TmpPArray))*exp(-1*((tow/eta)^beta));
PArrayZero=PArray(1);
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
TmpPArray(r)=(TmpPArray(r)*rf+PArrayZero*(1-
rf))*(beta/eta)*((TmpIndArray(r)/eta)^(beta-1))*exp(-
1*(TmpIndArray(r)/eta)^beta);
end
intout(k)=intout(k)+h*(TmpPArray(1)+4*sum(TmpPArray(2:2:(numel(TmpPArray)-
1)))+2*sum(TmpPArray(3:2:(numel(TmpPArray)-
2)))+TmpPArray(numel(TmpPArray)))/3;
else
TmpPSigmaArray=PSigmaArray(1:fix(tow/h+1));
TmpIndArray=IndArray(1:fix(tow/h+1));
intout(k)=TmpPSigmaArray(numel(TmpPSigmaArray))*exp(-
1*((tow/eta)^beta));
PSigmaArrayZero=PSigmaArray(1);
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
TmpPSigmaArray(r)=(TmpPSigmaArray(r)*rf+PSigmaArrayZero*(1-
rf))*(beta/eta)*((TmpIndArray(r)/eta)^(beta-1))*exp(-
1*(TmpIndArray(r)/eta)^beta);
end
intout(k)=intout(k)+h*(TmpPSigmaArray(1)+4*sum(TmpPSigmaArray(2:2:(numel(TmpP
SigmaArray)-1)))+2*sum(TmpPSigmaArray(3:2:(numel(TmpPSigmaArray)-
2)))+TmpPSigmaArray(numel(TmpPSigmaArray)))/3;
end;
end
end
%for P1(0)
function first_func=FirstFunction(x,tow,beta,eta)
if x==0
first_func=exp(-1*(tow/eta)^beta);
else
first_func=exp(-1*(((x+tow)/eta)^beta-(x/eta)^beta));
end;
end
end
function first_func=FirstFunction(x,tow,beta,eta,a,b)
if x==0
F=@(y) (a*(beta/eta).*(y/eta).^(beta-1)).*exp((-b.*y)-(y/eta).^beta);
else
F=@(y) (a*(beta/eta).*((x+y)/eta).^(beta-1)).*exp((-b.*(x+y))-
((x+y)/eta).^beta+(x/eta).^beta);
end;
first_func=quad(F,0,tow);
end
end
% for R1(0)
function first_func=FirstFunction(x,tow,beta,eta,a,b)
if x==0
F=@(y) (1-a*exp(-1*b*y))*(beta/eta).*(y/eta).^(beta-1).*exp(-
(y/eta).^beta);
else
F=@(y) ((1-a*exp(-1*b*(x+y)))*(beta/eta).*((x+y)/eta).^(beta-1)).*exp(-
((x+y)/eta).^beta+(x/eta).^beta);
end;
first_func=quad(F,0,tow);
APPENDIX Q 256
end
end
% for e1(0)
function first_func=FirstFunction(x,tow,beta,eta)
if x==0
F=@(y) exp(-1.*(y/eta).^beta);
else
F=@(y) exp(-1.*(((y+x)/eta).^beta-(x/eta).^beta));
end;
first_func=quad(F,0,tow);
end
function GN = generalVs(tow)
beta=[1.3,1.1,2.1,1.8,1.7];
eta=[3.5,4.6,6,10,3.6];
% prob of minimal repair at etha=0.6
a=[0.9,0.9,0.9,0.9,0.9];
b=[0.2317,0.1763,0.1352,0.0811,0.2253];
for i=1:numel(beta);
intout
=Simp_int_integral_GeneralInfinite_V(tow,beta(i),eta(i),0.001,a(i),b(i),0.000
1)
N(i)=intout;
end;
GN=N;
xlswrite('C:\Project\paper 4\excel\Vs\Generalv12.xls',N);
end
APPENDIX Q 257
function intout
=Simp_int_integral_GeneralInfinite_V(tow,beta,eta,h,a,b,epsilon)
tic;
ContWhile=true;
k=1;
while ContWhile %ContWhile
i=0;
stopcon=tow;
for x=0:h:stopcon;
i=i+1;
if k==1
cell{k}(i)=1; %FirstFunction(x,tow,beta,eta);
else
if x > 0
for l=1:1:k-1
y=(k-l)*tow+x;
if l==1
cell{l}(round(y/h+1))=1; %FirstFunction(y,tow,beta,eta);
else
TmpPArray=cell{l-1}(round(y/h+1):round((y+tow)/h+1));
TmpIndArray=[y:h:y+tow];
cell{l}(round(y/h+1))=TmpPArray(end)*exp(-
1*(((y+tow)/eta)^beta-(y/eta)^beta));
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
TmpPArray(r)=(TmpPArray(r)*rf)*(beta/eta)*((TmpIndArray(r)/eta)^(beta-
1))*exp(-1*((TmpIndArray(r)/eta)^beta-(y/eta)^beta));
end;
cell{l}(round(y/h+1))=cell{l}(round(y/h+1))+h*(TmpPArray(1)+4*sum(TmpPArray(2
:2:(numel(TmpPArray)-1)))+2*sum(TmpPArray(3:2:(numel(TmpPArray)-
2)))+TmpPArray(numel(TmpPArray)))/3;
end
end % for l=1:1:k-1
end % if x > 0
TmpPArray=cell{k-1}(round(x/h+1):round((x+tow)/h+1));
TmpIndArray=[x:h:x+tow];
APPENDIX Q 258
cell{k}(i)=TmpPArray(end)*exp(-1*(((x+tow)/eta)^beta-
(x/eta)^beta));
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
% rf=1;
TmpPArray(r)=(TmpPArray(r)*rf)*(beta/eta)*((TmpIndArray(r)/eta)^(beta-
1))*exp(-1*((TmpIndArray(r)/eta)^beta-(x/eta)^beta));
end;
cell{k}(i)=cell{k}(i)+h*(TmpPArray(1)+4*sum(TmpPArray(2:2:(numel(TmpPArray)-
1)))+2*sum(TmpPArray(3:2:(numel(TmpPArray)-
2)))+TmpPArray(numel(TmpPArray)))/3;
end; % if k==1
end; % for x=0:h:stopcon;
cell{k}(1)
if k > 2
checkterm= cell{k}(1)*cell{2}(1) /(1-cell{2}(1));
if checkterm < epsilon
ContWhile=false;
Sumterms=0;
for cnt=1:1:k;
Sumterms=Sumterms+cell{cnt}(1);
end;
k
intout =Sumterms
end
end
k=k+1;
end % while
toc;
end % function
function intout
=Simp_int_integral_GeneralInfinite_n(tow,beta,eta,h,a,b,epsilon,V2)
tic;
ContWhile=true;
APPENDIX Q 259
k=1;
while ContWhile %ContWhile
i=0;
stopcon=tow;
for x=0:h:stopcon;
i=i+1;
if k==1
cell{k}(i)=FirstFunction(x,tow,beta,eta,a,b);
else
if x > 0
for l=1:1:k-1
y=(k-l)*tow+x;
if l==1
cell{l}(round(y/h+1))=FirstFunction(y,tow,beta,eta,a,b);
else
TmpPArray=cell{l-1}(round(y/h+1):round((y+tow)/h+1));
TmpIndArray=[y:h:y+tow];
cell{l}(round(y/h+1))=TmpPArray(end)*exp(-
1*(((y+tow)/eta)^beta-(y/eta)^beta));
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
TmpPArray(r)=(TmpPArray(r)*rf)*(beta/eta)*((TmpIndArray(r)/eta)^(beta-
1))*exp(-1*((TmpIndArray(r)/eta)^beta-(y/eta)^beta));
end;
cell{l}(round(y/h+1))=cell{l}(round(y/h+1))+h*(TmpPArray(1)+4*sum(TmpPArray(2
:2:(numel(TmpPArray)-1)))+2*sum(TmpPArray(3:2:(numel(TmpPArray)-
2)))+TmpPArray(numel(TmpPArray)))/3;
end
end % for l=1:1:k-1
end % if x > 0
TmpPArray=cell{k-1}(round(x/h+1):round((x+tow)/h+1));
TmpIndArray=[x:h:x+tow];
cell{k}(i)=TmpPArray(end)*exp(-1*(((x+tow)/eta)^beta-
(x/eta)^beta));
for r=1:numel(TmpIndArray);
rf=a*exp(-b*TmpIndArray(r));
TmpPArray(r)=(TmpPArray(r)*rf)*(beta/eta)*((TmpIndArray(r)/eta)^(beta-
1))*exp(-1*((TmpIndArray(r)/eta)^beta-(x/eta)^beta));
end;
cell{k}(i)=cell{k}(i)+h*(TmpPArray(1)+4*sum(TmpPArray(2:2:(numel(TmpPArray)-
1)))+2*sum(TmpPArray(3:2:(numel(TmpPArray)-
2)))+TmpPArray(numel(TmpPArray)))/3;
end; % if k==1
APPENDIX Q 260
cell{k}(1)
if k > 2
checkterm= cell{k}(1)*V2 /(1-V2);
if checkterm < epsilon
ContWhile=false;
Sumterms=0;
for cnt=1:1:k;
Sumterms=Sumterms+cell{cnt}(1);
end;
k
intout =Sumterms
end
end
k=k+1;
end % while
toc;
end % function
function first_func=FirstFunction(x,tow,beta,eta,a,b)
if x==0
F=@(y) (a*(beta/eta).*(y/eta).^(beta-1)).*exp((-b.*y)-(y/eta).^beta);
else
F=@(y) (a*(beta/eta).*((x+y)/eta).^(beta-1)).*exp((-b.*(x+y))-
((x+y)/eta).^beta+(x/eta).^beta);
end;
first_func=quad(F,0,tow);
end
function intout
=Simp_int_integral_GeneralInfinite_e(tow,beta,eta,h,a,b,epsilon,V2)
…
…
…
end % function
function first_func=FirstFunction(x,tow,beta,eta)
if x==0
F=@(y) exp(-1.*(y/eta).^beta);
else
F=@(y) exp(-1.*(((y+x)/eta).^beta-(x/eta).^beta));
end;
first_func=tow-quad(F,0,tow);
end
% Function GCalculator_Minimal_Hard_improve3(totaltime,tau,t,beta1,eta1,s1,
% beta2,eta2,s2,beta3,eta3,h) is the function which its performance has
% been improved to calculate faster the expected number of minimal repairs
% of a component with soft failure in the case of opportunistic
% inspections at hard failures. This function calculates recursively
% the expected value of minimal repairs. Simpson's rule is used to
% calculate the multi-dimenstional intergrals, however, to have a good
% precision, smaller value of h is required which needs a long execution
% time for this function.
