Você está na página 1de 40

INTRODUCTION

1. Quality can be defined in many different ways and may mean different
things for different people

2. A key factor for business success manufacturing, service, or others

3. Dimensions of Quality
Performance
Reliability
Durability
Serviceability
Aesthetics
Features
Perceived Quality
Conformance to Standards

4. Definition 1: Quality means fitness for use


Products or services meet the needs of customers and make customers
satisfied
Quality of Design Intentional design differences by:
Material used
Specifications on the components
Reliability

1
Appearances
Product performance

Quality of Conformance How well the product conforms to


specifications required by the design. This is determined by:
Manufacturing processes
Training and supervision of workforce
Quality assurance procedures and standards

5. Definition 2: Quality is inversely proportional to variability (for large


number of same or similar products and over long time period)
Less variability among products, better quality level
This is a more operational effective definition

The automotive transmission example:

2
6. Definition 3: Quality improvement is the reduction of variability in
processes and products
Less variability => better quality products
Quality improvement Elimination of waste

7. Quality Characteristics
Physical: structure, length, weight, etc
Sensory: taste, appearance, color
Time oriented: reliability, durability, serviceability
o Attributes data generally discrete (number of counts, etc.)
o Variables data generally continuous (length, voltages, etc.)

8. Quality Engineering using statistical methods to ensure that the quality


characteristics of a product are at the nominal or required levels
Nominal or target value desired value for the quality characteristic
Specifications desired measurement range for the quality characteristic
Upper specification limit (USL) largest allowable value
Lower specification limit (LSL) smallest allowable value
When a component or product does not meet specifications, they are
considered to be nonconforming or defective
A nonconforming product is a defective product
Defects are nonconformities that may affect the safe or effective use of
the product.
A defective or nonconforming product must have one or more defects.
A product with defects (nonconformities) may or may not be defective
(nonconforming)

3
9. Concurrent Engineering and Product Quality
Team approach in product and process design

10. Statistical Methods for Quality Control and Improvement


Statistical process control (SPC)
Acceptance sampling
Design of experiments (DOX)

11. Statistical Process Control (SPC)


Control charts are used for process monitoring and variability reduction.
An on-line quality control tool.

12. Acceptance Sampling

4
Inspection and classification of a sample of the product selected at
random from a larger batch or lot and the ultimate decision about
disposition of the lot.
Two types:
Outgoing inspection - follows production
Incoming inspection - before use in production
Should be used as an audit tool to ensure quality level

13. Design of Experiments


an approach to systematically varying the controllable input factors in the
process and determine their effects on the output
They are off-line quality tools
Crucial for variability reduction

14. Quality Control Systems

5
specifications of what is needed
design to meet the specifications
production to realize the specifications
inspections for conformance to specifications
review to revise the specifications

15. Quality Costs


Prevention Costs quality planning and engineering, product and process
design, etc.
Appraisal Costs inspection and testing, equipment and materials
Internal Failure Costs scrap, rework, downtime, etc.
External Failure Costs warranty charges, liability costs, etc.

6
16. Other Aspects of Quality Control and Improvement
Total Quality Management (TQM)
Quality Standards and Registration
Six Sigma
origins at Motorola in 1987, expanded impact during 1990s to present
focus on process improvement with an emphasis on achieving
significant business impact
A process is an organized sequence of activities that produces an
output that adds value to the organization
All work is performed in (interconnected) processes
Any process can be improved
An organized approach to improvement is necessary
The process focus is essential to Six Sigma

7
Six Sigma focus
Initially in manufacturing
Commercial applications
Banking
Finance
Public sector
Services

DFSS Design for Six Sigma


New process design
New product design (engineering)
A disciplined and analytical approach to process and product improvement

8
Specialized roles for people; Champions, Master Black Belts, Black Belts,
Green Belts
Top-down driven (Champions from each business)
BBs and MBBs have responsibility (project definition, leadership,
training/mentoring, team facilitation)
Involves a five-step process (DMAIC)
DMAIC solves problems by using Six Sigma tools
Define problems in processes
Measure performance
Analyze causes of problems
Improve processes to remove variations and non-value added
activities
Control processes so problems do not recur

DMAIC is closely related to the Shewhart cycle (variously called the


Deming cycle, or the PDCA cycle)

Successful implementations characterized by:


