CC BY-NC-ND
0Available online at www.sciencedirect.com
Procedía Engineering
ELSEVIER
Procedía Engineering 53 (2013) 607 - 615
www.elsevier.com/locate/procedia
Malaysian Technical Universities Conference on Engineering & Technology 2012, MUCET 2012 Part 2 - Mechanical And Manufacturing Engineering
CONWIP Based Control of a Semiconductor End of Line Assembly
Advancement of technology and trends in globalization has resulted in higher customer demands and expectations. Manufacturers now offer mass customization to stay competitive. In the semiconductor industry, where product mix and volume are high, production is further complicated by the different process routes and processing times for different product families. Coupled with rapid changeovers of products, it is essential to keep the work in process (WIP) low in order to reduce the inventory level on the shop floor. CONWIP is a production control strategy applicable in many manufacturing environment, that uses cards to control WIP level. In this paper, discrete event simulation models for processes at the End of Line (EOL) assembly in a semiconductor manufacturing company were developed. Experiments were conducted using these models to compare the current system with the single loop and multi loop CONWIP control mechanisms. Performance parameters of throughput, cycle time and WIP level were compared in all experiments. The result, firstly, shows that, generally, CONWIP production control is more effective to reduce WIP level compared to the current system. The reduction in WIP is accompanied by corresponding improvements of cycle times. S econdly, the multiloop system performs better than the single loop system with higher cycle time reduction. Multi loop control is also more robust and provides a better control mechanism compared to the single loop system.
© 2013 The Authors. Published by Elsevier Ltd.
Selection and peer-review- under responsibility of the Research Man agement & Innovation Centre,Universiti Malaysia Keywords: CONWIP, Production Control, Semiconductor, End of Line, Simulation.
1. Introduction
The semiconductor industry is characterized by high product mix and high production volume. The processes in this industry are typically complicated and time consuming [1,2]. At the End of Line (EOL) of semiconductor manufacturing, the assembly process is further complicated by the different, routes and processing times for the numerous product families. The accumulations of WIP and control policies are often a challenge to the semiconductor industry over the years. To be competitive, WIP in EOL assembly needs to be effectively controlled in order to minimize inventory and reduce cycle time.
* Corresponding author. E-mail address: kuaneng@utem.edu.my
Chong Kuan Enga* and, Lim Ke Sina
department of Manufacturing Management Faculty of Manufacturing Engineering Universiti Teknikal Malaysia Melaka Hang Tuah Jaya, 76100 Durian Tunggal, Melaka, Malaysia.
Abstract
1877-7058 © 2013 The Authors. Published by Elsevier Ltd.
Selection and peer-review under responsibility of the Research Management & Innovation Centre, Universiti Malaysia Perlis doi: 10.1016/j.proeng.2013.02.078
Production control systems are typically categorized as the push and pull systems. Example of push system includes MRP, MRP II, whereas examples of the push system include kanban and CONWIP. Some advantages of implementing pull over push system include the reduction of cycle time variability and the flexibility to make any engineering and design changes. Comparing between the two main pull systems, kanban maintain a tighter WIP control through individual cards at each workstation whereas CONWIP is generally easier to implement and regulate as only one set of production cards are used to manage WIP.
This study was conducted in a case company, which is a local semiconductor manufacturer. Data collection was focused on the EOL assembly that involved six main processes namely, moulding, curing, tie bar cut, manual inspection, trim and form, and test, as shown in the Fig 1. One major product family is considered for the purpose of this study. Currently, the production control employed is by using the conventional push approach and the company wishes to improve cycle time and reduce high WIP at the EOL assembly.
Fig. 1. EOL assembly process flow.
The remainder of this paper is organized as follows. In Section 2, the relevant literature review will be discussed. In section 3, the details for model development, verification and validation are presented. In section 4, the result and discussion of the experiments are provided. Finally in Section 5, the conclusion and future research opportunities will be discussed.
