- Email: [email protected]
Collaborative production line control for collaborative supply networks
Proceedings of the 14th IFAC Symposium on Information Control Problems in Manufacturing Bucharest, Romania, May 23-25, 2012
Collaborative production line control for collaborative supply networks Rodrigo Reyes Levalle. Manuel Scavarda. Shimon Y. Nof PRISM Center, School of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA ([email protected], [email protected], [email protected]) Abstract: Industries with automated discrete production lines struggle with intrinsic process variability affected, among other factors, by the rules that coordinate interactions among machines. This problem involves uncertain and dynamic conditions, and the need for collaboration and conflict resolution mechanisms to minimize disruptions in the production flow and achieve high process efficiency to meet customer demand with minimum inventory levels. A Collaborative Control Theory approach based on three of its principles is developed to solve this problem. Its performance is compared to common industry standard control methods. A case study shows that Collaborative Production Line Control (CPLC) outperform traditional models, achieving a 43% reduction in throughput variability, 24% reduction in WIP, and a 0.5% increase in service level. Keywords: Automation, Collaborative Control Theory, Production Systems, Inventory Control, WIP.
1. INTRODUCTION Since the development of the assembly line in the early ¶V, industry has sought ways to improve production OLQHV¶SHUIRUPDQFHWRdeliver more products, with lower lead time and better quality. Economic sustainability has become a central issue in many industries, especially those producing commodity products, where companies compete on price and customer service. From the manufacturing perspective, economic sustainability through higher profit margins has been pursued in the early years by seeking economies of scale. As a result, many production lines acquired the capability of producing a high volume of products in short periods of time, making discrete production lines more similar to continuous processes, where automation plays a key role. As demand for commodity products became asymptotic, marginal gains obtained by investing in state-ofthe-art manufacturing equipment banished and a new era of improvement by reducing waste began. To achieve higher performance in production lines it became necessary to standardize operating procedures, attain greater coordination among automated, high-volume processes, and develop and implement optimal buffer management techniques. Several models for buffer management and production control can be found in industrial engineering literature but most are applicable only to assembly lines with low production rates or high stability. In most cases, high volume production lines that have not reached stability face a dearth of methodologies to deal with intrinsic process variability, especially in real time. 2. LITERATURE REVIEW 2.1 Traditional approaches to production line control Many researchers have studied the problem of production line control under uncertain machine behavior and proposed different methodologies to increase throughput while minimizing WIP. One of the most basic control policies is buffer management through a base-stock policy. Different 978-3-902661-98-2/12/$20.00 © 2012 IFAC
methods for defining target base-stock levels have been proposed by Güllü (1998), Yazdan Shenas et al. (2009), and Duenyas and Patana-Anake (1998). Performance of the basestock policy and Kanban was compared by Duri et al. (2000) FRQFOXGLQJWKDWWKHODWWHU³«LVHDVLHUWRLPSOHPHQWDQGOLPLWV WKHPD[LPXPDPRXQWRI:,3´ Kanban, the WIP control mechanism from the Toyota Production System, introduced by Sugimori et al. (1977) is perhaps the most widely known buffer management system. An extensive study of the literature on the Kanban System, its principles and operating conditions can be found in the work of Sendil Kumar and Panneerselvam (2006). A classification of thirty two Kanban variations can be found in the work of Lage Junior and Godinho Filho (2010). Andijani (1997) showed that the allocation rules for kanban cards can affect the average throughput rate. Deleersnydhr et al. (1989) studied the effects of machine reliability and demand variation on Kanban systems, and Savsar (1996) studied the effects of fixed cycle and fixed quantity allocation policies on throughput and WIP. CONWIP, described by Sendil Kumar and Panneerselvam (2006) DV ³«D JHQHUDOL]HG IRUP RI NDQEDQ´ ZDV SURSRVHG by Spearman et al. (1990). A study of the differences between CONWIP and Kanban can be found in the work of Pettersen and Segerstedt (2009). Bonvik et al. (1997) analyzed the performance of Kanban, base-stock, and CONWIP policies and concluded that Kanban amplifies the process variability while the other policies dampen it. Yang et al. (2006) addressed &21:,3¶V ODFN RI DGDSWLYH capabilities by introducing Dynamic WIP, a solution approach that uses real time information about machine status and demand estimations to dynamically update WIP maximum levels in CONWIP control loops. Although DWIP fails to benefit from collaboration among parallel subassembly lines, performance of the DWIP algorithm is proven to be better than CONWIP on systems with low and high disturbances. Dynamic control strategy-updating based on system status was also studied by Paternina-Arboleda and
487
10.3182/20120523-3-RO-2023.00316
INCOM 2012, May 23-25, 2012 Bucharest, Romania
Das (2001), who proposed a learning agent methodology to control a production line based on the system status.