function intout
=GCalculator_Minimal_Hard_improve3(totaltime,tau,t,beta1,eta1,s1,beta2,eta2,s
2,beta3,eta3,h)
function [out1,out2]=makeFxMatrix(l,Tauh)
expfX=zeros(Tauh,l+1);
fX=zeros(Tauh,l+1);
j = 0:1:l;
for i=0:Tauh;
iarray=repmat(i,1,l+1);
expfX(i+1,:)= exp((j*h/eta1).^beta1-((iarray+j)*h/eta1).^beta1);
fX(i+1,:)=(beta1/eta1)*((iarray+j)*h/eta1).^(beta1-1).*expfX(i+1,:);
end;
out1=fX;
out2=expfX;
end;
function [out1,out2]=makeFzMatrix(l,Tauh)
expfZ=zeros(l,l);
fZ=zeros(l,l);
APPENDIX Q 262
j = 0:1:l-1;
for i=1:l;
iarray=repmat(i,1,l);
expfZ(i,:)=exp(((j*h)/eta2).^beta2-
(((iarray+j)*h)/eta2).^beta2+((j*h)/eta3).^beta3-
(((iarray+j)*h)/eta3).^beta3);
fZ(i,:)= ((beta2/eta2)*(((iarray+j)*h)/eta2).^(beta2-
1)+(beta3/eta3)*(((iarray+j)*h)/eta3).^(beta3-1)).*expfZ(i,:);
end;
out1=fZ;
out2=expfZ;
end;
function out=Gk(k,inspinterval,t,s1,s2,Gi,Gj,h)
function out=CaldblIntegral(t,s1,s2,i,Gi,j)
Func_Z=zeros(1,i+1);
Func_X=zeros(1,i+1);
for i1=1:i;
if (i1==i)
for i2=0:i1;
Func_X(i2+1)=fXmatrix(i2+1,j+1);
end;
else
for i2=0:i1;
Func_X(i2+1)=(1+M1(i-i1,i2+1))*fXmatrix(i2+1,j+1);
end;
end
switch i1+1
case 2
Func_Z(i1+1)=(h/2)*(Func_X(1)+Func_X(2));
case 3
Func_Z(i1+1)=(h/3)*(Func_X(1)+4*Func_X(2)+Func_X(3));
case 4
Func_Z(i1+1)=(3*h/8)*(Func_X(1)+3*Func_X(2)+3*Func_X(3)+Func_X(4));
otherwise
if mod(i1+1,2)==0 %even
Func_Z(i1+1)=h/3*(Func_X(1)+4*sum(Func_X(2:2:(i1)))+2*sum(Func_X(3:2:(i1-
1)))+Func_X(i1+1));
else % odd
Func_Z(i1+1)=h/3*(Func_X(1)+4*sum(Func_X(2:2:(i1-
3)))+2*sum(Func_X(3:2:(i1-4)))+Func_X(i1-2))+(3*h/8)*(Func_X(i1-
2)+3*Func_X(i1-1)+3*Func_X(i1)+Func_X(i1+1));
end;
end
Func_Z(i1+1)=Func_Z(i1+1)*fZmatrix(i1,Gi);
end;
APPENDIX Q 263
out=Simpson(Func_Z,h);
end;
function out=CalSingleIntegralGk(t,s1,s2,l,Gi,j,h)
Func_Z=zeros(1,l+1);
for i=1:l-1;
Func_Z(i+1)=M1(l-i,i)*fZmatrix(i,Gi)*expfXmatrix(i+1,j+1);
end;
out=Simpson(Func_Z,h);
end;
function out=CalNoIntegralGk(k,t,s1,s2,l,Gi,Gj,h)
Func_X=zeros(1,l+1);
for j=0:l;
Func_X(j+1)=(1+G(k-1,j+1))*fXmatrix(j+1,Gj+1);
end;
switch l+1
case 2
out1=(h/2)*(Func_X(1)+Func_X(2));
case 3
out1=(h/3)*(Func_X(1)+4*Func_X(2)+Func_X(3));
case 4
out1=(3*h/8)*(Func_X(1)+3*Func_X(2)+3*Func_X(3)+Func_X(4));
otherwise
if mod(l+1,2)==0 %even
out1=h/3*(Func_X(1)+4*sum(Func_X(2:2:(l)))+2*sum(Func_X(3:2:(l-
1)))+Func_X(l+1));
else % odd
out1=h/3*(Func_X(1)+4*sum(Func_X(2:2:(l-
3)))+2*sum(Func_X(3:2:(l-4)))+Func_X(l-2))+(3*h/8)*(Func_X(l-2)+3*Func_X(l-
1)+3*Func_X(l)+Func_X(l+1));
end;
end
PZSigma_given_S=expfZmatrix(l,Gi);
out1=out1*PZSigma_given_S;
PXSigma_given_t=expfXmatrix(l+1,Gj+1);
out2=PZSigma_given_S*PXSigma_given_t*(G(k-1,l+1));
out=out1+out2;
APPENDIX Q 264
end;
if k==1
out=G1(inspinterval,t,s1,s2,Gi,Gj,h);
else
l=inspinterval/h;
M1=zeros(l,l);
for i=1:l;
for j=0:l-i;
M1(i,j+1)=CaldblIntegral(t+j*h,s1+(l-i)*h,s2+(l-i)*h,i,Gi+l-
i+1,Gj+j);
end;
end;
out=M1(l,1);
SingleIntegral=CalSingleIntegralGk(t,s1,s2,l,Gi+1,Gj,h);
NoIntegral=CalNoIntegralGk(k,t,s1,s2,l,Gi+1,Gj,h);
out=out+SingleIntegral+NoIntegral;
Denominator=((beta2/eta2)*(s1/eta2)^(beta2-
1))+((beta3/eta3)*(s2/eta3)^(beta3-1));
switch l+1
case 2
out=out/(1-Denominator*h/2);
case 3
out=out/(1-Denominator*h/3);
case 4
out=out/(1-Denominator*h*3/8);
otherwise
out=out/(1-Denominator*h/3);
end;
end;
end;
function out=Si(funccall,x,z)
switch funccall
case 1
if x < z
APPENDIX Q 265
out=1;
else
out=0;
end;
otherwise
end;
end;
Simps=(3*h/8)*(InpArray(1)+3*InpArray(2)+3*InpArray(3)+InpArray(4));
otherwise
if mod(Dimension,2)==0 %even
Simps=h/3*(InpArray(1)+4*sum(InpArray(2:2:(Dimension-
1)))+2*sum(InpArray(3:2:(Dimension-2)))+InpArray(Dimension));
else % odd
Simps=h/3*(InpArray(1)+4*sum(InpArray(2:2:(Dimension-
4)))+2*sum(InpArray(3:2:(Dimension-5)))+InpArray(Dimension-
3))+(3*h/8)*(InpArray(Dimension-3)+3*InpArray(Dimension-
2)+3*InpArray(Dimension-1)+InpArray(Dimension));
end;
end
end;
function out=G1(sigma,t,s1,s2,Gi,Gj,h)
function out=CaldblIntegral(t,s1,s2,i,Gi,Gj,h)
Func_Z=zeros(1,i+1);
Func_X=zeros(1,i+1);
for i1=1:i;
if (i1==i)
for i2=0:i1;
Func_X(i2+1)=fXmatrix(i2+1,Gj+1);
end;
else
for i2=0:i1;
Func_X(i2+1)=(1+M1(i-i1,i2+1))*fXmatrix(i2+1,Gj+1);
end;
end
switch i1+1
case 2
Func_Z(i1+1)=(h/2)*(Func_X(1)+Func_X(2));
case 3
APPENDIX Q 266
Func_Z(i1+1)=(h/3)*(Func_X(1)+4*Func_X(2)+Func_X(3));
case 4
Func_Z(i1+1)=(3*h/8)*(Func_X(1)+3*Func_X(2)+3*Func_X(3)+Func_X(4));
otherwise
if mod(i1+1,2)==0 %even
Func_Z(i1+1)=h/3*(Func_X(1)+4*sum(Func_X(2:2:(i1)))+2*sum(Func_X(3:2:(i1-
1)))+Func_X(i1+1));
else % odd
Func_Z(i1+1)=h/3*(Func_X(1)+4*sum(Func_X(2:2:(i1-
3)))+2*sum(Func_X(3:2:(i1-4)))+Func_X(i1-2))+(3*h/8)*(Func_X(i1-
2)+3*Func_X(i1-1)+3*Func_X(i1)+Func_X(i1+1));
end;
end
Func_Z(i1+1)=Func_Z(i1+1)*fZmatrix(i1,Gi);
end;
out=Simpson(Func_Z,h);
end;
function out=CalSingleIntegral(t,s1,s2,l,Gi,j,h)
Func_Z=zeros(1,l+1);
for i=1:l-1;
Func_Z(i+1)=M1(l-i,i)*fZmatrix(i,Gi)*expfXmatrix(i+1,j+1);
end;
out=Simpson(Func_Z,h);
end;
function out=CalNoIntegral(sigma,t,s1,s2,Gi,h)
F=@(x) (beta1/eta1).*((t+x)/eta1).^(beta1-1).*exp((t/eta1).^beta1-
((t+x)/eta1).^beta1);
out=quad(F,0,sigma);
out=out*expfZmatrix(sigma/h,Gi);
end;
l=sigma/h;
M1=zeros(l,l);
for i=1:l;
for j=0:l-i;
M1(i,j+1)=CaldblIntegral(t+j*h,s1+(l-i)*h,s2+(l-i)*h,i,Gi+l-
i+1,Gj+j,h);
end;
end;
SingleIntegral=CalSingleIntegral(t,s1,s2,l,Gi+1,Gj,h);
NoIntegral=CalNoIntegral(sigma,t,s1,s2,Gi+1,h);
APPENDIX Q 267