Committed leadership
Use of top talent
Supporting infrastructure
Formal project selection process
Formal project review process
Dedicated resources
Financial system integration

9
Project-by-project improvement strategy (borrowed from Juran)

DFSS matches customer needs with capability


Mean and variability affects product performance and cost
Designers can predict costs and yields in the design phase
Consider mean and variability in the design phase
Establish top level mean, variability and failure rate targets for a
design
Rationally allocate mean, variability, and failure rate targets to
subsystem and component levels
Match requirements against process capability and identify gaps
Close gaps to optimize a producible design
Identify variability drivers and optimize designs or make designs
robust to variability
Process capability impact design decisions
DFSS enhances product design methods

10
Lean focuses on waste elimination
Definition
A set of methods and tools used to eliminate waste in a process
Lean helps identify anything not absolutely required to deliver a quality
product on time.
Lean methods help reduce inventory, lead time, and cost
Lean methods increase productivity, quality, on time delivery, capacity, and
sales

17. Quality Responsibilities


Marketing -- customer's needs, wants and price they can pay
Product engineering design the products to meet the customer's needs
(design, tolerance, material, safety, designed quality vs. controlled
quality)
Purchasing - quality material and components (inspection, contract,
standard, cost)
Manufacturing - process selection, development, planning, production
Inspection and test - measuring equipment, standard procedures,
inspector training, quality control leading to quality improvement, quality
is not inspected into the product or the process
Packaging and shipping - protect the quality product
Product service - help customers to fully realize the intended function of
the product

11
18. Legal Aspects of Quality
Product liability exposure
Concept of strict liability
1. Responsibility of both manufacturer and seller/distributor
2. Advertising must be supported by valid data

19. Implementing Quality Improvement


A strategic management process, focused along the eight dimension of
quality
Suppliers and supply chain management must be involved
Must focus on all three components: Quality Planning, Quality
Assurance, and Quality Control and Improvement

20. Quality and Reliability


to function effectively over a certain period of time
built in through quality of design

12
Probability and Statistical Methods for Quality Control

1. Frequency Distribution

Sample average, sample variance, histograms etc.

2. Probability Distributions

Continuous distributions - Variable being measured is expressed on a


continuous scale. Normal, Exponential, Gamma, Weibull, etc.
Discrete distributions - Variable being measured is expressed on a discrete
scale. Binomial, Hypergeometric, Poisson, etc.
Mean, variance or standard deviation

Normal Distribution
Symmetric, bell-shaped
Mean:
Standard deviation (SD):
1 x 2
1 ( )
Density function: f ( x ) e 2
x
2
1 x
1
a ( )2
Cumulative function: F ( a ) P ( x a ) e 2
dx
2
Standardized normal distribution:
x
Let z

then x z
For xa
it leads to x z a

13
and z a
a
leads to z

a
1
z 1 2u2
Then: F (a ) P( x a ) P( z ) e du
2
a a
or: P( x a ) P( z ) ( )

Z is normally distributed with mean 0 and standard deviation, 1.
Use the standard normal distribution to find probabilities when the original
population or sample of interest is normally distributed
The normal distribution has about
68.26% of the observations lying within one SD of the mean,
95.44% within two SDs and
99.73% within three SDs.
This allows for a simple description of where most values are to be found

0.4

0.3

68% of obs.
0.2

95% of obs.
0.1
99.7% of obs.

0.0
-4 -3 -2 -1 0 1 2 3 4
Z score

14
Example

15
Normal Distribution Table

16
Example

(note: P should be 0.9192 according to the Normal Distribution Table.)

17
Normal distribution properties:

Mean and Variance of a linear combination of normally distributed variables


If x1 , x2 , , x n are normally and independently distributed random variables

with means 1 , 2 ,, n and variances 1 , 2 ,, n , respectively, then the


2 2 2

variable of ( a1 , a2 ,...an are constants):

y a1 x1 a2 x2 ... an xn
is normally distributed with mean:
y a11 a2 2 ... an n

and variance:
y2 a12 12 a22 22 ... an2 n2

The Central Limit Theorem


If x1 , x2 , , x n are independent random variables with means 1 , 2 ,, n and

variances 1 , 2 ,, n , respectively, for the random variable


2 2 2

y x1 x2 ... xn
then the distribution of:
n
y i
z i 1
n


i 1
i
2

approaches the N(0,1) distribution when n approaches infinity.