Table 1. Traditional and collaborative methodologies summary
Table 1 categorizes the reviewed methods based on four factors; buffer management, production rates management, real-time decision making, and system-state evolution forecast, clarifying the similarity among them. 2.2 Collaborative Control Theory approach to production line control Buffer management methods described in the previous section focus their control over the production line on a single variable, WIP levels. Alternative methods for line control based on two or more variables have been developed, relying on certain level of collaboration between processes to overcome unexpected downtimes, maintaining throughput while simultaneously reducing the required level of WIP. Collaborative Control Theory, CCT, introduced by Nof (2003; 2007), comprises six design principles which provide a common analysis framework to enable the control of the collaboration level of different systems. Ma and Koren (2004) LQWURGXFHG WKH ³2SWLPDO &RQWUROOHU 6HOHFWLRQ´ PHWKRG based on the optimization of the system performance for each possible system state using real time data. This methodology partly applies the CCT principles of cooperation requirement planning, by using optimization models to obtain optimal configuration for each system status and enable collaboration between the production line processes to achieve target throughput and minimize WIP, and fault tolerance by increasing WKH V\VWHP¶V UREXVWQHVV when running the optimization under failure conditions. To further enable fault tolerance in 0DDQG.RUHQµV work, a method for anticipating system state duration is needed. Such method is introduced in the Simulation-based Real-time Decision-Making, SRDM, methodology proposed by Dalal et al. (2003). Similar to the CCT framework, the SRDM methodology applies the principle of conflict and error GHWHFWLRQ DQG SURJQRVWLFV E\ VLPXODWLQJ WKH V\VWHP¶V behavior under a given number of different policies, allowing decision makers to decide on the best policy based on expected performance under future problems. Table 1 categorizes the reviewed collaborative methods according to the four factors used for traditional approaches. Evidence from the reviewed papers regarding the potential benefits of developing a production control methodology based on the principles of Collaborative Control Theory is clear. Therefore, the objective of this research paper is to introduce CCT principles to the solution of the production line control problem under uncertain demand and equipment behavior to create a novel solution methodology. The study is organized as follows: In the methodology section a mathematical definition of the problem is introduced. Second, the CCT based solution methodology is developed. Third, an experimental comparison of the proposed methodology and other solution approaches is made to evaluate its performance. Finally, a conclusion from the results obtained in the experiments and future work directions are presented.
3.
METHODOLOGY
3.1 Standardized mathematical model for production lines Production lines can be highly different across industries; moreover, production lines within the same industry can be very dissimilar when considering the possibility of having different technologies to perform a given operation or diverse machine arrangements and layouts. Nevertheless, all production lines share features that can be treated as equal in order to model their behavior using a common mathematical model. To do so, it is necessary to divide the production line into its elemental components, processes, and buffers which describe the two main activities through which raw materials and WIP can flow: transformations, and storage. 3.1.1 The building blocks: Process and buffer models A process is any activity on materials that adds value and, in general terms, has physical and logical inputs and outputs. Physical I/O are related to material flow, i.e., raw materials and items produced. A set of functions relates physical inputs to outputs, e.g., rate of production. Logical I/O comprises the variables and parameters required by the process control logic and the state variables that describe the process at any given point in time. Buffers, defined as storage systems or areas that follow a given process (Fig. 1) have two main purposes: store the finished parts of the preceding process, and smooth system throughput. Buffer levels, Lb, are controlled by line or
488
INCOM 2012, May 23-25, 2012 Bucharest, Romania
process interaction policies whose goal is to achieve and hold a target level, TLb.
regulated by control policies g . When changes in system state occur, different g may be enforced to regulate the V\VWHP¶VSHUIRUPDQFH/et E be the set of gr applied on the system, where t is the instant of enforcement for control policy i, then E is defined as the control strategy for the production line. If a production line is run for a given period of time where state changes occur and their durations ¿ r` are known a priori, the best control strategy can be calculated based on the available information. E L <r, â å â r` =, where
Fig. 1. Production process and buffer models 3.1.2 Constructing a production line model Process/buffer pairs can be arranged in combinations to form a line. Several lines can be set in parallel and connected prior to or after an assembly operation or a common storage area. An intersection point is a physical location where two or more lines connect. Intersection points have bill of materials constraints and a control policy to decide which of the exiting lines is served, when more than one exist.
rc
(3)
L : á ¿ á å á á ¿ ; rc
rc
r`
r`
However, due to the stochastic nature of the system, information on state changes is unknown a priori. Consequently, the question to be solved is: How should a control strategy be selected to maintain constant throughput rate while simultaneously minimizing WIP inventory? 3.2 Collaborative Production Line Control (CPLC) To explore the benefits of enabling collaborative resolution for the problem of production line control, a control methodology named Collaborative Production Line Control (CPLC) was designed based on three principles of Collaborative Control Theory. i. Cooperation Requirement Planning (CRP) ii. Conflict and Error Detection and Prognostics (CEDP) iii. Fault-Tolerance by Teaming (FTT) An extensive explanation of the principles comprising CCT can be found in Nof (2007; 2009).
Fig. 2. Production lines and intersection point (IP) 3.1.3 Enabling interaction between the building blocks: Production line control At any given point in time t, the production line can be characterized by the status of the processes and the status of the buffers. In a network with n processes, there will be t system state variablesir characterizing the system at any r point in time t. Let6n?5 define process p according to sá LEOKG ç O6ã?5 L\ L L sá tá å á J rá LEOB=EHA@
CPLC starts with the CRP-I step by defining the best static control policy (Rp and TLb) for every possible combination of process states assuming that this condition is in steady state. To illustrate this process, consider the example production line of Fig. 3. Taking into account only processes, there are eight different states, i.e., all process working (steady state), one process failed (three states, one for each line), two processes failed (three states) and all processes failed.
(1)
ç Correspondingly, let O6Õ define buffer b according to
sá .Õ F 6.Õ Q ÜÕ ç O6Õ L\ > L sá tá å á J rá .Õ F 6.Õ P ÜÕ
(2)
where A` is the tolerance for TLb deviation.
Fig. 3. Example production line
Let r be a vector of t components such that r L r ; :5r á å á 6l then the system state at time t is fully described r by . Since all ir are binary, the number of possible system states at a given time t is t6á , therefore the number of states grows exponentially with increasing system complexity.