2. Literature review
Industries have been consistently putting in effort to reduce WIP as it incurs the inventory cost. These efforts can be seen from the numerous past researches on WIP issues. For instances, Bertrand and Ooijen [3] have conducted research on work release order to study the influence on the throughput, throughput time and WIP. In this study, they aim to reduce the sum of the cost for these three factors. The job shop is examined by manipulating the WIP as compared to the total cost under instant release of arrival orders. From the experiment, the results indicated that cost reduction is considerable for small shops and for high ratios of WIP cost to lead time costs.
Accumulation of the WIP is a kind of waste for company especially for the semiconductor industry. Product cycles of semiconductor are short as the changeovers of the product are very fast. Qiu [4] researched the semiconductor company by using simulation to assess the WIP control system. The WIP control algorithms were tested by applying real time operation, where WIP levels, and machine utilization were evaluated in this study.
Sharma and Agrawal [5], employed AHP-algorithm in the control policy under different demand situation for the manufacturing system. The policy that they studied included, KANBAN, CONWIP and hybrid versions. Different demand patterns were studied by applying different control policies in order to obtain the behavior of each performance level. Other than this, Lin et al. [6] focused on finding an optimal WIP value of wafer fabrication processes by developing an algorithm integrating an artificial neural network (ANN) and the sequential quadratic programming (SQP) method. In this research paper, these methods offered an effective and systematic way to identify an optimal WIP level. Lin and Lee used queuing network-based algorithm to determine total standard of WIP and used a fixed WIP to implement the release control policy. The fixed controls WIP achieved the targeted throughput rate and the cycle times were kept relatively low. Isakow and Golany. [7], Yang et al. [8], and Li et al. [9] presented the study on constant work in process (CONWIP). The study proposed a model with the closed loop queue network and an approximate method was proposed in order to evaluate the system's performance. As the results were fairly precise, the genetic algorithm was designed to the proposed model to obtain the optimal results. In the review paper of Framinan et al. [10], a comprehensive and extensive of application CONWIP at the shop floor to control the WIP were discussed. Ip et al. [11] compared single loop and multi loop of CONWIP control for the lamp assembly production line which produce different kinds of products with discrete distribution processing time and demand. With respect to total cost and service level, a model was constructed with a novel rule-based genetic algorithm approach to determine the optimum parameter setting for multi loop CONWIP system. Results of the experiment showed that the single loop have better efficiency
control than the multi loop system whereby the total cost can be greatly decrease and WIP with zero shortage probability.
For developing and executing complex WIP models, the simulation approach has been the favored option for many researchers. Leitch [12] used the simulation approach to study the effect of stochastically, capacity, lead time on WIP and throughput. In his paper, it was found that these factors led to the cost drivers effect on WIP and throughput. Mirbirjandian and Wong [13] compared Kanban, CONWIP and Base-Stock in two kinds of essentially different production lines by using simulation study. In the simulation results, CONWIP showed the fewest average WIP and average WIP holding time as compared to the other systems. While for Morrice et al. [14] simulation analysis was used to count and predict the effect of internal on-time delivery, inventory and WIP change on the customer order fulfillment service level. In this way suitable control levels at various stages in a semiconductor supply chain, based on inventory and service level metrics can be identified.
From past researches, it was observed that most work was done for the wafer fabrication, and backend of the manufacturing line. There is not much work done to study WIP control at EOL Assembly of the semiconductor industry. In this paper, the analysis and study on the CONWIP based system in the EOL assembly is presented. The objective of this research is to identify the suitable level of WIP float in the EOL in order to decrease the holding of WIP and improve the cycle time. As the system to be studied is too complicated for analytical methods, models were developed an analyzed using the WITNESS simulation software.
3. Model Development, verification & validation
In the semiconductor industry, the products are high mix with each product family undergo different process routes. These product route, process times, machines are included in the simulation models. Setup time, breakdown time and conversion time for each of the machine are also defined. The distribution patterns for the unit per hour (UPH) of each machine for different products and material handling times are analyzed using the Minitab statistical software.