Approximation improves as n becomes larger
If xi s are from the same distribution, not very far from normal, it will
work well for n greater than 3 or 4.

18
If xi s are from the same distribution, then
n
y i
y n y n
z i 1

n
n 2 n

i 1
i
2

Exponential distribution

f ( x ) e x x 0
1
Mean:

1
Variance:
2

2
Cumulative probability:
a
F ( a ) P( x a ) e t dt 1 e a , a 0
0

This function may be used for time to failure probability in reliability models
where is the failure rate and 1 / is the mean-time-to-failure (MTTF)

19
Bernoulli distribution
p x 1
p( x )
1 p q x0

Mean: E ( x ) p

Variance: Var( x ) 2 p(1 p )

Binomial distribution

If x is the number of successes in n Bernoulli trials, its probability is:

n! n
p( X x ) p x (1 p ) n x = p x (1 p )n x
x!( n x )! x

Mean: E ( x ) np

Variance: Var( x ) 2 np(1 p )

20
Example

If we use Minitab, this can be done by:

21
When x is binomial with E ( x ) np and Var( x ) np(1 p ) , then the variable
x np(1 p ) p(1 p )
p is also binomial with E ( p ) p and Var( p )
n n2 n

x
( na )
n x
P( p a ) P a P( x na ) p (1 p )n x
n x 0 x
Example. A random sample of 100 units is drawn from a production process every
half hour. The fraction of nonconforming products is 0.03. Calculate the
probability of p 0.04 .
Solution. Binomial with p 0.03 and n 100
4
100
P( p 0.04) P x 4 (0.03) x (0.97)100 x 0.818
x 0 x

If we use Minitab, this can be done by:

22
Hypergeometric distribution
D N D

x n x
p( x ) , x = 0,1,2, min (n,D)
N

n

nD nD D N n
mean: , variance: 1
2

N N N N 1
Example. A lot of N 100 items with D 5 nonconforming. A random sample of
n 10 items were taken without replacement. Then the probability that there is one

or less non-conforming items in the sample is


5 95 5 95

p( x 0) p( x 1)
0 10 1 9
100 100

10 10
5! 95! 5! 95!
0! (5 0)! 10! (95 10)! 1! (5 1)! 9! (95 9)!
0.923
100! 100!
10! (100 10)! 10! (100 10)!

Using Minitab:

23
Poisson distribution
e x
p( x ) , x 0,1,2,....
x!
mean: , variance: 2

Example
The number of wire-bonding defects per unit in a semiconductor device is Poisson
distributed with 4 . Then the probability that a unit has 2 or less defects will be:
2
e 4 4 x
P ( x 2) 0.2380
x 0 x!

24
3. Approximations

Binomial approximation to Hypergeometric

N=200, D=5, n=10, x=0


5 195

0 10
p (0 ) 0.7717
200

10

Since n/N=10/200=0.005 is small, use binomial with p=D/N=5/200=0.025 and


n=10 to calculate:
10
p(0) (0.025)0 (0.975)10 0.7763
0

Poisson approximation to binomial with np

25
Normal approximation to binomial with np and np (1 p )
2

a 12 np a 12 np
P{x a}
np(1 p ) np(1 p )

and
b 12 np
a 2 np
1
P{a x b}
np(1 p ) np(1 p )

For sample fraction nonconforming ( p ) :


p p
p(1 p )
p2
n
v p u p
P{u p v}
p(1 p ) / n p(1 p ) / n

Normal approximation to Poisson with and 2

a 12 a 12
P{x a}

and
b 12 a 12
P{a x b}

26
4. Sample Statistics and Sampling

Sample: A set of data taken from the process


Sample Mean:
n

x x ... xn x i
x 1 2 i 1
n n
Sample Variance:
n

(x i x )2
S2 i 1

n 1
Sample Standard Deviation:
n

(x i x )2
S i 1

n 1
Sampling from a normal distribution
o If x ~ N ( , 2 ) and x1 , x2 ,..., xn is a sample of size n from x, then for
x1 x2 ... xn
x
n
we have: x ~ N ( , 2 / n )
o If x1 , x2 ,..., xn is a sample of size n from x following any distribution, not
necessarily normal, then x ~ N ( , 2 / n ) will hold for large n.
o If x1 , x2 ,..., xn are normally and independently distributed with
x ~ N (0,1) , then the random variable

y x12 x22 ... xn2

follows chi-square ( n2 ) distribution with n degrees of freedom:


1 y2
f ( y) y ( 2 )1e 2 y0
n

n
2 2
n

27
o If x1 , x2 ,..., xn is a random sample from x ~ N ( , ) , then the random
2

n
( xi x ) 2
variable y follows n2 distribution with n-1 degree of
i 1 2

freedom. In addition, we have:


n
( xi x ) 2 n
( xi x ) 2 ( n 1) S 2 ( n 1)
y
i 1 2 i 1 ( n 1 ) 2
2
S 2 ( n 1)
so we have ~ n21 .
2

o If x is a standard normal random variable and if y is a chi-square random


variable with k degrees of freedom, then
x
t
y/k

is t-distributed with k degrees of freedom.


Sampling from a Bernoulli distribution
If x1 , x2 ,..., xn , is a random sample from a Bernoulli process with x p and

x2 p(1 p ) , then the sum of the sample:


n
x xi
i 1

is binomial with parameters p and n . The sample mean:


n

x x2 ... xn x i
x
x p 1 i 1

n n n

is a discrete random variable with mean x p and variance x2 p(1 p ) / n .


The probability P( x a ) is calculated by:

x na n

P( p a ) P( a ) P( x na ) p x (1 p )n x

n x 0 x

28
5. Point Estimation of Process Parameters

Use sample mean and sample variance as estimated mean and estimated
variance 2 for the real mean and real variance 2

x , 2 S 2
if x is normally distributed, then can be estimated by S
c4
If x1 , x 2 , , x n are from a normal distribution and let

R max( xi ) min( xi ) x max x min

Then can be estimated by


R
d2

c4 and d 2 can be found in given tables.

29
6. Statistical Inference

Null hypothesis H 0 : 0

Alternative hypothesis H1 : 0

Producers risk
P{type I error}=P{reject H 0 | H 0 is true}
Consumers risk
P{type II error}=P{fail to reject H 0 | H 0 is false}

Power of test
Power = 1 = P{reject H 0 | H 0 is false}

Central Limit Theorem Revisited


When x1 , x2 , , x n are independent random variables with means

1 = 2 = n = and variances 1 = 2 = n = , then for


2 2 2 2

y x1 x2 ... xn
we have
n n
y i
y
n i
y i i 1
i 1
n n n x
i 1

n n
n
i2
i 1
i2
i 1 n n
n
approach N(0,1) distribution when n approaches infinity.

Hypothesis testing on when is known (for any distribution)

Let Z 2 is a point on Z axis of the N(0,1) distribution such that

P( Z Z 2 ) 2 and P( Z Z 2 ) 2 . To test

30
H 0 : 0

H1 : 0

1. Take a sample of size n, x1 , x 2 , x 3 ,..., x n


x 0
2. Calculate: Z0
n
3. For two-sided testing, we expect that 100( 1 )% of Z 0 would be in the

range between Z 2 and Z 2 . So we reject H 0 if | Z 0 | Z 2 .

4. For one-sided testing, If the alternate hypothesis is H 1 : 0 (smaller

is better), we expect that 100( 1 )% of Z 0 would be in the area outside

Z . So we reject H 0 if Z 0 Z . If the alternate hypothesis is H 1 : 0

(bigger is better), we expect that 100( 1 )% of Z 0 would be in the area

outside Z . So we reject H 0 when Z 0 Z .

31
Normal Distribution Table

Confidence interval on when is known (for any distribution)


x
For a given producers risk , since P ( Z 2 Z 2 ) 1 ,
n
from

x Z 2 x Z 2
n n
we can state that should be in the range of:

x Z 2 x Z 2
n n

with 100( 1 )% probability. This is called 100( 1 )% confidence


interval for the unknown .

32
P -values for hypothesis testing

2[1 | Z 0 |] H 0 : 0 , H 1 : 0

P 1 ( Z 0 ) H 0 : 0 , H 1 : 0
( Z ) H 0 : 0 , H 1 : 0
0

For the problem in Example 3.1, we calculated that Z 0 2.66 . Since the test

was for H 0 : 0 , H 1 : 0 , the P -values is:


P 1 ( Z 0 ) 1 ( 2.66) 0.0039

It means that the null hypothesis will be rejected at any level of significance
of P 0.0039

33
Hypothesis testing on when is unknown (for normal distribution)
From the discussion on sampling from normal distributions, it can be
verified that the testing of

H 0 : 0
H1 : 0
must be done based on t-distribution with n-1 degree of freedom.
Calculate:
x 0
t0
S n

where S is the sample standard deviation.