The CRP-I step involves determining the bottleneck in each state, setting all non-bottleneck process speeds according to the bottleneck capacity to cooperate in maximizing throughput, and all buffers targets to low values, ideally zero. The results obtained are stored in a database to be retrieved by the algorithm of the second phase, CRP-II.
The interactions among the lines that constitute the network and among processes that compose each of the lines are
489
INCOM 2012, May 23-25, 2012 Bucharest, Romania
CRP-II takes place during the execution of the balanced line configuration, defined in CRP-I, by making adjustments to Rp and TLb based on the current system progress. To adjust the control variables, information is gathered from the system, the CRP-I database and an early conflict detection tool (ECDT). The principle of CEDP is applied to create the ECDT which feeds from historical machine time between failures and provides an estimate for the probability of each machine to be the next to fail. To illustrate how this information is used in CRP-II, consider the system of Fig. 3 operating in steady state and information from the ECDT indicating a high probability of failure in process on line 3. Information from the CRP-I database shows that under this failure condition throughput is reduced. As a result, the algorithm increases speed in line 2 process. This has a double objective; to reduce the level in line 1 buffer in order to prevent bottleneck idling, and to maintain system throughput as long as possible after the failure occurs. Following this logic, the decision making algorithm is able to anticipate the next system state change, enabling conflict prevention and collaboration among machines to provide the next expectedto-fail machine a better condition to face the likely failure.
decide in real time which is the best control strategy. Often, the lack of tools to assist them in this task prevents them from making decisions that will lead to a higher throughput and lower WIP inventory. A list of the most relevant interaction policies indicating equipment involved, trigger conditions, resulting actions and ZKHWKHU WKH SROLF\ LV DXWRPDWLF FRQWUROOHG E\ SOF¶V RU manual (implemented by the operator) is shown in Table 2 Table 2. Main interaction policies summary
Fig. 4. Collaborative Production Line Control algorithm 4. EXPERIMENTAL RESULTS A tissue converting line was modeled in ARENA in order to test the feasibility of CPLC and compare its performance to other control methods.
To test the performance of the proposed methodology, an ARENA model was run using real data from a certain industry. Five different control policies were tested using the model; by simulating one week of production after one day of warm up in order to obtain steady state indicators. Each simulation run was replicated 100 times to obtain statistically valid inferences on the indicators. The results obtained are shown in Table 3 and Fig. 6. Table 3. Simulation Results
Fig. 5. Tissue converting line model The simulated line, depicted in Fig. 5 is a high volume automated production system which is often challenged by the succession and interaction of machine failures that lead to line downtimes and reduction of throughput. The extensive set of control logics that regulate the interaction of equipment within the line makes it difficult for job floor leaders to 490
INCOM 2012, May 23-25, 2012 Bucharest, Romania
Experimental conclusions, though validated only for the real system data analyzed in the case study, may be applicable to other systems and system configurations, but further studies to generalize the conclusions are required. Table 5. Methods ranking and contributions summary
Fig. 6. Boxplot for WIP and throughput (cs=cases) In order to compare service levels and WIP from different policies, two-sample unpooled t-tests considering unequal variances were performed with a significance level of 1%. The results are summarized in Table 4 where each cell shows the result of the hypothesis test comparing the model in the column with the model in the row. Table 4. Hypotheses tests summary
5.1 From Collaborative Production Collaborative Supply Networks 5. CONCLUSIONS Production line control under uncertainty conditions requires a highly adaptive and anticipative methodology that enables collaboration between the different components of the line to overcome failures and maintain high throughput while keeping WIP low. Traditional models for buffer management and production control lack the required characteristics. The proposed methodology, CPLC EDVHG RQ &&7¶V SULQFLSOHV, constitutes a more advanced and complete solution approach allowing both adaptability and anticipation, more suitable to face the challenges of high volume production line control under uncertainty conditions. Based on the observations made during the simulation runs and the conclusions for the case study, the methods analyzed can be classified as shown in Table 5. Statistical comparisons of the results in this case indicate that CPLC outperforms Base-stock, CONWIP, DWIP and Kanban in terms of service level with a significance level of 1%. Also, the hypothesis tests show that it is not possible to conclude that there are statistically significant differences in service level within the group of Kanban, DWIP, CONWIP and Base-stock. The statistical analysis also allows clustering models in three performance groups in terms of WIP; i.e. Base-stock and CONWIP, DWIP and Kanban, and CPLC, with a difference at significance level of 1%. As for throughput variability, it can be concluded that CPLC reduces variability, measured as the ratio between throughput standard deviation and mean. Performing the statistical comparison for comparing CPLC with the remaining methods shows that variability is reduced in every case.
Line
Control
to
Collaborative Production Line Control enables a production line to achieve higher throughput with lower variability. In other words, materials flow through the production lines faster and more evenly. The proposed methodology detects possible conflicts emerging in the near future and addresses them by utilizing information from CRP-I. This means that CPLC is capable of predicting where material flow interruptions may occur in a production line, finding alternative ways to maintain product flow, and even minimize its reduction. In Collaborative Supply Networks (CSN) material flow through the supply chain can also be controlled by using the same principles contained in CPLC. In the case of a CSN (Fig. 7) flow interruptions may appear as product stockouts in DCs and XDs, raw material shortage from suppliers, lack of production capacity in a given facility, among others. These situations are analogous to downtimes in equipment in a production line that prevent materials from flowing across them. In a CSN, CRP-I will consider alternative ways to convey products from the first supplier in tier n to the retailers in the opposite side of the supply chain, under different flow interruption situations. The optimal supply flow in each case should be defined based on the cost of moving all the required materials through the different chosen points of the supply chain in order to deliver products to the retailers. As for the early conflict detection capabilities, visibility through the entire supply chain can be embedded in the model, in order to provide decision makers with the necessary status information to be able to predict possible flow interruption in the near future. Therefore, an information sharing system enabling parties to share insights must be implemented to create an ECDT for CSN.
