Você está na página 1de 7

Classical Control Theory Analysis of a Lean Manufacturing System

N.H. Ben Fong, John P. Shewchuk, and Robert H. Sturges Department of Industrial and Systems Engineering Virginia Polytechnic Institute and State University Blacksburg, VA 24061, USA Abstract
Lean manufacturing is a philosophy that seeks to minimize waste or non-value added to finished products in manufacturing processes. In this paper, a method to design and analyze a lean manufacturing system is examined using the transfer functions of classical control theory. This method is used to design and adjust system parameters by predicting and analyzing the dynamic characteristics of a kanban/pull system model. Transient-response system parameters such as undamped natural frequency and damping ratio are introduced to determine system time constant, settling time, and steady-state value. The effects of significant system variables on particular measuring responses are discussed. Keywords: Kanban system, transfer function, block diagram, transient responses.

1. Introduction
Taiichi Ohno introduced the Just-in-Time (JIT) philosophy at the Toyota Motor Company in Japan in the 1960s [1,2,3]. JIT is based on the concept of producing the kind of products needed, at the time needed and in the quantities needed for each subsequent stage of production. The most significant benefit of the JIT system is to reduce inventory, work-in-process (WIP), and production lead times. Lean manufacturing has evolved from the JIT system. The principles of lean are: define customer value, set the value stream map, make value flow by eliminating waster, let the customer pull product, and pursue for perfection by continuous improvement [2]. According to the MEP of NIST [3], lean manufacturing is defined as, a systematic approach to identifying and eliminating waste (non-value-added activities) through continuous improvement by flowing the product at the pull of the customer in pursuit of perfection. The foundation of the lean manufacturing concept is implemented using the following tools: seven wastes, 5S system, pull system/Kanban, Kaizen, cellular/flow manufacturing, value stream mapping, TPM, SMED, poka-yoke, quality tools, one-piece-flow, and working balancing. The essence of lean manufacturing is to provide quick time response from getting the production order through shipping the finished products. Quick time response gives greater productivity, shorter delivery times, lower costs, and increased customer satisfaction. The use of computer simulation, such as discrete event simulation (DES) and/or system dynamics (SD) approaches, to implement lean principles to model and analyze manufacturing systems has been considered successfully [3,4,5,6]. However, due to fast-changing make-to-order market demand, the DES method may require too much time to model and evaluate a particular manufacturing system. In addition, DES does not predict apriori the dynamic characteristics of a system model under non-steady state conditions, such as oscillation, settling time, and steadystate error. Although an SD approach can simulate the dynamic performance of a system model through the use of causal loop diagrams and stock-and-flow diagrams, it does not find key system parameters such as the loop gain or the location of the closed-loop poles of the plant structure for design and analysis purposes. In this paper, a method to design and analyze a lean manufacturing system is examined using transfer function (TF) of the classical control theory. The dynamic model to be used is extracted and modified from OCallaghans paper [5], a basic structure of a SD kanban-based manufacturing system. By converting the SD kanban-based model into block diagram representation and transfer functions, we show how to predict and analyze the dynamic characteristics of the lean system. A linear second-order differential equation is derived from this model for determining dynamic characteristic measures, such as undamped natural frequency, damping ratio, time constant, settling time, and steady-state inventory value.

2. A SD Kanban System Model


The Japanese word kanban refers to a card. The intent of kanban is to signal a preceding process that the next process requires parts/material. The kanban system can be considered as an information system that controls the lean production. In this paper, we have used the basic structure of a SD kanban system model that extracted from OCallaghans paper [5]. This kanban system model has been slightly modified and constructed by Vensim software as shown in Fig.1. Vensim [7] is a visual modeling tool that allows one to conceptualize, document, simulate, analyze, and optimize models of dynamic systems. Vensim provides a simple and flexible way of building simulation models from causal loop diagrams and/or stock and flow diagrams.