The input data for the base model were collected from the shop floor of the company. The base model was verified using the process flow diagram and by personnel in the shop floor. This is followed by the validation process. Operational validity is important to ensure the model's output behavior has the accuracy required for the models intended purpose [15].The throughput from a total of 10 replications runs were collected and tested with the Minitab statistical software to validate the model. The duration for each replication run is for a period of nine months. The warm up period is estimated to be six months.
Once the base model has been verified and validated, the experiments to test the CONWIP control policy were developed. Some assumptions made for the experimentations were that operators and material transportations were not included in the models and that all machines were available for all products.
Lot Arrival
Release Lot
Lot Leaving
Production Card
Fig. 2. CONWIP control system.
In Fig 2, the movements of the lots in the CONWIP system are shown. When a job arrives at the planning department, decision will be made as to whether to release the job for production. In this CONWIP system, the numbers of lots to be released to the EOL assembly is limited to the number of production cards available at the beginning of EOL assembly.
The logic flow diagram for the CONWIP system is shown in Fig 3. Based on the targeted throughput rate the numbers of production card are determined by the simulation model. The number of cards determines the WIP level for the assembly line. The study consists of two experimental models i.e. single loop and multi-loop CONWIP models.
Fig. 3. Framework of CONWIP flow 3(a) single loop CONWIP 3(b) multi loop CONWIP
For Fig 3(a) shows the single loop control which a uses single universal card in the system. Every machine or buffer shares and uses the same card. So when a lot arrive, the system will check if there are any cards available to authorize the lot to enter into the system. Each production card can only authorized one lot to be released into the system. The card is released back to the front the EOL assembly once the lot is withdrawn from the system. If there a queue in the buffer, the precedence will give to the same product family first.
Fig 3 (b) illustrates the multi loop CONWIP system. In this example there are 3 loops with 3 different production cards. Each card only can only be released back to the beginning of the respective loop once the lot leaves that specific
It shares the same concept as the single loop system but with each loop utilizing its own set of cards which cannot be used at other loops. For both the single and multi-loop system, there are a limited numbers of cards circulating to control the amount of WIP allowable in the system at any one time.
The objective for single loop experiment is to determine the appropriate CONWIP level for a targeted throughput rate. A CONWIP level ranging from 400 to 650 were analyzed.
Loop 1
Loop 2
Loop 3
Fig. 4. Processes of multi loop in CONWIP system.
While for the multi loop CONWIP experiment, three different loops are developed in the EOL assembly. The first loop includes the Moulding, Curing and Tie Bar Cut processes. While Loop 2 includes process by the curing process only and the last loop will includes the Manual Inspection, Trim & Form and Test processes. Figure 4 illustrates the processes in each loop.
The WIP in each loops are controlled at different levels and adjusted to find the best combination to meet the targeted throughput. For loop 1, the WIP are controlled from 200 to 220 lots, and the loop 2 controlled from 220 to 220 lots, and the last loop from 50 to 70 lots. The throughput rates are maintained at around 193 lots based on the production data. The screen capture of the developed base model in Witness simulation software is shown in fig 5. The results of the experiments are discussed in the next section.
Fig. 5. Witness simulation representation of the base model.
4. Result And Discussion.
For both experiments, the models were executed for ten replications runs. For each experiment, the average reading will be taken from each of the ten replication run. Different level of CONWIP will result in different output performance. The performance criteria for both experiments are: Work In Process (WIP), Cycle Time (CT), and Throughput (TH).
From the Table 1, single loop CONWIP level is represented by the number of production cards (ranged between 400 to 650). In the base model, average WIP was 575 lots, with CT of 3984 minutes and TH of 193 lots. This throughput rate is set as the target for the experimental models.
From the single loop CONWIP experiment, the WIP and CT increased with the numbers of production cards (please refer to Table 1 and Fig 6). The result is consistent with Little's law. However at the WIP of 440 levels, the throughput target is achieved. The CT is 3265.621, an 18.1% improvement from the base model. The improvement of WIP is 23% lower compared to the base model.