491
INCOM 2012, May 23-25, 2012 Bucharest, Romania
Fig. 7. Collaborative Supply Network (CSN): DC=distribution center; XD=Cross dock Implementation of a Collaborative Control Theory approach to CSN control can result in major improvement in service level and reduction in lead time and total supply cost. 6. FUTURE RESEARCH It has been proven that production line control based on CCT can produce relatively high line performance while keeping low WIP levels in high volume production lines under uncertainty conditions. The methodology introduced in this study considers only the first system state evolution options to define the new control strategy. More complex path evolution analysis models for depicting possible future system states and their associated probabilities of occurrence can be integrated into CPLC to extend its prediction accuracy, enabling it to reach better control decisions. Although CPLC was proven to be an effective buffer control policy for the case study at hand, it is still a theoretical methodology and future research should define the necessary requisites for it to be viable for implementation in plants. Another important research direction is collaborative supply networks with the extension of the CPLC methodology presented in this study to the general case of CSNs. Additional research must also validate the CPLC in CSNs and quantify the improvements in service level, reduction in lead time and total supply cost that can be achieved through a Collaborative Control Theory approach to CSN control. ACKNOWLEDGEMENT This research has been developed with partial support from the PRISM (Production, Robotics, and Integration Software for Manufacturing & Management) Center at Purdue University. REFERENCES Andijani, A. (1997). Trade-off between maximizing throughput rate and minimizing system time in kanban systems. International Journal of Operations & Production Management, Vol. 17 No. 5, pp. 429-445. Bonvik, A., Couch, C., and Gershwin, S. (1997). A comparison of production-line control mechanisms. International Journal of Production Research, Vol. 35 No. 3, pp. 789-804. Dalal, M., Groel, B., and Prieditis, A. (2003). Real-time decision making using simulation. Proceedings of the 2003 Winter Simulation Conference, pp. 1456-1464. Deleersnydhr, J., Hodgson, T., Muller, H., and O'Grady, P. (1989). Kanban controlled pull systems: An analytic
approach. Management Science Vol. 35, No. 9, pp. 1079-1091. Duenyas , I., and Patana-Anake, P. (1998). Base-stock control for single-product tandem make-to-stock systems. IIE Transactions, Vol 30 No. 1 , pp. 31-39. Duri, C., Frein, Y., and Di Mascolo, M. (2000). Comparison among three pull control policies: kanban, base stock, and generalized kanban. Annals of Operations Research, Vol. 93, pp. 41±69. Güllü, R. (1998). Base stock policies for production/inventory problems with uncertain capacity levels. European Journal of Operational Research, Vol. 105 No. 1, pp. 43-51. Lage Junior, M., and Godinho Filho, M. (2010). Variations of the Kanban system: Literature review and classification. International Journal of Production Economics, Vol. 125 No. 1, pp. 13-21. Ma, Y., and Koren, Y. (2004). Operation of manufacturing systems with work-in-process inventory and production control. CIRP Annals - Manufacturing Technology Vol. 53 No. 1, pp. 361-365. Nof, S. Y. (2003). Design of effective e-Work: review of models, tools, and emerging challenges. Production Planning & Control, Vol. 14 No.8, pp. 681-703. Nof, S. Y. (2007). Collaborative Control Theory for e-Work, e-Production, and e-Service. Annual Reviews in Control IFAC, Vol. 31 No.2, pp. 281-292. Nof, S. Y. (2009). Springer Handbook of Automation, pp. 1549. Springer, United States. Paternina-Arboleda, C., and Das, T. (2001). Intelligent dynamic control policies for serial production lines. IIE Transactions, Vol. 33 No. 1, pp. 65-77. Pettersen, J., and Segerstedt, A. (2009). Restricted work-inprocess: A study of differences between Kanban and CONWIP. International Journal of Production Economics, Vol. 118 No. 1, pp. 199±207. Savsar, M. (1996). Effects of Kanban withdrawal policies and other factors on the performance of JIT systems - A simulation study. International Journal of Production Research, Vol. 34, No. 10, pp. 2879-2899. Sendil Kumar, C., and Panneerselvam, R. (2006). Literature review of JIT-KANBAN system. International Journal of Advanced Manufacturing Technology 2007, Vol. 32 No. 3-4 , pp. 393-408. Spearman, M., Woodruff, D., and Hopp, W. (1990). CONWIP: a pull alternative to kanban. International Journal of Production Research, Vol. 28 No. 5, pp. 879894. Sugimori, Y., Kusunoki, K., Cho, F., and Uchikawa, S. (1977). Toyota production system and Kanban system: Materialization for just-in-time and respect-for-human system. International Journal of Production Research, Vol. 15 No. 6, pp. 553-564. Yadzan Shenas, N., Eshraghniaye Jahromi, A., and Modarres, M. (2009). A new approach to find an optimal solution for base stock policies. Journal of Applied Sciences, Vol. 9 No. 4, pp. 789-793. Yang, R., Subramaniam, V., and Gershwin, S. (2006). Setting real time WIP levels in production lines. Innovation in Manufacturing Systems and Technology ± MIT.