Lead Time LT Work-InProcess WIP Production Completion Rate PCR + - Production Orders PO + Kanban Cycle KC
Figure 1: A SD kanban system model

Shipment Time ST Finished Inventory FI Shipment Rate SR + <Container Size>

Production Start Rate PSR +

Desired Production Rate DPR -

Total Number of Kanban TNK

<Input>

Referring to Figure 1, the work-in-process (WIP) level is the accumulation of a difference between the production start rate (PSR) and the production completion rate (PCR) during a certain production lead- time (LT). The finished inventory (FI) determines the stock level between the production completion rate less the shipment rate (SR) over an average shipment time (ST). The total number of kanban defines the inventory allowed of the entire system. Any kanban has to be either attached to the stock container (WIP or FI) or the dispatching post (i.e. kanban receiving box). Each time a unit is withdrawn from the finished inventory, its kanban is detached and put in the collection box. Based on a certain time interval, the detached kanbans found in the collection box will be taken to the dispatching post, where they become production orders (PO). OCallaghan called this time interval the kanban cycle (KC), and it determines how fast the system reacts to changes in production rate. He further asserted that the kanban cycle is a time constant. However, in a later section of this paper, we will demonstrate that kanban cycle of this particular system model only represents portion of the entire system time constant via the transfer function. Finally, the backlog of PO determines the desired production rate (DPR). The total number of kanban (TNK) is defined as the number of kanban multiplied by each container size. In this kanban system, TNK has to be equal to the sum of the number of WIP kanban, the number of FI kanban and the number of PO kanban at all times. For example, we arbitrarily set the following system parameter values: number of kanban=10, container size=10 units, LT=0.5 day, KC=0.5 day, ST=5 days (assume that production works at a rate of 20 hours per day). By giving a step input of planned inventory to the system (i.e. 10*10=100 units), the FI system output response behaves similar to a goal seeking feedback loop structure as shown in Fig.2a [5,6]. The goal is to reach the planned inventory of 100 units. For the given set of parameters, the steady-state finished inventory reaches 83.33 units instead. The reasons for this effect are not immediately clear from Fig. 1 alone, but will be seen from the corresponding transfer functions. The WIP level has reached 37 units at its early stage and reduces down to 8.33 units after 3 days. Fig. 2b shows the production rates response subjected to the given step input. The PSR starts producing at a rate of 200 units/day and drops down to 16.67 units/day, whereas the PCR takes 0.5 day to reach its peak at 74 units/day and reduces down to 16.67 units/day after 4 days.

100 90 80

200 180 160 140

(Stock of Units)

(Units/Day)

70 60 50 40 30 20 10 0

Production Start Rate PSR Production Completion Rate PCR

Work-In-Process WIP Finished Inventory FI

120 100 80 60 40 20 0

Time (Day)

10

Time (Day)

10

Figures 2a and 2b: A step input response of a kanban system model

3. Classical Control Theory Approach


In control engineering, the mathematical model of a system is typically represented as a set of differential equations used to represent a physical system. As stated in Ogata [8,9], the transfer function (TF) of a linear, time-invariant, differential equation system is defined as the ratio of the Laplace transform of the output (response function) to the Laplace transform of the input (driving function) under all zero initial conditions. Like the SD representation of Sterman [14], a block diagram (BD) of a system is a graphical representation of the cause-and-effect relations operating in a particular system [8,9,10]. In a BD, all system variables are linked to each other through functional blocks. The functional block is a symbol for the mathematical operation on the input signal to the block that produces the output. The transfer functions of the components are usually entered in the corresponding blocks that are connected by arrows to indicate the direction of the flow of signals. One of the major advantages of using BD representation is to form the overall BD for the entire system by connecting the blocks of the components according to the signal flow in evaluating the overall performance and the interaction of each system components. In addition, the algebraic representation of the systems equations in terms of transfer functions allows easy manipulation for design and analysis purposes. We will now show in Figures 3 and 4 the mathematical equivalence of translating these causal-loop diagrams and stock-and-flow diagrams terminologies in modeling the kanban system. As given in Fig. 1, TF and BD representation from classical control theory follows:
r
+