out=M1(l,1)+SingleIntegral+NoIntegral;
Denominator= ((beta2/eta2)*(s1/eta2)^(beta2-
1))+((beta3/eta3)*(s2/eta3)^(beta3-1));
switch l+1
case 2
out=out/(1-Denominator*h/2);
case 3
out=out/(1-Denominator*h/3);
case 4
out=out/(1-Denominator*h*3/8);
otherwise
out=out/(1-Denominator*h/3);
end;
end;
% for k
tStart=tic;
k=fix(totaltime/tau);
sigma=rem(totaltime,tau);
if sigma > 0
k=k+1;
end
l3=tau/h;
[fXmatrix,expfXmatrix]=makeFxMatrix(totaltime/h,l3);
[fZmatrix,expfZmatrix]=makeFzMatrix(totaltime/h,l3);
G=zeros(k,(k-1)*l3);
for i3=1:k;
if sigma > 0 && i3==1
inspinterval=sigma;
else
inspinterval=tau;
end;
for j3=0:(k-i3)*l3;
G(i3,j3+1)= Gk(i3,inspinterval,t+j3*h,s1+(k-i3)*tau,s2+(k-i3)*tau,(k-
i3)*l3,j3,h);
end;
end;
APPENDIX Q 268
xlswrite('C:\opportunistic
maintenance\matlab\Opportunistic\bothMinimal\Excel\Out.xls',G);
intout=G(k,1);
tElapsed=toc(tStart)
end
function out=RunAll()
TotalTime=12;
for Tau=TotalTime:-1:1;
SO = SimulationOpportunisticMRSoftFile(TotalTime,Tau)
end
out=Tau;
end
function SO = SimulationOpportunisticMRSoftFile(TotalTime,Tau)
function Out=OneSimulationRun()
NextFailuresSoft=zeros(1,SoftCompNo);
CurrentAgesSoft=zeros(1,SoftCompNo);
NextFailuresHard=zeros(1,HardCompNo);
CurrentAgesHard=zeros(1,HardCompNo);
for i=1:SoftCompNo;
z= random('uniform',0,1);
APPENDIX Q 269
NextFailuresSoft(i)=SoftParams(i,2)*power(power(CurrentAgesSoft(i)/SoftParams
(i,2),SoftParams(i,1))-log(z),1/SoftParams(i,1))-CurrentAgesSoft(i);
i=i+1;
end;
for i=1:HardCompNo;
z= random('uniform',0,1);
NextFailuresHard(i)=HardParams(i,2)*power(power(CurrentAgesHard(i)/HardParams
(i,2),HardParams(i,1))-log(z),1/HardParams(i,1))-CurrentAgesHard(i);
i=i+1;
end;
CurrentTime=0;
NextInspection=Tau;
OutSoft(i,3)=OutSoft(i,3)+NextFailuresSoft(i);
rf=SoftParams(i,3)*exp(-SoftParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextFailuresSoft(i);
else % replacement
OutSoft(i,2)=OutSoft(i,2)+1;
CurrentAgesSoft(i)=0;
end
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,3)=OutSoft(i,3)+MinHard;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+MinHard;
NextFailuresSoft(i)=NextFailuresSoft(i)-MinHard;
end
end
CurrentTime=CurrentTime+MinHard;
CurrentAgesHard=CurrentAgesHard+MinHard;
APPENDIX Q 270
NextfailureofHardfailed=GenerateNextFailure(HardParams(MinIdx,1),HardParams(M
inIdx,2),CurrentAgesHard(MinIdx));
NextFailuresHard=NextFailuresHard-MinHard;
NextFailuresHard(MinIdx)=NextfailureofHardfailed;
NextInspection=NextInspection-MinHard;
[MinHard,MinIdx]=min(NextFailuresHard);
end
for i=1:SoftCompNo;
if NextFailuresSoft(i) < NextInspection
rf=SoftParams(i,3)*exp(-SoftParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
OutSoft(i,3)=OutSoft(i,3)+NextFailuresSoft(i);
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,3)=OutSoft(i,3)+NextInspection;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextInspection;
NextFailuresSoft(i)=NextFailuresSoft(i)-NextInspection;
end
end
CurrentAgesHard=CurrentAgesHard+NextInspection;
CurrentTime=CurrentTime+NextInspection;
NextFailuresHard=NextFailuresHard-NextInspection;
if CurrentTime - Tau*(CurrentTime/Tau)== 0
Sigma=TotalTime-CurrentTime;
if Sigma < Tau
NextInspection=Sigma;
else
NextInspection=Tau;
end
end
end
Out=OutSoft;
APPENDIX Q 271
end
function nextfailure=GenerateNextFailure(beta,eta,CurrentAge)
z= random('uniform',0,1);
nextfailure=eta*power(power(CurrentAge/eta,beta)-log(z),1/beta)-
CurrentAge;
end
Data = xlsread('C:\opportunistic
maintenance\matlab\Opportunistic\Simulation\Minimal_Replacement\Excel\SoftCom
ponents.xls');
RecordsNo=numel(Data(:,2));
rowindx=0;
SoftCompNo=rowindx;
Data = xlsread('C:\opportunistic
maintenance\matlab\Opportunistic\Simulation\Minimal_Replacement\Excel\HardCom
ponents.xls');
RecordsNo=numel(Data(:,2));
rowindx=0;
HardCompNo=rowindx;
SimOut=zeros(SoftCompNo,3);
SimulationRuns=1000000;
APPENDIX Q 272
MinimalMatrix=zeros(SimulationRuns,SoftCompNo);
ReplacementMatrix=zeros(SimulationRuns,SoftCompNo);
UpMatrix=zeros(SimulationRuns,SoftCompNo);
for iteration=1:SimulationRuns;
Out=OneSimulationRun();
MinimalMatrix(iteration,:)=Out(:,1)';
ReplacementMatrix(iteration,:)=Out(:,2)';
UpMatrix(iteration,:)=Out(:,3)';
SimOut=SimOut+Out;
end;
for i=1:SoftCompNo;
AvgSTDMinimals(i,1)=mean(MinimalMatrix(:,i));
AvgSTDMinimals(i,2)=std(MinimalMatrix(:,i));
AvgSTDMinimals(i,3)=std(MinimalMatrix(:,i))/sqrt(SimulationRuns);
AvgSTDReplacement(i,1)=mean(ReplacementMatrix(:,i));
AvgSTDReplacement(i,2)=std(ReplacementMatrix(:,i));
AvgSTDReplacement(i,3)=std(ReplacementMatrix(:,i))/sqrt(SimulationRuns);
AvgSTDUp(i,1)=mean(UpMatrix(:,i));
AvgSTDUp(i,2)=std(UpMatrix(:,i));
AvgSTDUp(i,3)=std(UpMatrix(:,i))/sqrt(SimulationRuns);
end;
SimOut=SimOut/SimulationRuns;
SO=SimOut;
f1=strcat('AvgSTDMinimals-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement','Excel', f1);
xlswrite(f,AvgSTDMinimals);
f1=strcat('AvgSTDReplacement-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement','Excel', f1);
xlswrite(f,AvgSTDReplacement);
f1=strcat('AvgSTDUp-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement','Excel', f1);
xlswrite(f,AvgSTDUp);
end
APPENDIX Q 273
% Function SimulationOpportunisticFile(TotalTime,Tau)
% is similar to function SimulationOpportunisticMRSoftFile, except that
% it runs the simulation model for the opportunistic case with the
% assumption of minimal repair of the components with soft and hard
% failures.