492
Collaborative production line control for collaborative supply networks Rodrigo Reyes Levalle. Manuel Scavarda. Shimon Y. Nof PRISM Center, School of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA ([email protected], [email protected], [email protected]) Abstract: Industries with automated discrete production lines struggle with intrinsic process variability affected, among other factors, by the rules that coordinate interactions among machines. This problem involves uncertain and dynamic conditions, and the need for collaboration and conflict resolution mechanisms to minimize disruptions in the production flow and achieve high process efficiency to meet customer demand with minimum inventory levels. A Collaborative Control Theory approach based on three of its principles is developed to solve this problem. Its performance is compared to common industry standard control methods. A case study shows that Collaborative Production Line Control (CPLC) outperform traditional models, achieving a 43% reduction in throughput variability, 24% reduction in WIP, and a 0.5% increase in service level. Keywords: Automation, Collaborative Control Theory, Production Systems, Inventory Control, WIP.
1. INTRODUCTION Since the development of the assembly line in the early ¶V, industry has sought ways to improve production OLQHV¶SHUIRUPDQFHWRdeliver more products, with lower lead time and better quality. Economic sustainability has become a central issue in many industries, especially those producing commodity products, where companies compete on price and customer service. From the manufacturing perspective, economic sustainability through higher profit margins has been pursued in the early years by seeking economies of scale. As a result, many production lines acquired the capability of producing a high volume of products in short periods of time, making discrete production lines more similar to continuous processes, where automation plays a key role. As demand for commodity products became asymptotic, marginal gains obtained by investing in state-ofthe-art manufacturing equipment banished and a new era of improvement by reducing waste began. To achieve higher performance in production lines it became necessary to standardize operating procedures, attain greater coordination among automated, high-volume processes, and develop and implement optimal buffer management techniques. Several models for buffer management and production control can be found in industrial engineering literature but most are applicable only to assembly lines with low production rates or high stability. In most cases, high volume production lines that have not reached stability face a dearth of methodologies to deal with intrinsic process variability, especially in real time. 2. LITERATURE REVIEW 2.1 Traditional approaches to production line control Many researchers have studied the problem of production line control under uncertain machine behavior and proposed different methodologies to increase throughput while minimizing WIP. One of the most basic control policies is buffer management through a base-stock policy. Different 978-3-902661-98-2/12/$20.00 © 2012 IFAC
methods for defining target base-stock levels have been proposed by Güllü (1998), Yazdan Shenas et al. (2009), and Duenyas and Patana-Anake (1998). Performance of the basestock policy and Kanban was compared by Duri et al. (2000) FRQFOXGLQJWKDWWKHODWWHU³«LVHDVLHUWRLPSOHPHQWDQGOLPLWV WKHPD[LPXPDPRXQWRI:,3´ Kanban, the WIP control mechanism from the Toyota Production System, introduced by Sugimori et al. (1977) is perhaps the most widely known buffer management system. An extensive study of the literature on the Kanban System, its principles and operating conditions can be found in the work of Sendil Kumar and Panneerselvam (2006). A classification of thirty two Kanban variations can be found in the work of Lage Junior and Godinho Filho (2010). Andijani (1997) showed that the allocation rules for kanban cards can affect the average throughput rate. Deleersnydhr et al. (1989) studied the effects of machine reliability and demand variation on Kanban systems, and Savsar (1996) studied the effects of fixed cycle and fixed quantity allocation policies on throughput and WIP. CONWIP, described by Sendil Kumar and Panneerselvam (2006) DV ³«D JHQHUDOL]HG IRUP RI NDQEDQ´ ZDV SURSRVHG by Spearman et al. (1990). A study of the differences between CONWIP and Kanban can be found in the work of Pettersen and Segerstedt (2009). Bonvik et al. (1997) analyzed the performance of Kanban, base-stock, and CONWIP policies and concluded that Kanban amplifies the process variability while the other policies dampen it. Yang et al. (2006) addressed &21:,3¶V ODFN RI DGDSWLYH capabilities by introducing Dynamic WIP, a solution approach that uses real time information about machine status and demand estimations to dynamically update WIP maximum levels in CONWIP control loops. Although DWIP fails to benefit from collaboration among parallel subassembly lines, performance of the DWIP algorithm is proven to be better than CONWIP on systems with low and high disturbances. Dynamic control strategy-updating based on system status was also studied by Paternina-Arboleda and
487
10.3182/20120523-3-RO-2023.00316
INCOM 2012, May 23-25, 2012 Bucharest, Romania
Das (2001), who proposed a learning agent methodology to control a production line based on the system status.