TNK
_

1 KC

+ _

1/s

1 LT

+ _

1/s

FI

1 LT

1 ST

Figure 3: Complete block diagram representation of the SD kanban system in Fig. 1


r
1 1 TNK KC LT 1 1 1 1 1 1 1 1 1 s2 + + + s+ + + LT KC ST LT ST KC ST KC LT

FI

Figure 4: Block diagram reduction into a single transfer function block


FI(s) = r(s) 1 1 TNK KC LT 1 1 1 1 1 1 1 1 1 s2 + + + s + LT ST + KC ST + KC LT LT KC ST
2 K wn Y(s) = 2 2 X(s) s + (2zw n )s + w n

(1)

( )

(2)

As illustrated in Fig.3, there are a total four feedback loops found. The first two feedback loops contain the loop gain value of unity, whereas the third and the fourth loop have a loop gain of 1/LT and 1/ST, respectively. The 1/s term is simply an integrator. The complete block diagram representation as shown in Fig. 3 is mathematically equivalent to the kanban system model that is built by Vensim as shown in Fig. 1. For this multiple feedback loop block diagram as shown in Fig. 3, we can simplify it into a single transfer function by a step-by-step rearrangement called the block diagram reduction technique. Details of the technique are found in references [8,10,11]. After the block diagram reduction is applied, Fig. 3 representation reduces down into a single TF block as shown in Fig. 4. The resulting single TF block can be expressed under the Laplace transform domain and it yields a secondorder closed-loop transfer function as stated in equation (1). The denominator of equation (1) is the characteristic equation of the closed-loop kanban system model. Equation (2) gives a specific form of a second-order system differential equation with nonunity gain, where K is the proportional gain factor, w n is the undamped natural frequency, and z is the damping ratio of the system. Equation (2) can become a standard form of a second-order system equation by taking off the K term. The dynamic behavior of a second-order system can be described in terms of two parameters wn and z. If 0<z<1, the closed-loop poles are complex conjugates and its transient response is oscillatory. This system is called underdamped. If z=0, the transient response does not die out, it will oscillate forever. If z=1, the system is called critically damped; exponential behavior occurs if z>1, the system is called overdamped. Overshoot or oscillation will not occur unless z<= 0.707. To better understand the concept of the classical control theory, please read references [9,10,11,12]. For any linear second-order system, the time constant is computed as : t = (1/zwn). By comparing equation (1) to equation (2), it yields, Undamped natural frequency: Damping ratio:
z =

wn =
1

1 1 1 1 1 1 + + LT ST KC ST KC LT

(3) (4)

1 2

1 1 1 + + 1 1 1 1 1 1 LT KC ST + + LT ST KC ST KC LT

Damped natural frequency:

wd = wn 1 - z 2
2

(5) (6) (7)

Roots of characteristic equation: s1,2 = - zw n jw n 1 - z Time constant:


t=

1 2 = 1 1 zw n 1 LT + KC + ST

Kanban systems with minimal inventory storage cannot respond instantaneously and will exhibit transient responses when they are subjected to inputs or disturbances. The performance characteristics of a kanban system can be specified in terms of the transient response to a unit-step input. The transient response of a second-order dynamic system often shows damped oscillation before it gets to the steady state condition. It is common to determine the following indicators: delay time, peak time, maximum overshoot, rise time, and settling time. According to Fig. 2, we may not be able to determine three of these indicators due to its lack of oscillation. Settling time, ts is the time required for the response curve to reach and stay within a range about the final value. Rise time, tr is the time needed for the response to rise from 10% to 90% of its goal value. Settling time (2% criterion): Rise time (standard):
t s = 4t = 4 zw n

(8)

wd p - b 1 (9) tan -1 - zw = w wd n d -1 w d where (10) b = tan zw n Due to the non-standard form of a second-order system derived in equations (1), we have to use the following approximation instead of applying equation (9) to calculate the rise time [10]. tr =

Rise time (approx.):

t r, 10, 90% @

0.8 + 2.5z , wn

0 z 1

(11)

The inverse Laplace transform of equation (2) with a unit step input is obtained as:
1-z 2 e -zw n t y(t) = K 1 sin w d t + tan -1 z 1-z 2
1 1 KC LT wn

for t 0

(12)

As time goes to infinity, the steady-state value of the kanban system is computed from equations (1,2,12):