% The outputs which are the expected number of minimal repairs and uptime
% for all components with soft failures are exported to the excel files
% The main difference in the code of this function is in function
% OneSimulationRun()
function SO = SimulationOpportunisticFile(TotalTime,Tau)
function Out=OneSimulationRun()
NextFailuresSoft=zeros(1,SoftCompNo);
CurrentAgesSoft=zeros(1,SoftCompNo);
NextFailuresHard=zeros(1,HardCompNo);
CurrentAgesHard=zeros(1,HardCompNo);
for i=1:SoftCompNo;
z= random('uniform',0,1);
NextFailuresSoft(i)=SoftParams(i,2)*power(power(CurrentAgesSoft(i)/SoftParams
(i,2),SoftParams(i,1))-log(z),1/SoftParams(i,1))-CurrentAgesSoft(i);
i=i+1;
end;
for i=1:HardCompNo;
z= random('uniform',0,1);
NextFailuresHard(i)=HardParams(i,2)*power(power(CurrentAgesHard(i)/HardParams
(i,2),HardParams(i,1))-log(z),1/HardParams(i,1))-CurrentAgesHard(i);
i=i+1;
end;
CurrentTime=0;
NextInspection=Tau;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextFailuresSoft(i);
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,2)=OutSoft(i,2)+MinHard;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+MinHard;
NextFailuresSoft(i)=NextFailuresSoft(i)-MinHard;
end
end
CurrentTime=CurrentTime+MinHard;
CurrentAgesHard=CurrentAgesHard+MinHard;
NextfailureofHardfailed=GenerateNextFailure(HardParams(MinIdx,1),HardParams(M
inIdx,2),CurrentAgesHard(MinIdx));
NextFailuresHard=NextFailuresHard-MinHard;
NextFailuresHard(MinIdx)=NextfailureofHardfailed;
NextInspection=NextInspection-MinHard;
[MinHard,MinIdx]=min(NextFailuresHard);
end
for i=1:SoftCompNo;
if NextFailuresSoft(i) < NextInspection
OutSoft(i,1)=OutSoft(i,1)+1;
OutSoft(i,2)=OutSoft(i,2)+NextFailuresSoft(i);
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextFailuresSoft(i);
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,2)=OutSoft(i,2)+NextInspection;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextInspection;
NextFailuresSoft(i)=NextFailuresSoft(i)-NextInspection;
end
end
CurrentAgesHard=CurrentAgesHard+NextInspection;
CurrentTime=CurrentTime+NextInspection;
NextFailuresHard=NextFailuresHard-NextInspection;
if mod(CurrentTime,Tau)== 0
Sigma=TotalTime-CurrentTime;
if Sigma < Tau
NextInspection=Sigma;
else
NextInspection=Tau;
end
end
end
Out=OutSoft;
APPENDIX Q 275
end
end
function SO = SimulationOpportunisticMRSoftHardFile(TotalTime,Tau)
function [Out1,Out2]=OneSimulationRun()
NextFailuresSoft=zeros(1,SoftCompNo);
CurrentAgesSoft=zeros(1,SoftCompNo);
NextFailuresHard=zeros(1,HardCompNo);
CurrentAgesHard=zeros(1,HardCompNo);
for i=1:SoftCompNo;
z= random('uniform',0,1);
NextFailuresSoft(i)=SoftParams(i,2)*power(power(CurrentAgesSoft(i)/SoftParams
(i,2),SoftParams(i,1))-log(z),1/SoftParams(i,1))-CurrentAgesSoft(i);
i=i+1;
end;
for i=1:HardCompNo;
z= random('uniform',0,1);
NextFailuresHard(i)=HardParams(i,2)*power(power(CurrentAgesHard(i)/HardParams
(i,2),HardParams(i,1))-log(z),1/HardParams(i,1))-CurrentAgesHard(i);
i=i+1;
end;
CurrentTime=0;
NextInspection=Tau;
APPENDIX Q 276
OutSoft(i,3)=OutSoft(i,3)+NextFailuresSoft(i);
rf=SoftParams(i,3)*exp(-SoftParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextFailuresSoft(i);
else % replacement
OutSoft(i,2)=OutSoft(i,2)+1;
CurrentAgesSoft(i)=0;
end
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,3)=OutSoft(i,3)+MinHard;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+MinHard;
NextFailuresSoft(i)=NextFailuresSoft(i)-MinHard;
end
end
CurrentTime=CurrentTime+MinHard;
CurrentAgesHard=CurrentAgesHard+MinHard;
rf=HardParams(MinIdx,3)*exp(-
HardParams(MinIdx,4)*CurrentAgesHard(MinIdx));
repairz= random('uniform',0,1);
NextfailureofHardfailed=GenerateNextFailure(HardParams(MinIdx,1),HardParams(M
inIdx,2),CurrentAgesHard(MinIdx));
NextFailuresHard=NextFailuresHard-MinHard;
NextFailuresHard(MinIdx)=NextfailureofHardfailed;
NextInspection=NextInspection-MinHard;
[MinHard,MinIdx]=min(NextFailuresHard);
end
for i=1:SoftCompNo;
if NextFailuresSoft(i) < NextInspection
rf=SoftParams(i,3)*exp(-SoftParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
OutSoft(i,3)=OutSoft(i,3)+NextFailuresSoft(i);
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,3)=OutSoft(i,3)+NextInspection;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextInspection;
NextFailuresSoft(i)=NextFailuresSoft(i)-NextInspection;
end
end
CurrentAgesHard=CurrentAgesHard+NextInspection;
CurrentTime=CurrentTime+NextInspection;
NextFailuresHard=NextFailuresHard-NextInspection;
if CurrentTime - Tau*(CurrentTime/Tau)== 0
Sigma=TotalTime-CurrentTime;
if Sigma < Tau
NextInspection=Sigma;
else
NextInspection=Tau;
end
end
end
Out1=OutSoft;
Out2=OutHard;
APPENDIX Q 278
end
function nextfailure=GenerateNextFailure(beta,eta,CurrentAge)
z= random('uniform',0,1);
nextfailure=eta*power(power(CurrentAge/eta,beta)-log(z),1/beta)-
CurrentAge;
end
Data = xlsread('C:\opportunistic
maintenance\matlab\Opportunistic\Simulation\Minimal_Replacement_Hard\Excel\So
ftComponents.xls');
RecordsNo=numel(Data(:,2));
rowindx=0;
SoftCompNo=rowindx;
Data = xlsread('C:\opportunistic
maintenance\matlab\Opportunistic\Simulation\Minimal_Replacement_Hard\Excel\Ha
rdComponents.xls');
RecordsNo=numel(Data(:,2));
rowindx=0;
HardCompNo=rowindx;
SimOutSoft=zeros(SoftCompNo,3);
SimOutHard=zeros(HardCompNo,2);
APPENDIX Q 279
SimulationRuns=1000000;
MinimalMatrix=zeros(SimulationRuns,SoftCompNo);
ReplacementMatrix=zeros(SimulationRuns,SoftCompNo);
UpMatrix=zeros(SimulationRuns,SoftCompNo);
MinimalMatrixHard=zeros(SimulationRuns,HardCompNo);
ReplacementMatrixHard=zeros(SimulationRuns,HardCompNo);
for iteration=1:SimulationRuns;
[out1,out2]=OneSimulationRun();
MinimalMatrix(iteration,:)=out1(:,1)';
ReplacementMatrix(iteration,:)=out1(:,2)';
UpMatrix(iteration,:)=out1(:,3)';
SimOutSoft=SimOutSoft+out1;
MinimalMatrixHard(iteration,:)=out2(:,1)';
ReplacementMatrixHard(iteration,:)=out2(:,2)';
SimOutHard=SimOutHard+out2;
end;
for i=1:SoftCompNo;
AvgSTDMinimals(i,1)=mean(MinimalMatrix(:,i));
AvgSTDMinimals(i,2)=std(MinimalMatrix(:,i));
AvgSTDMinimals(i,3)=std(MinimalMatrix(:,i))/sqrt(SimulationRuns);
AvgSTDReplacement(i,1)=mean(ReplacementMatrix(:,i));
AvgSTDReplacement(i,2)=std(ReplacementMatrix(:,i));
AvgSTDReplacement(i,3)=std(ReplacementMatrix(:,i))/sqrt(SimulationRuns);
AvgSTDUp(i,1)=mean(UpMatrix(:,i));
AvgSTDUp(i,2)=std(UpMatrix(:,i));
AvgSTDUp(i,3)=std(UpMatrix(:,i))/sqrt(SimulationRuns);
end;
for i=1:HardCompNo;
AvgSTDMinimalsHard(i,1)=mean(MinimalMatrixHard(:,i));
AvgSTDMinimalsHard(i,2)=std(MinimalMatrixHard(:,i));
AvgSTDMinimalsHard(i,3)=std(MinimalMatrixHard(:,i))/sqrt(SimulationRuns);
AvgSTDReplacementHard(i,1)=mean(ReplacementMatrixHard(:,i));
AvgSTDReplacementHard(i,2)=std(ReplacementMatrixHard(:,i));
AvgSTDReplacementHard(i,3)=std(ReplacementMatrixHard(:,i))/sqrt(SimulationRun
s);
end;
APPENDIX Q 280
SimOutSoft=SimOutSoft/SimulationRuns;
SO=SimOutSoft;
f1=strcat('AvgSTDMinimals-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement_Hard','Excel', f1);
xlswrite(f,AvgSTDMinimals);
f1=strcat('AvgSTDReplacement-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement_Hard','Excel', f1);
xlswrite(f,AvgSTDReplacement);
f1=strcat('AvgSTDUp-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement_Hard','Excel', f1);
xlswrite(f,AvgSTDUp);
f1=strcat('AvgSTDMinimalsHard-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement_Hard','Excel', f1);
xlswrite(f,AvgSTDMinimalsHard);
f1=strcat('AvgSTDReplacementHard-tau-',num2str(Tau),'.xls');
f = fullfile('C:', 'opportunistic maintenance', 'matlab',
'Opportunistic','Simulation','Minimal_Replacement_Hard','Excel', f1);
xlswrite(f,AvgSTDReplacementHard);
end
% Function SimulationOpportunisticMRSoftHardPMFile(TotalTime,Tau)
% is similar to function SimulationOpportunisticMRSoftHardFile, except that
% it runs the simulation model for the opportunistic case with the
% additional assumption of preventive replacement of the components with
% hard failures at periodic inspections.
