- Email: [email protected]
Simulation of a Hybrid Product-Driven System for Manufacturing Planning and Control*
11th IFAC Workshop on Intelligent Manufacturing Systems The International Federation of Automatic Control May 22-24, 2013. São Paulo, Brazil
WeFT1.3
Simulation of a Hybrid Product-Driven System for Manufacturing Planning and Control ? Carlos Herrera ∗ Andr´ e Thomas ∗∗ V´ıctor Vera ∗∗∗ ∗
Departamento de Ingenier´ıa Industrial, Universidad de Concepci´ on, Chile (e-mail: [email protected]). ∗∗ Centre de Recherche en Automatique de Nancy (CRAN), CNRS UMR 7039, France (e-mail: [email protected]) ∗∗∗ Departamento de Ingenier´ıa Industrial, Universidad de Concepci´ on, Chile (e-mail: [email protected]). Abstract: Manufacturing planning and control systems (MPCS) incorporate processes that consider several levels of product aggregation and different time horizons for decision making. The decisions rendered on each level do not always have similar objectives. In the context of intelligent manufacturing systems (IMS), the coordination of decisions on different levels is a fundamental problem. Extensive research on IMS, specifically regarding coordination among decision levels in product-driven control systems (PDCS), is nonexistent. Therefore, simulations of the planning and control processes are proposed to analyze the coordination of multilevel objectives. The proposed implementation simulates the coordination between tactical and operational levels. At the tactical level, production plans are obtained through a system based on advanced planning and scheduling (APS). At the operational level, a decentralized system, which is based on distributed decision rules, is implemented. The simulation considers decentralized decisions that are managed by production lots, which are modeled as holons, and based on an industrial study case. The results indicate that coordination is feasible and highlight the importance of the reactivity caused by the distributed decisions made by the active lots. The proposed simulation schema can also be used to compare conventional and holonic collaborative approaches. Keywords: Intelligent manufacturing systems, industrial production systems, manufacturing systems, mathematical programming, production control. 1. INTRODUCTION Currently, holonic manufacturing systems (HMS) are a feasible alternative for improving the flexibility and adaptability of manufacturing. These characteristics are fundamental because of the complexity and dynamism of the current systems (Valckenaers et al., 2007). In a holonic system, entities (machines, robots, AGVs or workers) are modeled as holons, which consist of a physical component and an information processing component. Holons can be individual entities or can be composed of other holons. A set of organized holons is named a holarchy. The objective is to achieve collaborative behavior among these holons, in real or simulated systems, which enables them to make decisions that reflect their environment. A considerable amount of research has been devoted to the application of this concept for scheduling decisions ? The authors gratefully acknowledge the financial support of the CPER 2007-2013. “Structuration du Pˆ ole de Comp´ etitivit´ e Fibres Grand’Est” (Competitiveness Fiber Cluster), through local (Conseil G´ en´ eral des Vosges), regional (R´ egion Lorraine), national (DRRT and FNADT) and European (FEDER) funds.