Table 1. Traditional and collaborative methodologies summary
Table 1 categorizes the reviewed methods based on four factors; buffer management, production rates management, real-time decision making, and system-state evolution forecast, clarifying the similarity among them. 2.2 Collaborative Control Theory approach to production line control Buffer management methods described in the previous section focus their control over the production line on a single variable, WIP levels. Alternative methods for line control based on two or more variables have been developed, relying on certain level of collaboration between processes to overcome unexpected downtimes, maintaining throughput while simultaneously reducing the required level of WIP. Collaborative Control Theory, CCT, introduced by Nof (2003; 2007), comprises six design principles which provide a common analysis framework to enable the control of the collaboration level of different systems. Ma and Koren (2004) LQWURGXFHG WKH ³2SWLPDO &RQWUROOHU 6HOHFWLRQ´ PHWKRG based on the optimization of the system performance for each possible system state using real time data. This methodology partly applies the CCT principles of cooperation requirement planning, by using optimization models to obtain optimal configuration for each system status and enable collaboration between the production line processes to achieve target throughput and minimize WIP, and fault tolerance by increasing WKH V\VWHP¶V UREXVWQHVV when running the optimization under failure conditions. To further enable fault tolerance in 0DDQG.RUHQµV work, a method for anticipating system state duration is needed. Such method is introduced in the Simulation-based Real-time Decision-Making, SRDM, methodology proposed by Dalal et al. (2003). Similar to the CCT framework, the SRDM methodology applies the principle of conflict and error GHWHFWLRQ DQG SURJQRVWLFV E\ VLPXODWLQJ WKH V\VWHP¶V behavior under a given number of different policies, allowing decision makers to decide on the best policy based on expected performance under future problems. Table 1 categorizes the reviewed collaborative methods according to the four factors used for traditional approaches. Evidence from the reviewed papers regarding the potential benefits of developing a production control methodology based on the principles of Collaborative Control Theory is clear. Therefore, the objective of this research paper is to introduce CCT principles to the solution of the production line control problem under uncertain demand and equipment behavior to create a novel solution methodology. The study is organized as follows: In the methodology section a mathematical definition of the problem is introduced. Second, the CCT based solution methodology is developed. Third, an experimental comparison of the proposed methodology and other solution approaches is made to evaluate its performance. Finally, a conclusion from the results obtained in the experiments and future work directions are presented.
3.
METHODOLOGY
3.1 Standardized mathematical model for production lines Production lines can be highly different across industries; moreover, production lines within the same industry can be very dissimilar when considering the possibility of having different technologies to perform a given operation or diverse machine arrangements and layouts. Nevertheless, all production lines share features that can be treated as equal in order to model their behavior using a common mathematical model. To do so, it is necessary to divide the production line into its elemental components, processes, and buffers which describe the two main activities through which raw materials and WIP can flow: transformations, and storage. 3.1.1 The building blocks: Process and buffer models A process is any activity on materials that adds value and, in general terms, has physical and logical inputs and outputs. Physical I/O are related to material flow, i.e., raw materials and items produced. A set of functions relates physical inputs to outputs, e.g., rate of production. Logical I/O comprises the variables and parameters required by the process control logic and the state variables that describe the process at any given point in time. Buffers, defined as storage systems or areas that follow a given process (Fig. 1) have two main purposes: store the finished parts of the preceding process, and smooth system throughput. Buffer levels, Lb, are controlled by line or
488
INCOM 2012, May 23-25, 2012 Bucharest, Romania
process interaction policies whose goal is to achieve and hold a target level, TLb.
regulated by control policies g . When changes in system state occur, different g may be enforced to regulate the V\VWHP¶VSHUIRUPDQFH/et E be the set of gr applied on the system, where t is the instant of enforcement for control policy i, then E is defined as the control strategy for the production line. If a production line is run for a given period of time where state changes occur and their durations ¿ r` are known a priori, the best control strategy can be calculated based on the available information. E L <r, â å â r` =, where
Fig. 1. Production process and buffer models 3.1.2 Constructing a production line model Process/buffer pairs can be arranged in combinations to form a line. Several lines can be set in parallel and connected prior to or after an assembly operation or a common storage area. An intersection point is a physical location where two or more lines connect. Intersection points have bill of materials constraints and a control policy to decide which of the exiting lines is served, when more than one exist.
rc
(3)
L : á ¿ á å á á ¿ ; rc
rc
r`
r`
However, due to the stochastic nature of the system, information on state changes is unknown a priori. Consequently, the question to be solved is: How should a control strategy be selected to maintain constant throughput rate while simultaneously minimizing WIP inventory? 3.2 Collaborative Production Line Control (CPLC) To explore the benefits of enabling collaborative resolution for the problem of production line control, a control methodology named Collaborative Production Line Control (CPLC) was designed based on three principles of Collaborative Control Theory. i. Cooperation Requirement Planning (CRP) ii. Conflict and Error Detection and Prognostics (CEDP) iii. Fault-Tolerance by Teaming (FTT) An extensive explanation of the principles comprising CCT can be found in Nof (2007; 2009).
Fig. 2. Production lines and intersection point (IP) 3.1.3 Enabling interaction between the building blocks: Production line control At any given point in time t, the production line can be characterized by the status of the processes and the status of the buffers. In a network with n processes, there will be t system state variablesir characterizing the system at any r point in time t. Let6n?5 define process p according to sá LEOKG ç O6ã?5 L\ L L sá tá å á J rá LEOB=EHA@
CPLC starts with the CRP-I step by defining the best static control policy (Rp and TLb) for every possible combination of process states assuming that this condition is in steady state. To illustrate this process, consider the example production line of Fig. 3. Taking into account only processes, there are eight different states, i.e., all process working (steady state), one process failed (three states, one for each line), two processes failed (three states) and all processes failed.
(1)
ç Correspondingly, let O6Õ define buffer b according to
sá .Õ F 6.Õ Q ÜÕ ç O6Õ L\ > L sá tá å á J rá .Õ F 6.Õ P ÜÕ
(2)
where A` is the tolerance for TLb deviation.
Fig. 3. Example production line
Let r be a vector of t components such that r L r ; :5r á å á 6l then the system state at time t is fully described r by . Since all ir are binary, the number of possible system states at a given time t is t6á , therefore the number of states grows exponentially with increasing system complexity.