% The outputs which are the expected number of minimal repairs,
% replacements, and uptime for all components with soft failures, and the
% expected numbers of minimal repairs and replacements inside, and at the
% inspection intervals for hard failures are exported to the excel files
% The main addition in the code of this function is in function
% OneSimulationRun()
function SO = SimulationOpportunisticMRSoftHardPMFile(TotalTime,Tau)
function [Out1,Out2]=OneSimulationRun()
APPENDIX Q 281
NextFailuresSoft=zeros(1,SoftCompNo);
CurrentAgesSoft=zeros(1,SoftCompNo);
NextFailuresHard=zeros(1,HardCompNo);
CurrentAgesHard=zeros(1,HardCompNo);
for i=1:SoftCompNo;
z= random('uniform',0,1);
NextFailuresSoft(i)=SoftParams(i,2)*power(power(CurrentAgesSoft(i)/SoftParams
(i,2),SoftParams(i,1))-log(z),1/SoftParams(i,1))-CurrentAgesSoft(i);
i=i+1;
end;
for i=1:HardCompNo;
z= random('uniform',0,1);
NextFailuresHard(i)=HardParams(i,2)*power(power(CurrentAgesHard(i)/HardParams
(i,2),HardParams(i,1))-log(z),1/HardParams(i,1))-CurrentAgesHard(i);
i=i+1;
end;
CurrentTime=0;
NextInspection=Tau;
OutSoft(i,3)=OutSoft(i,3)+NextFailuresSoft(i);
rf=SoftParams(i,3)*exp(-SoftParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextFailuresSoft(i);
else % replacement
OutSoft(i,2)=OutSoft(i,2)+1;
APPENDIX Q 282
CurrentAgesSoft(i)=0;
end
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,3)=OutSoft(i,3)+MinHard;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+MinHard;
NextFailuresSoft(i)=NextFailuresSoft(i)-MinHard;
end
end
CurrentTime=CurrentTime+MinHard;
CurrentAgesHard=CurrentAgesHard+MinHard;
rf=HardParams(MinIdx,3)*exp(-
HardParams(MinIdx,4)*CurrentAgesHard(MinIdx));
repairz= random('uniform',0,1);
NextfailureofHardfailed=GenerateNextFailure(HardParams(MinIdx,1),HardParams(M
inIdx,2),CurrentAgesHard(MinIdx));
NextFailuresHard=NextFailuresHard-MinHard;
NextFailuresHard(MinIdx)=NextfailureofHardfailed;
NextInspection=NextInspection-MinHard;
[MinHard,MinIdx]=min(NextFailuresHard);
end
for i=1:SoftCompNo;
if NextFailuresSoft(i) < NextInspection
rf=SoftParams(i,3)*exp(-SoftParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
OutSoft(i,3)=OutSoft(i,3)+NextFailuresSoft(i);
OutSoft(i,1)=OutSoft(i,1)+1;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextFailuresSoft(i);
else % replacement
OutSoft(i,2)=OutSoft(i,2)+1;
CurrentAgesSoft(i)=0;
end
NextFailuresSoft(i)=GenerateNextFailure(SoftParams(i,1),SoftParams(i,2),Curre
ntAgesSoft(i));
else
OutSoft(i,3)=OutSoft(i,3)+NextInspection;
CurrentAgesSoft(i)=CurrentAgesSoft(i)+NextInspection;
NextFailuresSoft(i)=NextFailuresSoft(i)-NextInspection;
end
end
CurrentAgesHard=CurrentAgesHard+NextInspection;
CurrentTime=CurrentTime+NextInspection;
NextFailuresHard=NextFailuresHard-NextInspection;
if CurrentTime - Tau*(CurrentTime/Tau)== 0
% Applying PM for components with hard failure
for i=1:HardCompNo;
rf=HardParams(i,3)*exp(-HardParams(i,4)*CurrentAgesSoft(i));
repairz= random('uniform',0,1);
NextFailuresHard(i)=GenerateNextFailure(HardParams(i,1),HardParams(i,2),Curre
ntAgesHard(i));
end
end
Sigma=TotalTime-CurrentTime;
if Sigma < Tau
NextInspection=Sigma;
else
NextInspection=Tau;
end
end
end
Out1=OutSoft;
Out2=OutHard;
end
end
APPENDIX Q 284
function GN = generalLike()
First2=true;
First3=true;
First4=true;
tStart2=tic;
tStart3=tic;
tStart4=tic;
tStart5=tic;
Params(1,1)=-0.7;
Params(1,2)=1;
RecordsNo=numel(Data(:,3));
rowindx=1;
UnitNo=1;
aHistoryL(1,1)=Data(rowindx,2);
aHistoryU(1,1)=Data(rowindx,3);
k=2;
rowindx=rowindx+1;
while (rowindx <= RecordsNo) & (Data(rowindx,1) ~= 1)
aHistoryL(1,k)=Data(rowindx,2);
aHistoryU(1,k)=Data(rowindx,3);
k=k+1;
rowindx=rowindx+1;
end
L{UnitNo}=aHistoryL(1:k-1);
U{UnitNo}=aHistoryU(1:k-1);
UnitNo=UnitNo+1;
end
APPENDIX Q 285
NoofSystems=UnitNo-1;%3;
l=1;
AllParams(l,:)=Params(1,:);
ContWhile=true;
Oncemore=0;
if Oncemore==1
ContWhile=false;
end;
%
% N(1)='Ci';
% N(2)='(yn)^beta';
% N(3)='((yn)^beta)*log(yn)';
% N(4)='((yn)^beta)*[log(yn)]^2';
Sumni=0;
for i=1:NoofSystems;
Sumni=Sumni+numel(L{i});
[C,B1,B2,B3,B4] =
Simp_int_integral_GeneralLikelihoodX(numel(L{i}),Params(1,2),Params(1,1),0.00
1,L{i},U{i});
N(1,i)=C;
N(2,i)=B1;
N(3,i)=B2;
N(4,i)=B3;
N(5,i)=B4;
end
% Construct Q' array
Q_Prime(1,1)=Sumni-exp(Params(1,1))*sum(N(2,:));
Q_Prime(2,1)=(1/Params(1,2))*Sumni+sum(N(5,:))-exp(Params(1,1))*sum(N(3,:));
NewParams=Params'-inv(Q_Second)*Q_Prime
l=l+1;
APPENDIX Q 286
AllParams(l,1)=NewParams(1,1);
AllParams(l,2)=NewParams(2,1);
AllParams(l,3)=abs(NewParams(1,1)-Params(1,1));
AllParams(l,4)=abs(Params(1,2)-NewParams(2,1));
AllParams(l,5)=Q_Prime(1,1);
AllParams(l,6)=Q_Prime(2,1);
if NewParams(2,1) < 0
NewParams(2,1)=0.1;
end
Params(1,:)=NewParams(:,1)';
end; % while
GN=N;
xlswrite('C:\Paper 5 new\excel\NewSet\OneStep\Expected-
BatterySubset3.xls',N);
xlswrite('C:\Paper 5 new\excel\NewSet\OneStep\AllParams-
BatterySubset3.xls',AllParams);
end
% here are C1, B1, B2, B3, and B4 with the following definitions
% C1 1
% B1 (yn)^beta
% B2 ((yn)^beta)*log(yn)
% B3 ((yn)^beta)*[log(yn)]^2
% B4 Sum(yi) i=1,...,n
function [C,B1,B2,B3,B4]
=Simp_int_integral_GeneralLikelihoodX(k,beta0,alpha0,h,L,U)
beta1=beta0;
alpha1=alpha0;
%{
%
Ccell:
C1 1
Bcell:
{1} B1 (yn)^beta
{2} B2 ((yn)^beta)*log(yn)
{3} B3 ((yn)^beta)*[log(yn)]^2
{4} B4 Sum(yi) i=1,...,n
%}
if k==1
intout(1)=FirstFunctionC(0,L(k),U(k),beta0,alpha0);
intout(2)=FirstFunctionB1(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
intout(3)=FirstFunctionB2(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
intout(4)=FirstFunctionB3(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
intout(5)=FirstFunctionB4(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
else
for j=k-1:-1:1;
i=0;
startcon=sum(L(1:j));
stopcon=sum(U(1:j));
for x=startcon:h:stopcon;
i=i+1;
if j==k-1
IndArray(i)=x;
Ccell{1}(i)= FirstFunctionC(x,L(j+1),U(j+1),beta0,alpha0);
Bcell{1}(i)=
FirstFunctionB1(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
Bcell{2}(i)=
FirstFunctionB2(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
Bcell{3}(i)=
FirstFunctionB3(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
APPENDIX Q 288
Bcell{4}(i)=
FirstFunctionB4(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
else
TmpPArrayC=Ccell{1}(INDICES);
TmpPArrayB1=Bcell{1}(INDICES);
TmpPArrayB2=Bcell{2}(INDICES);
TmpPArrayB3=Bcell{3}(INDICES);
TmpPArrayB4=Bcell{4}(INDICES);
for r=1:numel(TmpIndArray);
TmpPArrayB4(r)=(log(TmpIndArray(r))*TmpPArrayC(r)+TmpPArrayB4(r))*(beta0*exp(
alpha0))*(TmpIndArray(r)^(beta0-1))*exp(exp(alpha0)*(x^beta0-
TmpIndArray(r)^beta0));
TmpPArrayC(r)=TmpPArrayC(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
TmpPArrayB1(r)=TmpPArrayB1(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
TmpPArrayB2(r)=TmpPArrayB2(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
TmpPArrayB3(r)=TmpPArrayB3(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
end;
Ccell{1}(i)=h*(TmpPArrayC(1)+4*sum(TmpPArrayC(2:2:(numel(TmpPArrayC)-
1)))+2*sum(TmpPArrayC(3:2:(numel(TmpPArrayC)-