Table 1: single loop conwip results
No. of Production Cards TH CT WIP
400 192.2489 2981.763 400
410 192.5044 3051.663 410
420 192.7437 3121.006 420
430 192.9148 3192.657 430
440 193.0148 3264.621 440
450 193.1396 3336.135 450
500 193.2826 3695.78 500
550 193.3004 3922.818 545
600 193.3074 3976.837 568.7
650 193.3004 3983.964 574.9
Figure 6: Throughput and cycle time versus WIP. (a) Throughput versus WIP level in the CONWIP system. (b) Cycle time versus WIP level
in the CONWIP system.
Table 2 shows the result from the multi loop CONWIP model. Production Card in first 3 columns of table 2 represents the initial setting of WIP level for each loop while the average number of cards used is shown in columns 4-
6. For instances, the first 3 columns at the first row in table 2, the number of cards were set at 200, 200 and 50 for loops, L1, L2, L3 respectively.
The average number of cards used is actually 200, 190, and 39 respectively. Hence when L2 achieved WIP level 190, the remaining 10 cards are not utilized.
Table 2 Multi loop conwip result
Production Card (initial setting) Average Number of Cards Used (output from model) Output Parameters
L1 L2 L3 L1 L2 L3 CT WIP OUT-PUT
200 200 50 200 190 39 3240.46 430 190.75
200 200 60 200 189 38 3216.84 427 192.08
200 200 70 200 189 40 3212.93 428 192.29
200 210 50 200 192 40 3200.84 431 192.68
200 210 60 200 190 42 3194.15 431 192.78
200 210 70 200 190 38 3190.96 429 192.8
200 220 50 200 190 42 3185.59 432 192.84
200 220 60 197 192 44 3177.18 433 192.83
200 220 70 196 194 44 3175.97 434 192.84
210 200 50 210 190 42 3313.65 442 191.13
210 200 60 210 194 36 3290.16 440 192.4
210 200 70 210 190 40 3286.09 440 192.65
210 210 50 210 193 43 3274.31 446 192.9
210 210 60 210 188 40 3266.71 439 192.94
210 210 70 210 186 45 3267.33 441 192.84
210 220 50 210 182 45 3260.3 436 192.99
210 220 60 210 198 41 3254.3 449 193.04
210 220 70 210 191 42 3251.11 443 193
220 200 50 220 194 40 3387.23 454 191.36
220 200 60 219 198 35 3364.29 452 192.6
220 200 70 220 192 41 3359.1 453 192.79
220 210 50 220 187 43 3347.71 450 192.95
220 210 60 220 189 39 3337.57 449 193.04
220 210 70 220 192 44 3339.47 456 193
220 220 50 220 191 44 3331.82 455 193.13
220 220 60 216 184 44 3325.52 444 193.07
220 220 70 212 189 46 3324.24 447 193.09
Figure 7: Throughput and cycle time versus WIP. (a) Throughput versus WIP level in the CONWIP system. (b) Cycle time versus WIP level in the
CONWIP system.
The graph pattern for CT and WIP for multi loop CONWIP is shown in fig 7. For the CT graph, a series of peaks were observed for each setting of loop 1 when the WIP level of loop 3 is at minimum setting point (refer to labels in Table 2 and corresponding points in Fig 7). The values of peaks point in the fig 7 are highlighted in Table 2. This phenomenon occurs when there are queue in front of the loop 1 and loop 3 are limited to lesser WIP level to process the lots. These reasons caused to the waiting times of the part resulting in the increased cycle time. For the multi-loop experiment, the results show that at combination of 210, 220, 70 at loop 1, 2, 3 respectively, the CONWIP achieved the targeted throughput at CT of 3251.111. At this CONWIP level the CT is better than the CT of the base model (improvement of 18.4%). The improvement of WIP is 23% compared to the base model. For both the experiments, the results show the increasing upward trend of WIP and CT with the increase of production cards. As more cards are allowed in circulation the buffer size will become lager resulting in longer cycle times.For a fixed throughput of 193 lots, we compared the best average cycle time readings for the single loop (3264.621 min) and the multi loop (3251.111 mins.). Results show that the multi loop CONWIP performed better with a slightly lower CT (by 13.5 minutes) when compared to the best result for single loop system.