The CRP-I step involves determining the bottleneck in each state, setting all non-bottleneck process speeds according to the bottleneck capacity to cooperate in maximizing throughput, and all buffers targets to low values, ideally zero. The results obtained are stored in a database to be retrieved by the algorithm of the second phase, CRP-II.
The interactions among the lines that constitute the network and among processes that compose each of the lines are
489
INCOM 2012, May 23-25, 2012 Bucharest, Romania
CRP-II takes place during the execution of the balanced line configuration, defined in CRP-I, by making adjustments to Rp and TLb based on the current system progress. To adjust the control variables, information is gathered from the system, the CRP-I database and an early conflict detection tool (ECDT). The principle of CEDP is applied to create the ECDT which feeds from historical machine time between failures and provides an estimate for the probability of each machine to be the next to fail. To illustrate how this information is used in CRP-II, consider the system of Fig. 3 operating in steady state and information from the ECDT indicating a high probability of failure in process on line 3. Information from the CRP-I database shows that under this failure condition throughput is reduced. As a result, the algorithm increases speed in line 2 process. This has a double objective; to reduce the level in line 1 buffer in order to prevent bottleneck idling, and to maintain system throughput as long as possible after the failure occurs. Following this logic, the decision making algorithm is able to anticipate the next system state change, enabling conflict prevention and collaboration among machines to provide the next expectedto-fail machine a better condition to face the likely failure.
decide in real time which is the best control strategy. Often, the lack of tools to assist them in this task prevents them from making decisions that will lead to a higher throughput and lower WIP inventory. A list of the most relevant interaction policies indicating equipment involved, trigger conditions, resulting actions and ZKHWKHU WKH SROLF\ LV DXWRPDWLF FRQWUROOHG E\ SOF¶V RU manual (implemented by the operator) is shown in Table 2 Table 2. Main interaction policies summary
Fig. 4. Collaborative Production Line Control algorithm 4. EXPERIMENTAL RESULTS A tissue converting line was modeled in ARENA in order to test the feasibility of CPLC and compare its performance to other control methods.
To test the performance of the proposed methodology, an ARENA model was run using real data from a certain industry. Five different control policies were tested using the model; by simulating one week of production after one day of warm up in order to obtain steady state indicators. Each simulation run was replicated 100 times to obtain statistically valid inferences on the indicators. The results obtained are shown in Table 3 and Fig. 6. Table 3. Simulation Results
Fig. 5. Tissue converting line model The simulated line, depicted in Fig. 5 is a high volume automated production system which is often challenged by the succession and interaction of machine failures that lead to line downtimes and reduction of throughput. The extensive set of control logics that regulate the interaction of equipment within the line makes it difficult for job floor leaders to 490
INCOM 2012, May 23-25, 2012 Bucharest, Romania
Experimental conclusions, though validated only for the real system data analyzed in the case study, may be applicable to other systems and system configurations, but further studies to generalize the conclusions are required. Table 5. Methods ranking and contributions summary
Fig. 6. Boxplot for WIP and throughput (cs=cases) In order to compare service levels and WIP from different policies, two-sample unpooled t-tests considering unequal variances were performed with a significance level of 1%. The results are summarized in Table 4 where each cell shows the result of the hypothesis test comparing the model in the column with the model in the row. Table 4. Hypotheses tests summary
5.1 From Collaborative Production Collaborative Supply Networks 5. CONCLUSIONS Production line control under uncertainty conditions requires a highly adaptive and anticipative methodology that enables collaboration between the different components of the line to overcome failures and maintain high throughput while keeping WIP low. Traditional models for buffer management and production control lack the required characteristics. The proposed methodology, CPLC EDVHG RQ &&7¶V SULQFLSOHV, constitutes a more advanced and complete solution approach allowing both adaptability and anticipation, more suitable to face the challenges of high volume production line control under uncertainty conditions. Based on the observations made during the simulation runs and the conclusions for the case study, the methods analyzed can be classified as shown in Table 5. Statistical comparisons of the results in this case indicate that CPLC outperforms Base-stock, CONWIP, DWIP and Kanban in terms of service level with a significance level of 1%. Also, the hypothesis tests show that it is not possible to conclude that there are statistically significant differences in service level within the group of Kanban, DWIP, CONWIP and Base-stock. The statistical analysis also allows clustering models in three performance groups in terms of WIP; i.e. Base-stock and CONWIP, DWIP and Kanban, and CPLC, with a difference at significance level of 1%. As for throughput variability, it can be concluded that CPLC reduces variability, measured as the ratio between throughput standard deviation and mean. Performing the statistical comparison for comparing CPLC with the remaining methods shows that variability is reduced in every case.
Line
Control
to
Collaborative Production Line Control enables a production line to achieve higher throughput with lower variability. In other words, materials flow through the production lines faster and more evenly. The proposed methodology detects possible conflicts emerging in the near future and addresses them by utilizing information from CRP-I. This means that CPLC is capable of predicting where material flow interruptions may occur in a production line, finding alternative ways to maintain product flow, and even minimize its reduction. In Collaborative Supply Networks (CSN) material flow through the supply chain can also be controlled by using the same principles contained in CPLC. In the case of a CSN (Fig. 7) flow interruptions may appear as product stockouts in DCs and XDs, raw material shortage from suppliers, lack of production capacity in a given facility, among others. These situations are analogous to downtimes in equipment in a production line that prevent materials from flowing across them. In a CSN, CRP-I will consider alternative ways to convey products from the first supplier in tier n to the retailers in the opposite side of the supply chain, under different flow interruption situations. The optimal supply flow in each case should be defined based on the cost of moving all the required materials through the different chosen points of the supply chain in order to deliver products to the retailers. As for the early conflict detection capabilities, visibility through the entire supply chain can be embedded in the model, in order to provide decision makers with the necessary status information to be able to predict possible flow interruption in the near future. Therefore, an information sharing system enabling parties to share insights must be implemented to create an ECDT for CSN.