2)))+TmpPArrayC(numel(TmpPArrayC)))/3;
Bcell{1}(i)=h*(TmpPArrayB1(1)+4*sum(TmpPArrayB1(2:2:(numel(TmpPArrayB1)-
1)))+2*sum(TmpPArrayB1(3:2:(numel(TmpPArrayB1)-
2)))+TmpPArrayB1(numel(TmpPArrayB1)))/3;
Bcell{2}(i)=h*(TmpPArrayB2(1)+4*sum(TmpPArrayB2(2:2:(numel(TmpPArrayB2)-
1)))+2*sum(TmpPArrayB2(3:2:(numel(TmpPArrayB2)-
2)))+TmpPArrayB2(numel(TmpPArrayB2)))/3;
Bcell{3}(i)=h*(TmpPArrayB3(1)+4*sum(TmpPArrayB3(2:2:(numel(TmpPArrayB3)-
1)))+2*sum(TmpPArrayB3(3:2:(numel(TmpPArrayB3)-
2)))+TmpPArrayB3(numel(TmpPArrayB3)))/3;
Bcell{4}(i)=h*(TmpPArrayB4(1)+4*sum(TmpPArrayB4(2:2:(numel(TmpPArrayB4)-
APPENDIX Q 289
1)))+2*sum(TmpPArrayB4(3:2:(numel(TmpPArrayB4)-
2)))+TmpPArrayB4(numel(TmpPArrayB4)))/3;
IndArray(i)=x;
end; % if j==k-1
end; % for x=startcon:h:stopcon;
TmpPArrayC=Ccell{1}(INDICES);
TmpPArrayB1=Bcell{1}(INDICES);
TmpPArrayB2=Bcell{2}(INDICES);
TmpPArrayB3=Bcell{3}(INDICES);
TmpPArrayB4=Bcell{4}(INDICES);
TmpIndArray=IndArray(INDICES);
for r=1:numel(TmpIndArray);
TmpPArrayB4(r)=(log(TmpIndArray(r))*TmpPArrayC(r)+TmpPArrayB4(r))*(beta0*exp(
alpha0))*(TmpIndArray(r)^(beta0-1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayC(r)=TmpPArrayC(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayB1(r)=TmpPArrayB1(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayB2(r)=TmpPArrayB2(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayB3(r)=TmpPArrayB3(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
end
intout(1)=h*(TmpPArrayC(1)+4*sum(TmpPArrayC(2:2:(numel(TmpPArrayC)-
1)))+2*sum(TmpPArrayC(3:2:(numel(TmpPArrayC)-
2)))+TmpPArrayC(numel(TmpPArrayC)))/3;
intout(2)=h*(TmpPArrayB1(1)+4*sum(TmpPArrayB1(2:2:(numel(TmpPArrayB1)-
1)))+2*sum(TmpPArrayB1(3:2:(numel(TmpPArrayB1)-
2)))+TmpPArrayB1(numel(TmpPArrayB1)))/3;
intout(3)=h*(TmpPArrayB2(1)+4*sum(TmpPArrayB2(2:2:(numel(TmpPArrayB2)-
1)))+2*sum(TmpPArrayB2(3:2:(numel(TmpPArrayB2)-
2)))+TmpPArrayB2(numel(TmpPArrayB2)))/3;
APPENDIX Q 290
intout(4)=h*(TmpPArrayB3(1)+4*sum(TmpPArrayB3(2:2:(numel(TmpPArrayB3)-
1)))+2*sum(TmpPArrayB3(3:2:(numel(TmpPArrayB3)-
2)))+TmpPArrayB3(numel(TmpPArrayB3)))/3;
intout(5)=h*(TmpPArrayB4(1)+4*sum(TmpPArrayB4(2:2:(numel(TmpPArrayB4)-
1)))+2*sum(TmpPArrayB4(3:2:(numel(TmpPArrayB4)-
2)))+TmpPArrayB4(numel(TmpPArrayB4)))/3;
end % if k==1
C=intout(1);
B1=intout(2)/C;
B2=intout(3)/C;
B3=intout(4)/C;
B4=intout(5)/C;
end
function first_func=FirstFunctionC(prev_y,lower,upper,beta0,alpha0)
if prev_y==0
F=@(x) (beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x) (beta0.*exp(alpha0)).*((prev_y+x).^(beta0-
1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
% for B1(0)
function
first_func=FirstFunctionB1(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) (x.^beta1).*(beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x) ((prev_y+x).^beta1).*(beta0.*exp(alpha0)).*((prev_y+x).^(beta0-
1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
APPENDIX Q 291
function
first_func=FirstFunctionB2(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) (x.^beta1).*log(x).*(beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x)
((prev_y+x).^beta1).*log(prev_y+x).*(beta0.*exp(alpha0)).*((prev_y+x).^(beta0
-1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function
first_func=FirstFunctionB3(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) (x.^beta1).*(log(x).^2).*(beta0.*exp(alpha0)).*(x.^(beta0-
1)).*exp(-1.*exp(alpha0).*x.^beta0);
else
F=@(x)
((prev_y+x).^beta1).*(log(prev_y+x).^2).*(beta0.*exp(alpha0)).*((prev_y+x).^(
beta0-1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function
first_func=FirstFunctionB4(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) log(x).*(beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x) log(prev_y+x).*(beta0.*exp(alpha0)).*((prev_y+x).^(beta0-
1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
APPENDIX Q 292
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function GN = generalLikeEM()
First2=true;
First3=true;
First4=true;
tStart2=tic;
tStart3=tic;
tStart4=tic;
tStart5=tic;
Params(1,1)=-0.7;
Params(1,2)=1;
RecordsNo=numel(Data(:,3));
rowindx=1;
UnitNo=1;
aHistoryL(1,1)=Data(rowindx,2);
aHistoryU(1,1)=Data(rowindx,3);
k=2;
rowindx=rowindx+1;
while (rowindx <= RecordsNo) & (Data(rowindx,1) ~= 1)
aHistoryL(1,k)=Data(rowindx,2);
aHistoryU(1,k)=Data(rowindx,3);
k=k+1;
rowindx=rowindx+1;
end
L{UnitNo}=aHistoryL(1:k-1);
U{UnitNo}=aHistoryU(1:k-1);
APPENDIX Q 293
UnitNo=UnitNo+1;
end
NoofSystems=UnitNo-1;
l=1;
AllParams(l,:)=Params(1,:);
ContWhile=true;
beta0=Params(1,2);
alpha0=Params(1,1);
beta1=beta0;
alpha1=alpha0;
Oncemore=0;
while ContWhile %ContWhile
if Oncemore==1
ContWhile=false;
end;
Params(1,2)=beta1;
Params(1,1)=alpha1;
% N(1)='Ci';
% N(2)='(yn)^beta';
% N(3)='((yn)^beta)*log(yn)';
% N(4)='((yn)^beta)*[log(yn)]^2';
Sumni=0;
for i=1:NoofSystems;
Sumni=Sumni+numel(L{i});
[C,B1,B2,B3,B4,B5,B6,B7,B8,B9,B10] =
Simp_int_integral_GeneralLikelihoodEMX(numel(L{i}),beta0,alpha0,beta1,alpha1,
0.001,L{i},U{i});
N(1,i)=C;
N(2,i)=B1;
N(3,i)=B2;
N(4,i)=B3;
N(5,i)=B4;
end
% Construct Q' array
APPENDIX Q 294
Q_Prime(1,1)=Sumni-exp(alpha1)*sum(N(2,:));
Q_Prime(2,1)=(1/beta1)*Sumni+sum(N(5,:))-exp(alpha1)*sum(N(3,:));
NewParams=Params'-inv(Q_Second)*Q_Prime
l=l+1;
if NewParams(2,1) < 0
NewParams(2,1)=0.1;
end
AllParams(l,1)=NewParams(1,1);
AllParams(l,2)=NewParams(2,1);
AllParams(l,7)=Q_Prime(1,1);
AllParams(l,8)=Q_Prime(2,1);
else
beta0=NewParams(2,1);
alpha0=NewParams(1,1);
beta1=beta0;
alpha1=alpha0;
end
else
AllParams(l,5)=abs(NewParams(1,1)-alpha1);
AllParams(l,6)=abs(beta1-NewParams(2,1));
beta1=NewParams(2,1);
alpha1=NewParams(1,1);
end
end; % while
GN=N;
xlswrite('C:\Paper 5 new\excel\NewSet\EM\Expected-ChassisSubset164-5-
4.xls',N); % first external-then internal
APPENDIX Q 295
xlswrite('C:\Paper 5 new\excel\NewSet\EM\AllParams-ChassisSubset164-5-
4.xls',AllParams);
end
function [C,B1,B2,B3,B4]
=Simp_int_integral_GeneralLikelihoodEMX(k,beta0,alpha0,beta1,alpha1,h,L,U)
%{
%
Ccell:
C1 1
Bcell:
{1} B1 (yn)^beta
{2} B2 ((yn)^beta)*log(yn)
{3} B3 ((yn)^beta)*[log(yn)]^2
{4} B4 Sum(yi) i=1,...,n
%}
if k==1
intout(1)=FirstFunctionC(0,L(k),U(k),beta0,alpha0);
intout(2)=FirstFunctionB1(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
intout(3)=FirstFunctionB2(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
intout(4)=FirstFunctionB3(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
intout(5)=FirstFunctionB4(0,L(k),U(k),beta1,alpha1,beta0,alpha0);
else
for j=k-1:-1:1;
i=0;
startcon=sum(L(1:j));
stopcon=sum(U(1:j));
for x=startcon:h:stopcon;
i=i+1;
if j==k-1
IndArray(i)=x;
Ccell{1}(i)= FirstFunctionC(x,L(j+1),U(j+1),beta0,alpha0);
APPENDIX Q 296
Bcell{1}(i)=
FirstFunctionB1(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
Bcell{2}(i)=
FirstFunctionB2(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
Bcell{3}(i)=
FirstFunctionB3(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
Bcell{4}(i)=
FirstFunctionB4(x,L(j+1),U(j+1),beta1,alpha1,beta0,alpha0);
else
TmpPArrayC=Ccell{1}(INDICES);