5. Conclusion and Recommendation
This study focused on the process in the EOL for a semiconductor company. As the manufacturing process is complex and involved a high mix of products, discrete-event simulation models using the WITNESS software were developed to evaluate the performance of the single loop and multi loop CONWIP systems. From the simulation results, both CONWIP systems outperform the existing system, by effectively controlling the WIP level and reducing the cycle time at the current throughput level. However the multi loop system performs better to reduce the CT as compared to the single loop system. The multi loop system is also more robust and easier to implement and control in the production line as compared to the single loop CONWIP system. In a multi loop system, cards are returned faster and the starvations due to bottleneck loops are minimized. As recommendations for future study, the multi loop CONWIP systems can be further subdivided for bottleneck processes or a hybrid control system can be embedded for better control of the production cards.
Acknowledgments
The authors wish to acknowledge that this project was supported by UTeM under grant number PJP/2011/FKP/ (22B) S00922.
References
[1] K. Miyashita and R. Gautam, "Multiagent coordination for controlling complex and unstable manufacturing processes." Expert System with Application, vol. 37(3), 2010, pp.1836-1845.
[2] Y. H. Lin and C. E. Lee, "A total standard WIP estimation method for wafer fabrication," European Journal of Operation Research, vol. 131(1), 2001, pp.78-94.
[3] JWM. Bertrand and V. HPG Ooijen, " Optimal work order release for make-to-order job shops with customer order lead-time costs, tardiness costs and work-in-^roess costs," vol. 116 International Journal of Production Economics, 2008, pp.233-241.
[4] R. G. Qiu, "Virtual production line based WIP control for semiconductor manufacturing systems," International Journal of Production Economics, vol. 95(2), 2005, pp.165-178.
[5] S. Sharma and N. Agrawal, "Selection of a pull production control policy under different demand situations for a manufacturing systems by AHP algorithm," Computers and Operations Research, vol. 36(5), 2009, pp.1622-1632.
[6] Y. H. Lin, C. C. Chiu and C.H. Tsai, "The study of applying ANP model to assess dispatching rules for wafer fabrication," Expert Systems with Applications: An International Journal, vol. 34 n.3, pp.2148-2163, April 2008.
[7] S. A. Isakow, and B. Golany, "Managing multi-project environments through constant work-in-process." International Journal of Project Management, vol. 21(1), 2003, pp.9-18.
[8] T. Yang, H. P. Fu, and K. Y. Yang, "An evolutionary-simulation approach for the optimization of multi-constant work-in-process strategy-A case study," Int. J. Production Economics, vol. 107(1), 2007, pp.104-114.
[9] N. Li, S. Yao, G. Liu, and, C. Zhuang, "Optimization of a multi-Constant Work-in-Process semiconductor assembly and test factory based on performance evaluation," Computers & Industrial Engineering, vol. 59, 2010, pp.314—322.
[10] J. M. Framinan, P. L. Fonzalez, and R. R Usano, "The CONWIP production control system; review and research issues," Production Planning and Control, vol.14, 2003, pp.255-265
[11] W.H. Ip, M. Huang, K.L. Yung, D. Wang and X. Wang, "CONWIP based control of a lamp assembly production line," Journal of Intelligent Manufacturing," vol. 18, no. 2,2007, pp. 261 -271.
[12] R. A. Leitch, "Effect of stochasticity, capacity and lead time cost drivers on WIP and throughput in a pull production environment," Management Accounting Research, vol. 12(2), 2001, pp.167—196.
[13] E. Mirbirjandian and K. Y. Wong, "A Simulation Study on Kanban Control Systems," 17th IBIMA Conference on Creating Global Competitive Economies: A 360-degree Approach, 2011.
[14] D. J. Momee, R. A. Valdez, Jr. J. P. Chida, and M. Eido, "Discrete event simulation in supply chain planning and inventory control at Freescale semiconductor, inc.," Proceedings of the 2005 Winter Simulation Conference, pp.1718-1724, 2005.
[15] R. G. Sargait, "Verification and Validation of Simulation Models' Proceedings of the 1998 Winter Simulation Conference, pp.121-130, 1998