For a given producers risk and for two-sided testing, if | t0 | t 2, n 1 ,
reject H 0 For one-sided testing, If H 1 : 0 , when t0 t , n 1 , reject H 0

If H 1 : 0 , when t0 t , n 1 , reject H 0 .

34
35
Confidence interval on when is unknown (for normal distribution)
It is similar to the known variance situation while should be replaced by
S and t-distribution should be used. For a given producers risk ,
should be in the range of:
S S
x t 2,n 1 x t 2, n 1
n n

with 100( 1 )% probability.


Hypothesis testing on for a normal distribution
H 0 : 2 02

H1 : 2 02
Use:
( n 1) S 2

2

02
0

where S 2 is the sample variance.


For a given producers risk and for two-sided testing, reject H 0 if

02 2 2 , n 1 or 02 12 2, n 1 . For one-sided testing:

If H 1 : 0 , when 02 2 , n 1 , reject H 0
2 2

If H 1 : 0 , when 02 12 , n 1 , reject H 0
2 2

To find out process variances is very important since they are indicators of
the process capabilities.
Confidence interval on for a normal distribution
For a given producers risk , 2 should be in the range of:
( n 1) S 2 ( n 1) S 2
2
2 2 , n 1 12 2 , n 1

with 100( 1 )% probability.

36
Inference on a population proportion
Hypothesis testing will be based on the normal approximation to binomial.
H 0 : p p0
H 1 : p p0

Use:
x 0.5 np0
np (1 p ) if x np0

Z0
0 0

x 0.5 np0 if x np0


np0 (1 p0 )

For two sided-testing, reject H 0 when | Z 0 | Z 2 . One-sided testing is similar

to other tests.

Example. Test: H 0 : p 0.04 ; H 1 p 0.04 . Take a sample of 50 units and


found that 1 unit is defective. Since x 1 50 0.04 2.0 , use:
x 0.5 np0 1 0.5 2.0 0.5
Z0 1.443 .
np0 (1 p0 ) 2.0(1 0.04) 2.0(1 0.04)

H 0 : p 0.04 will not be rejected if 0.05 ( Z 0.025 1.96 ).

Confidence interval on a population proportion


When n is large and p 0.1 , we can use normal approximation to binomial
for the testing. Let p x / n where x is the number of nonconforming items
and n is the sample size, calculate:
p (1 p ) p (1 p )
p Z 2 p p Z 2
n n
If n is small, use binomial distribution.
If n is large but p is small, use Poisson approximation to binomial.

37
Type II Error and OC Curves

For a process with known , we assume that the center of the process has
shifted from 0 to 1 0 . Perform a 2 sided hypothesis test with given
and a sample of n units from the process. We want to find out the
probability that such test does not reject the null hypothesis H 0 : 0 . This
is the type II error of the test. This is also called consumers risk

P{type II error}=P{fail to reject H 0 | H 0 is false}

For the described process, if the mean is 0 , then for the sample mean x , we
should have:

x 0
Z ~ N (0,1)

n

Since the true mean is 1 0 , then for the sample mean x1 , we have

38


x1 0 x1 0 x1 0 n
Z0 ~ N (0,1)


n n n n
or

n
Z0 ~ N ( ,1)

The consumers risk can be calculated by:

n n
( Z 2 ) ( Z 2 )

as shown in the above figure.

Example 3.7. A process has known 0.1 . Assume that the center of the process
has shifted by 0.1 . We want to perform a 2 sided hypothesis test with =0.05
with a sample of 9 units from the process. We want to find out the type II error of
this test. Since =0.05, from a normal distribution table, we can find out that
Z 2 Z 0.025 1.96 . We then calculate :

n n
( Z 2 ) ( Z 2 )=

0.1 9 0.1 9
(1.96 ) ( 1.96 ) = ( 1.04) ( 4.96) =0.1492
0.1 0.1

The consumers risk is 0.1492. This result tells us that about 14% time that such
test is not able to detect such shift of the process. The power of the test is
1 0.8508.

39
| |
OC curves: To give the values for different n, and d

( =0.05)

40

Você também pode gostar