TmpPArrayB1=Bcell{1}(INDICES);
TmpPArrayB2=Bcell{2}(INDICES);
TmpPArrayB3=Bcell{3}(INDICES);
TmpPArrayB4=Bcell{4}(INDICES);
for r=1:numel(TmpIndArray);
TmpPArrayB4(r)=(log(TmpIndArray(r))*TmpPArrayC(r)+TmpPArrayB4(r))*(beta0*exp(
alpha0))*(TmpIndArray(r)^(beta0-1))*exp(exp(alpha0)*(x^beta0-
TmpIndArray(r)^beta0));
TmpPArrayC(r)=TmpPArrayC(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
TmpPArrayB1(r)=TmpPArrayB1(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
TmpPArrayB2(r)=TmpPArrayB2(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
TmpPArrayB3(r)=TmpPArrayB3(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(exp(alpha0)*(x^beta0-TmpIndArray(r)^beta0));
end;
Ccell{1}(i)=h*(TmpPArrayC(1)+4*sum(TmpPArrayC(2:2:(numel(TmpPArrayC)-
1)))+2*sum(TmpPArrayC(3:2:(numel(TmpPArrayC)-
2)))+TmpPArrayC(numel(TmpPArrayC)))/3;
Bcell{1}(i)=h*(TmpPArrayB1(1)+4*sum(TmpPArrayB1(2:2:(numel(TmpPArrayB1)-
1)))+2*sum(TmpPArrayB1(3:2:(numel(TmpPArrayB1)-
2)))+TmpPArrayB1(numel(TmpPArrayB1)))/3;
Bcell{2}(i)=h*(TmpPArrayB2(1)+4*sum(TmpPArrayB2(2:2:(numel(TmpPArrayB2)-
1)))+2*sum(TmpPArrayB2(3:2:(numel(TmpPArrayB2)-
2)))+TmpPArrayB2(numel(TmpPArrayB2)))/3;
APPENDIX Q 297
Bcell{3}(i)=h*(TmpPArrayB3(1)+4*sum(TmpPArrayB3(2:2:(numel(TmpPArrayB3)-
1)))+2*sum(TmpPArrayB3(3:2:(numel(TmpPArrayB3)-
2)))+TmpPArrayB3(numel(TmpPArrayB3)))/3;
Bcell{4}(i)=h*(TmpPArrayB4(1)+4*sum(TmpPArrayB4(2:2:(numel(TmpPArrayB4)-
1)))+2*sum(TmpPArrayB4(3:2:(numel(TmpPArrayB4)-
2)))+TmpPArrayB4(numel(TmpPArrayB4)))/3;
IndArray(i)=x;
end; % if j==k-1
end; % for x=startcon:h:stopcon;
TmpPArrayC=Ccell{1}(INDICES);
TmpPArrayB1=Bcell{1}(INDICES);
TmpPArrayB2=Bcell{2}(INDICES);
TmpPArrayB3=Bcell{3}(INDICES);
TmpPArrayB4=Bcell{4}(INDICES);
TmpIndArray=IndArray(INDICES);
for r=1:numel(TmpIndArray);
TmpPArrayB4(r)=(log(TmpIndArray(r))*TmpPArrayC(r)+TmpPArrayB4(r))*(beta0*exp(
alpha0))*(TmpIndArray(r)^(beta0-1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayC(r)=TmpPArrayC(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayB1(r)=TmpPArrayB1(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayB2(r)=TmpPArrayB2(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
TmpPArrayB3(r)=TmpPArrayB3(r)*(beta0*exp(alpha0))*(TmpIndArray(r)^(beta0-
1))*exp(-1*exp(alpha0)*TmpIndArray(r)^beta0);
end
intout(1)=h*(TmpPArrayC(1)+4*sum(TmpPArrayC(2:2:(numel(TmpPArrayC)-
1)))+2*sum(TmpPArrayC(3:2:(numel(TmpPArrayC)-
2)))+TmpPArrayC(numel(TmpPArrayC)))/3;
intout(2)=h*(TmpPArrayB1(1)+4*sum(TmpPArrayB1(2:2:(numel(TmpPArrayB1)-
1)))+2*sum(TmpPArrayB1(3:2:(numel(TmpPArrayB1)-
2)))+TmpPArrayB1(numel(TmpPArrayB1)))/3;
APPENDIX Q 298
intout(3)=h*(TmpPArrayB2(1)+4*sum(TmpPArrayB2(2:2:(numel(TmpPArrayB2)-
1)))+2*sum(TmpPArrayB2(3:2:(numel(TmpPArrayB2)-
2)))+TmpPArrayB2(numel(TmpPArrayB2)))/3;
intout(4)=h*(TmpPArrayB3(1)+4*sum(TmpPArrayB3(2:2:(numel(TmpPArrayB3)-
1)))+2*sum(TmpPArrayB3(3:2:(numel(TmpPArrayB3)-
2)))+TmpPArrayB3(numel(TmpPArrayB3)))/3;
intout(5)=h*(TmpPArrayB4(1)+4*sum(TmpPArrayB4(2:2:(numel(TmpPArrayB4)-
1)))+2*sum(TmpPArrayB4(3:2:(numel(TmpPArrayB4)-
2)))+TmpPArrayB4(numel(TmpPArrayB4)))/3;
end % if k==1
C=intout(1);
B1=intout(2)/C;
B2=intout(3)/C;
B3=intout(4)/C;
B4=intout(5)/C;
end
function first_func=FirstFunctionC(prev_y,lower,upper,beta0,alpha0)
if prev_y==0
F=@(x) (beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x) (beta0.*exp(alpha0)).*((prev_y+x).^(beta0-
1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
% for B1(0)
function
first_func=FirstFunctionB1(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) (x.^beta1).*(beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x) ((prev_y+x).^beta1).*(beta0.*exp(alpha0)).*((prev_y+x).^(beta0-
1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
APPENDIX Q 299
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function
first_func=FirstFunctionB2(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) (x.^beta1).*log(x).*(beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
F=@(x)
((prev_y+x).^beta1).*log(prev_y+x).*(beta0.*exp(alpha0)).*((prev_y+x).^(beta0
-1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function
first_func=FirstFunctionB3(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) (x.^beta1).*(log(x).^2).*(beta0.*exp(alpha0)).*(x.^(beta0-
1)).*exp(-1.*exp(alpha0).*x.^beta0);
else
F=@(x)
((prev_y+x).^beta1).*(log(prev_y+x).^2).*(beta0.*exp(alpha0)).*((prev_y+x).^(
beta0-1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function
first_func=FirstFunctionB4(prev_y,lower,upper,beta1,alpha1,beta0,alpha0)
if prev_y==0
F=@(x) log(x).*(beta0.*exp(alpha0)).*(x.^(beta0-1)).*exp(-
1.*exp(alpha0).*x.^beta0);
else
APPENDIX Q 300
F=@(x) log(prev_y+x).*(beta0.*exp(alpha0)).*((prev_y+x).^(beta0-
1)).*exp(exp(alpha0).*(prev_y.^beta0-(prev_y+x).^beta0));
end;
if upper ~= -1
first_func=quadgk(F,lower,upper);
else
first_func=quadgk(F,lower,inf); % Case of infinity upper limit
end;
end
function GN = InformationMatrix()
alpha=-0.140950133;
beta=0.915588406;
RecordsNo=numel(Data(:,3));
rowindx=1;
UnitNo=1;
aHistoryL(1,1)=Data(rowindx,2);
aHistoryU(1,1)=Data(rowindx,3);
k=2;
rowindx=rowindx+1;
while (rowindx <= RecordsNo) & (Data(rowindx,1) ~= 1)
aHistoryL(1,k)=Data(rowindx,2);
aHistoryU(1,k)=Data(rowindx,3);
k=k+1;
rowindx=rowindx+1;
end
L{UnitNo}=aHistoryL(1:k-1);
U{UnitNo}=aHistoryU(1:k-1);
UnitNo=UnitNo+1;
end
NoofSystems=UnitNo-1;
APPENDIX Q 301
Sumni=0;
for i=1:NoofSystems;
Sumni=Sumni+numel(L{i});
[C,B1,B2,B3,B4,B5,B6,B7,B8,B9,B10] =
Simp_int_integral_GeneralLikelihood(numel(L{i}),beta,alpha,0.001,L{i},U{i});
% [C,B1,B2,B3,B4,B5,B6,B7,B8,B9,B10] =
Simp_int_integral_GeneralLikelihoodEM(numel(L{i}),beta0,alpha0,beta1,alpha1,0
.001,L{i},U{i});
N(1,i)=C;
N(2,i)=B1;
N(3,i)=B2;
N(4,i)=B3;
N(5,i)=B4;
N(6,i)=B5;
N(7,i)=B6;
N(8,i)=B7;
N(9,i)=B8;
N(10,i)=B9;
N(11,i)=B10;
end
Variance(1,1)=exp(2*alpha)*sum(N(8,:))-exp(2*alpha)*sum(N(2,:).^2);
Variance(1,2)=-
exp(alpha)*sum(N(11,:))+exp(alpha)*sum(N(2,:).*N(5,:))+exp(2*alpha)*sum(N(10,
:))-exp(2*alpha)*sum(N(2,:).*N(3,:));
Variance(2,1)=Variance(1,2);
Variance(2,2)=sum(N(7,:))-2*exp(alpha)*sum(N(6,:))+exp(2*alpha)*sum(N(9,:))-
sum(N(5,:).^2)+2*exp(alpha)*sum(N(5,:).*N(3,:))-exp(2*alpha)*sum(N(3,:).^2);
Q_Second_Complete=Q_Second+Variance;
FisherReverse=inv(-1*Q_Second_Complete);
APPENDIX Q 302
Error(1,1)=sqrt(FisherReverse(1,1));
Error(2,1)=sqrt(FisherReverse(2,2));
GN=Error;
xlswrite('C:\Paper 5 new\excel\NewSet\OneStep\InfoMatrix\Error-OneNR-
ChassisSubset164-3.xls',GN);
xlswrite('C:\Paper 5 new\excel\NewSet\OneStep\InfoMatrix\QP-OneNR-
ChassisSubset164-3.xls',Q_Prime);
end
% This function estimates the parameters which maxime the likelihood with
% the assumption that the midpoints of failure intervals are exact failure
% times. The input file is an excel sheet where an exact failure time
% is given as an interval with the same lower and upper bounds. The last
% right censoring interval is specified as an interval with the lower bound
% equal to the right censoring length, and the upper bound equal to -1.