(TNK )
lim y(t) = K =
t

(13)

4. Analysis of System Dynamics Responses


Given the set of system parameters described in section 2, the kanban system response subjected to a unit step input gives the following characteristics: wn= 2.19 cycle/day, z=0.9585, wd=0.62 cycle/day, t=0.476 day, tr=1.459 day, ts=1.905 day, and K=83.33 units. Given production runs 20 hours per day, it takes 0.456 day to complete a cycle, hence the kanban cycle frequency is 9.13 hours/cycle. wn is meaningful only when a system oscillates and therefore it is of little interest alone, whereas wd is a characteristic of the system responsiveness, and hence it has great interest to production planning. The damping ratio, z provides a way to determine whether inventory has been made over or under the target goal. The time constant, t tells how fast the kanban system response to any input or disturbance and it leads to calculate the settling time taken to reach the planned inventory target within 2%. K gives the final inventory level at the steady state condition. As shown in Figure 2, there is a steady-state offset of 100-83.33=16.67 units from the planned inventory. By setting up different system parameters sets, we will find whether the finished inventory response could reach its final value at 100 units. We have decided to apply the Design of Experiment (DOE) [15] technique to identify the key system variable(s) that influence(s) the finished inventory output measures. An advanced factorial design called Box-Behnken Design is used to construct a balanced incomplete block of threelevels-three-factors design. The three design variables of the kanban system include lead time (LT), kanban cycle (KC), and shipment time (ST). The design range of variables are selected to be: Lead Time: 0.25 day (low); 0.5 day (medium); 1 day (high) Kanban Cycle: 0.25 day (low); 0.5 day (medium); 1 day (high) Shipment Time: 2.5 days (low); 10 days (medium); 17.5 days (high) In this factorial experiment, there are total of 13 independent runs of parameters set. The measured responses of the experiment are: wn, z , t , ts, and K. We used Excel to generate all measurement responses into a matrix from the 13 different sets. By applying Matlab, we are able to compute and determine the most significant factors effect among the five measured response variables. The main factors effect and their interaction factors effect are all calculated via the regression coefficients of the matrix as shown in Table 1. Table 1: Box-Behnken Design Factors Effect Responses Summary
Undamped Natural Frequency(Wn) -0.7318 -0.7318 -0.1629 0.1308 0.1308 0.1155 0.2379 0.0053 0.0053 Damping Ratio (S) -0.0103 -0.0103 0.0358 0.0552 0.0552 -0.0254 -0.1144 0.0077 0.0077 Time Constant (T) 0.1632 0.1631 0.0209 0.0041 0.0041 -0.0206 0.1037 0.0121 0.0121 Settling Time (ts) 0.6526 0.6526 0.0837 0.0162 0.0162 -0.0823 0.415 0.0483 0.0483 Steady State Value(K) -3.7641 -3.7641 12.1431 -0.9109 -0.9109 -8.1435 0.1984 2.6595 2.6595

LT KC ST LT^2 KC^2 ST^2 LTxKC LTxST KCxST

The results indicate that both LT and KC are the most significant factors to influence the undamped natural frequency, wn. As LT and KC are reduced, the frequency of the kanban system increases to provide more cycles completed per day. The damping ratio, z is most influenced by the interaction factors of LTxKC. As LT increases

and KC decreases, or vice versa, it will affect the damping ratio significantly. The time constant, 1/zwn is significantly affected by LT and KC and their factors interaction. As LT and KC increases, the time constant increases. Similarly, the settling time, ts is also significantly affected by LT and KC and their factors interaction. Finally, the steady-state value, K is most influenced by ST. As the shipment time takes longer, the finished inventory is accumulating more units until it reaches the planned inventory (i.e. 100 units). Final comments on this particular SD kanban system used, as LT and KC are both reduced to very short periods (i.e. 30 minutes) and as the shipment time is made extremely long (i.e. 100 days), the steady-state value will very closely reach its final value of 99.95 units. However, it is not economically realistic to bump the production complete rate up to 1700 units per hour with the WIP level of 45 units at one point and wait for 100 days before shipment is made. For the given kanban system values chosen, the system does not reach the planned inventory under normal range of system parameters. Although this kanban model behaves like a critically damped system, as z goes to unity, the system should give the fastest time response under non-oscillatory conditions. Better response can be obtained by setting LT=0.25 day. With KC unchanged, the system response changes dramatically.