491
INCOM 2012, May 23-25, 2012 Bucharest, Romania
Fig. 7. Collaborative Supply Network (CSN): DC=distribution center; XD=Cross dock Implementation of a Collaborative Control Theory approach to CSN control can result in major improvement in service level and reduction in lead time and total supply cost. 6. FUTURE RESEARCH It has been proven that production line control based on CCT can produce relatively high line performance while keeping low WIP levels in high volume production lines under uncertainty conditions. The methodology introduced in this study considers only the first system state evolution options to define the new control strategy. More complex path evolution analysis models for depicting possible future system states and their associated probabilities of occurrence can be integrated into CPLC to extend its prediction accuracy, enabling it to reach better control decisions. Although CPLC was proven to be an effective buffer control policy for the case study at hand, it is still a theoretical methodology and future research should define the necessary requisites for it to be viable for implementation in plants. Another important research direction is collaborative supply networks with the extension of the CPLC methodology presented in this study to the general case of CSNs. Additional research must also validate the CPLC in CSNs and quantify the improvements in service level, reduction in lead time and total supply cost that can be achieved through a Collaborative Control Theory approach to CSN control. ACKNOWLEDGEMENT This research has been developed with partial support from the PRISM (Production, Robotics, and Integration Software for Manufacturing & Management) Center at Purdue University. REFERENCES Andijani, A. (1997). Trade-off between maximizing throughput rate and minimizing system time in kanban systems. International Journal of Operations & Production Management, Vol. 17 No. 5, pp. 429-445. Bonvik, A., Couch, C., and Gershwin, S. (1997). A comparison of production-line control mechanisms. International Journal of Production Research, Vol. 35 No. 3, pp. 789-804. Dalal, M., Groel, B., and Prieditis, A. (2003). Real-time decision making using simulation. Proceedings of the 2003 Winter Simulation Conference, pp. 1456-1464. Deleersnydhr, J., Hodgson, T., Muller, H., and O'Grady, P. (1989). Kanban controlled pull systems: An analytic
approach. Management Science Vol. 35, No. 9, pp. 1079-1091. Duenyas , I., and Patana-Anake, P. (1998). Base-stock control for single-product tandem make-to-stock systems. IIE Transactions, Vol 30 No. 1 , pp. 31-39. Duri, C., Frein, Y., and Di Mascolo, M. (2000). Comparison among three pull control policies: kanban, base stock, and generalized kanban. Annals of Operations Research, Vol. 93, pp. 41±69. Güllü, R. (1998). Base stock policies for production/inventory problems with uncertain capacity levels. European Journal of Operational Research, Vol. 105 No. 1, pp. 43-51. Lage Junior, M., and Godinho Filho, M. (2010). Variations of the Kanban system: Literature review and classification. International Journal of Production Economics, Vol. 125 No. 1, pp. 13-21. Ma, Y., and Koren, Y. (2004). Operation of manufacturing systems with work-in-process inventory and production control. CIRP Annals - Manufacturing Technology Vol. 53 No. 1, pp. 361-365. Nof, S. Y. (2003). Design of effective e-Work: review of models, tools, and emerging challenges. Production Planning & Control, Vol. 14 No.8, pp. 681-703. Nof, S. Y. (2007). Collaborative Control Theory for e-Work, e-Production, and e-Service. Annual Reviews in Control IFAC, Vol. 31 No.2, pp. 281-292. Nof, S. Y. (2009). Springer Handbook of Automation, pp. 1549. Springer, United States. Paternina-Arboleda, C., and Das, T. (2001). Intelligent dynamic control policies for serial production lines. IIE Transactions, Vol. 33 No. 1, pp. 65-77. Pettersen, J., and Segerstedt, A. (2009). Restricted work-inprocess: A study of differences between Kanban and CONWIP. International Journal of Production Economics, Vol. 118 No. 1, pp. 199±207. Savsar, M. (1996). Effects of Kanban withdrawal policies and other factors on the performance of JIT systems - A simulation study. International Journal of Production Research, Vol. 34, No. 10, pp. 2879-2899. Sendil Kumar, C., and Panneerselvam, R. (2006). Literature review of JIT-KANBAN system. International Journal of Advanced Manufacturing Technology 2007, Vol. 32 No. 3-4 , pp. 393-408. Spearman, M., Woodruff, D., and Hopp, W. (1990). CONWIP: a pull alternative to kanban. International Journal of Production Research, Vol. 28 No. 5, pp. 879894. Sugimori, Y., Kusunoki, K., Cho, F., and Uchikawa, S. (1977). Toyota production system and Kanban system: Materialization for just-in-time and respect-for-human system. International Journal of Production Research, Vol. 15 No. 6, pp. 553-564. Yadzan Shenas, N., Eshraghniaye Jahromi, A., and Modarres, M. (2009). A new approach to find an optimal solution for base stock policies. Journal of Applied Sciences, Vol. 9 No. 4, pp. 789-793. Yang, R., Subramaniam, V., and Gershwin, S. (2006). Setting real time WIP levels in production lines. Innovation in Manufacturing Systems and Technology ± MIT.
492