function GN = generalLike_Mid()
Params(1,2)=0.5;
Params(1,1)=-0.3;
RecordsNo=numel(Data(:,3));
rowindx=1;
UnitNo=1;
aHistoryL(1,1)=Data(rowindx,2);
aHistoryU(1,1)=Data(rowindx,3);
k=2;
rowindx=rowindx+1;
while (rowindx <= RecordsNo) & (Data(rowindx,1) ~= 1)
aHistoryL(1,k)=Data(rowindx,2);
aHistoryU(1,k)=Data(rowindx,3);
k=k+1;
rowindx=rowindx+1;
end
L{UnitNo}=aHistoryL(1:k-1);
U{UnitNo}=aHistoryU(1:k-1);
UnitNo=UnitNo+1;
end
APPENDIX Q 303
NoofSystems=UnitNo-1;%3;
l=1;
AllParams(l,:)=Params(1,:);
ContWhile=true;
Sumni=0;
for i=1:NoofSystems;
if U{numel(L{i})} ~= -1
Sumni=Sumni+numel(L{i});
else
Sumni=Sumni+numel(L{i})-1;
end;
[C,B1,B2,B3,B4,logC]
=Simp_int_integral_GeneralLikelihood_Midpoints(Params(1,2),Params(1,1),L{i},U
{i});
N(1,i)=C;
N(2,i)=B1;
N(3,i)=B2;
N(4,i)=B3;
N(5,i)=B4;
N(6,i)=logC;
end
NewParams=Params'-inv(Q_Second)*Q_Prime
l=l+1;
AllParams(l,:)=NewParams(:,1)';
if NewParams(2,1) < 0
NewParams(2,1)=0.1;
APPENDIX Q 304
end
Params(1,:)=NewParams(:,1)';
end; % while
GN=N;
xlswrite('C:\Paper 5 new\excel\ExpectedChassisSubset164-Mid-New.xls',N);
xlswrite('C:\Paper 5 new\excel\AllParamsChassisSubset164-Mid-
New.xls',AllParams);
end
% This function calculates iteratively the parameter alpha given that beta=1
function GN = generalLike_Beta1_2()
Params(1,2)=1;
Params(1,1)=-0.3;
RecordsNo=numel(Data(:,3));
rowindx=1;
UnitNo=1;
aHistoryL(1,1)=Data(rowindx,2);
aHistoryU(1,1)=Data(rowindx,3);
k=2;
rowindx=rowindx+1;
while (rowindx <= RecordsNo) & (Data(rowindx,1) ~= 1)
aHistoryL(1,k)=Data(rowindx,2);
aHistoryU(1,k)=Data(rowindx,3);
k=k+1;
rowindx=rowindx+1;
end
L{UnitNo}=aHistoryL(1:k-1);
U{UnitNo}=aHistoryU(1:k-1);
APPENDIX Q 305
UnitNo=UnitNo+1;
end
NoofSystems=UnitNo-1;
l=1;
AllParams(l,1)=Params(1,1);
ContWhile=true;
Sumni=0;
for i=1:NoofSystems;
Sumni=Sumni+numel(L{i});
[C,B1,logC] =
Simp_int_integral_GeneralLikelihoodBeta1_exact(Params(1,1),L{i},U{i});
N(1,i)=C;
N(2,i)=B1;
N(3,i)=logC;
end
Q_Prime=Sumni-exp(Params(1,1))*sum(N(2,:));
l=l+1;
AllParams(l,2)=NewParams(1,1);
AllParams(l,3)=Q_Prime;
ContWhile=false;
end
Params(1,:)=NewParams(:,1)';
end; % while
GN=N;
xlswrite('C:\Paper 5 new\excel\Expected-beta1exactAudibleSubset125-10-
7.xls',N);
xlswrite('C:\Paper 5 new\excel\AllParams-beta1exactAudibleSubset125-10-
7.xls',AllParams);
end
function [C,B1,logC]
=Simp_int_integral_GeneralLikelihoodBeta1_exact(alpha0,L,U)
lambda=exp(alpha0);
C=1;
B1=0;
logC=0;
for i=1:numel(L);
if U(i) ~= -1
Denom=exp(-1*exp(alpha0)*L(i))-exp(-1*exp(alpha0)*U(i));
else
Denom=exp(-1*exp(alpha0)*L(i));
end;
C=C*Denom;
logC=logC+log(Denom);
if U(i) ~= -1
B1=B1+(1/lambda)+(L(i)*exp(-lambda*L(i))-U(i)*exp(-
lambda*U(i)))/(exp(-lambda*L(i))-exp(-lambda*U(i)));
else
B1=B1+(1/lambda)+L(i);
end;
end;
end
INDEX
A D
absolute comparison, 28 delayed replacement, 5, 177, 182
age dependent probabilities, 5, 128, 154, 160, 168, 169, detectability, xvi, 25, 26, 35, 36, 52, 58, 179, 207
176, 177, 182 downtime, xv, xix, 5, 55, 105, 109, 117, 118, 121, 126, 127,
AHP. See Analytical Hierarchy Process 128, 138, 139, 140, 141, 144, 145, 146, 150, 152, 154,
Analytical Hierarchy Process, 27, 179 175, 177, 182, 246, 248, 249, 260
Availability, xvi, 32, 33, 34, 35, 38, 188, 206, 209
F
C
failure consequence, 42, 57, 121, 122
classification, iii, 4, 33, 34, 37, 39, 53, 54, 68, 98, 190 Failure frequency, xvi, 35, 38, 206, 209
CMMS. See Computerized Maintenance Management failure mode, 25, 35, 36, 40, 41, 142, 179, 189
System Failure Mode and Effect Analysis, xxiv, 25
Computerized Maintenance Management System, 3, 188 failure rate, 55, 60, 62, 102, 104, 106, 108
corrective maintenance, 13, 14, 16, 21, 55, 67, 68, 104 failures data, 73, 74
Criteria, iii, xi, xii, xiv, xxi, 3, 13, 14, 20, 27, 28, 29, 30, 31, FDA. See Food and Drug Administration
32, 33, 35, 36, 37, 38, 39, 40, 41, 42, 43, 46, 52, 53, 54, Finite Time Horizon, xii, 128, 245, 252
55, 56, 57, 58, 109, 179, 180, 183, 185, 194, 204, 205, FMEA. See Failure Mode and Effect Analysis
209 Food and Drug Administration, 37
Criticality score, 29, 30, 31, 42, 43, 56, 58
G
grade, 30, 31, 39, 40, 41, 43, 204
307
INDEX 308
O
J
Opportunistic inspection, 104, 177, 181
JACAHO. See Joint Commission on Accreditation of
opportunistic maintenance, xiii, xxi, 64, 104, 105, 141, 175,
Healthcare Organization
177, 182, 202, 231, 247, 268, 271, 272, 278, 280
Joint Commission on Accreditation of Healthcare
Organizations, xxiii, 2, 26, 191
P
L pairwise comparison, 28, 29, 30, 53, 58, 59, 183
penalty cost, xix, 110, 114, 117, 118, 119, 125, 126, 128,
Laplace trend test, 72, 73, 75, 76, 77, 78, 90, 98
129, 134, 140, 144, 151, 153, 166, 176, 182
likelihood ratio, 4, 73, 81, 82, 83, 92, 93, 96, 98, 181, 251
periodic inspection, 4, 5, 23, 100, 104, 106, 107, 128, 133,
144, 153, 175, 177, 181
M Periodic preventive maintenance, 100
MADM. See Multi-Attribute Decision Making placeholder function, 131, 138, 148, 155, 170
Maintenance requirements, xvi, 33, 37, 58, 179, 208 power law, xv, 4, 61, 62, 81, 82, 98, 115, 116, 134, 150,
maximum likelihood, 61, 62, 82, 86, 89, 90 151, 166, 173, 181
INDEX 309
179, 180 128, 129, 141, 142, 143, 144, 146, 151, 168
probability of survival, 132, 252 Soft failure, xv, xvii, 5, 68, 71, 78, 79, 108, 109, 110, 122,
142, 144, 145, 146, 147, 148, 150, 152, 155, 157, 158,
R 169, 173, 177, 181, 184, 213, 232, 235, 247, 248, 249,
261
ranking, 28, 30, 39, 53, 196 SPI. See Safety and Performance Inspection
RCM. See Reliability Centered Maintenance
recall, 27
T
relative measurement, 28, 31, 53, 58, 179, 197
Reliability Centered Maintenance, xxiv, 25 threshold, xix, 5, 103, 105, 109, 117, 118, 122, 123, 125,
replacement, xx, 5, 15, 16, 18, 21, 100, 102, 103, 105, 106, 126, 127, 175
108, 129, 131, 132, 134, 136, 154, 155, 157, 159, 160, time-based maintenance, 54, 55, 56
161, 167, 168, 169, 170, 173, 174, 176, 177, 182, 186, Trend analysis, iii, 3, 4, 23, 61, 62, 73, 81, 83, 90, 98, 180,
187, 188, 189, 190, 191, 192, 194, 198, 222, 247, 269, 183
W
Weibull, xv, xvi, xvii, 61, 62, 75, 76, 188, 216, 217, 218
work order, xviii, 19, 20, 21, 22, 64, 67, 68, 69, 213