5. Conclusions and Recommendations


In this paper, we have demonstrated an alternative method to model and analyze a dynamic lean manufacturing system using classical control theory. By translating a Vensim-based kanban model into block diagram and transfer function representations, we are now able to determine some key system parameters and corresponding transient response performance measures which the system dynamic approach does not offer. Firstly, the undamped natural frequency wn and the damping ratio z enables us to compute the kanban cycle frequency and its responsiveness subjected to the planned inventory. The authors have found the system time constant to be function of LT, KC, and ST, not only KC. Secondly, the damping ratio allows us to determine whether excessive inventory is being made (i.e. any overshoot) during the transient period. Although it is not shown in this particular kanban system, a typical second-order system should allow one to find the maximum inventory overshoot, plus its corresponding time to reach that peak value. Hence, we can design the damping ratio such that the inventory overshoot does not go over a particular inventory cost allowed. Thirdly, algebraically, we can now compute the rise time, the settling time, and the steady-state value as the kanban system reaches its steady-state condition where system dynamic approach could only find by running numerous computations to seek and obtain the same values. Finally, the Box-Behnken Design factorial experiment has given us a good statistical estimate about which variable factors affect most on the five chosen response variables. It shows that LT and KC have significant influence on wn , t , ts and their interaction factors have a significant effect on z ; whereas ST has its impact on reaching the steady-state value. The authors are currently working on developing different controllers to improve the transient response performance of any given lean manufacturing system. In addition, stability analysis based on the roots of characteristic equation, s1,2 shall be considered for higher order systems to provide quick lean response under stable conditions. Finally, we recognize that pure delays exist in real manufacturing systems and cannot be represented accurately with a second-order model. Higher order delays do mimic pure delays and will be the focus of future analysis and modeling.

References
[1] R.G. Askin and J.B. Goldberg, Design and Analysis of Lean Production Systems, John Wiley & Sons, 2002. [2] F.W. Breyfogle III, Implementing Six Sigma: Smarter Solutions Using Statistical Methods, John Wiley & Sons, Inc., 2003. [3] M.Adams, P.Componation, H.Czarnecki, and B.J. Schroer, Simulation as a tool for continous process improvement, Proceedings of the 1999 Winter Simulation Conference,p.766-773, 1999. [4] R.B. Detty and J.C. YingLing, Quantifying benefits of conversion to lean manufacturing with discrete event simulation: a case study, International Journal of Production Research, vol.38, no.2, pp. 429-445, 2000. [5] R. OCallaghan, A System Dynamics Perspective on JIT-Kanban, The 1986 International Conference of the System Dynamics Society, pp.960-1004, Oct 1986. [6] Y.P. Gupta and M.C. Gupta, A system dynamic model for a multi-stage multi-line dual-card JIT kanban system, International Journal of Production Research, vol.27, no.2, pp. 309-352, 1989. [7] Vensims user guide and reference manual, Ventana Systems, Inc., October 7, 2003. [8] K. Ogata, System Dynamics, 2nd edition, Prentice-Hall, Inc., 1992. [9] K. Ogata, Modern Control Engineering, 4th edition, Prentice-Hall, Inc., 2002. [10] W.J. Palm III, Modeling, Analysis, and Control of Dynamic Systems, 2nd edition, John Wiley & Sons, Inc., 1998 [11] C.L. Phillips and R.D. Harbor, Feedback Control Systems, 2nd edition, Prentice-Hall, Inc., 1991. [12] J.Van De Vegte, Feedback Control Systems, Prentice-Hall, Inc., 1986. [13] D.C. Montgomery, Design and Analysis of Experiments, 3rd edition, John Wiley & Sons, Inc., 1991. [14] J.D. Sterman, Business Dynamics: Systems Thinking and Modeling for a Complex World, McGraw-Hill Companies, 2000.

Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

Você também pode gostar