978-3-902823-33-5/13/$20.00 © 2013 IFAC
and manufacturing execution systems (MES). However, fewer studies have focused on the relationship between the operation and execution levels and the upper levels of decision making or enterprise resources planning (ERP) systems. For example, InteRRaP (Fischer, 1999) defines an agent as a set of functional layers linked by a control structure, which is based on communication. This architecture is based on agents using the social model BDI (Belief, Desires and Intentions) (Wooldridge, 2000). InteRRaP was proposed for flexible manufacturing systems (FMS). Functional layers correspond to the three basic activities that agents must perform in a FMS. Activities include coordination, problem solving and implementation of local plans. This architecture considers planning as the highest level in the decision tree. Decisions at this level are implemented offline without the potential for subsequent change. The architecture PROSA (Product-Resource-Order-Staff Architecture) (Van Brussel et al., 1998) is an architecture used for modeling and implementing a holonic manufacturing system. PROSA defines three basic types of holons: orders, products and resources. In addition, a holon staff can be defined as decision procedures or knowledge. All
69
holons in this holarchy are organized to conduct manufacturing activities. An interesting aspect of PROSA is its interpretation of a product as an active entity in the production process, which is a widespread practice in studies that address smart products (Pannequin et al., 2007). Nevertheless, the majority of the applications using PROSA as modeling framework for holonic systems concerns operational and control levels and, more specifically, in the context of modeling and implementation of MES systems. ExPlanTech is an agent-based technology for planning and production control (Pˇechouˇcek et al., 2007). This technology is based on Proplant (Marik et al., 2000), which is an architecture that was developed as a multi-agent system for project-based production systems. ExplanTech represents a generalization of Proplant for mass-production companies. The system functions by using a community of autonomous agents that represent entities or production information. A central feature of this technology is based on the premise that no centralized decision mechanism is utilized. PABADIS Promise is an architecture for production control based on a pyramid with three levels of automation (Wunsch and Bratukhin, 2007). One of its main objectives is to avoid the centralization of decisions by positioning the decision levels closer to the work-flow levels. Decision levels correspond to the ERP level (tactical level), MES level (operational level) and control level. Communication between ERP and MES is based on “web services” using ACL (agent communication language). The architecture is based on a manufacturing-order operation decomposition that is obtained from ERP. Although these approaches consider planning and control levels, they do not consider the inclusion of products as central entities in the decision-making process (except PROSA). On the contrary, PDCS (McFarlane et al., 2002; Morel et al., 2003, 2007) convert the role of products to active agents in the decision-making process, in which products can be also modeled as holons. On the other hand, decisions made at the planning level require the inclusion of medium-term horizons to prevent “myopia”. For that, the resources required for production (personnel, labor, raw materials, and machinery maintenance) should be necessarily planned in advance. Conversely, operational-level decisions are concerned with short-term horizons, which are inherently “myopic”. Finally, operational-level decisions must respond quickly and efficiently to disturbances (production blocking, machine breakdowns, and demand changes). Thus, planning and control systems should be robust, flexible and reactive with respect to short, medium, and long term decisions. In this context, Herrera (2011); Herrera et al. (2012) proposed an architecture that models a holarchy from products and sets of products at each level. This approach allows coordination among decision levels and their associated decision horizons, while focusing on the main objective of a production system (its products). Similar to other approaches in manufacturing (Tang et al., 2011), this architecture is based on the Viable System Model. The main advantage of this modeling framework is recursion, 978-3-902823-33-5/13/$20.00 © 2013 IFAC
that is, each composition level of the holarchy organization exhibits the same structure and organization on each of its levels. Recursion enables the replication of the same functions at each level, with only modifications to the objectives and decision methods. Therefore, the aim of this work is to analyze the coordination between centralized planning and decentralized control decisions in PDCS. Decentralized decisions are assumed to be performed by numerous holons. These holons detect disturbances in planning and trigger local changes that affect central planning. To accomplish this task, we developed an agent-based simulation of decisions that is based on an industrial case study. At the planning level, the goal is to preserve the stability of the plans. At the operational level, the goal is to minimize makespan degradation by satisfying buffer-stock constraints. From the viewpoint of PDCS, the objective of this approach is to demonstrate the advantages of a distributed decision and analyze its relationship with the objectives of other decision levels (tactical). The results are obtained through simulation. This approach aims to obtain quantitative comparisons and validation measures for suitable benchmarking regarding traditional techniques for production planning and control. This last point has been defined as a main objective of newer challenges proposed by the HMS community (Valckenaers et al., 2006). This paper is organized as follows: Section 2 is dedicated to all elements considered in the simulation experiment, and Section 3 presents the simulation results. Section 4 presents an analysis of the main results, and Section 5 states the conclusion and provides some research perspectives. 2. MATERIALS AND METHODS 2.1 Centralized and distributed decisions At different levels of MPCS, decisions are obtained by considering a rolling horizon. Levels are associated with different levels of an aggregation of products, such as product families, production orders, lots, and finished products or components. A major challenge is to preserve the coherence of decisions among the levels. But, in practice, when disturbances occur, the objectives for each level are not easily achieved, and disturbances may cause major planning changes. Note that frequent changes can be the source of considerable instability. In addition, these effects often cause reduced efficiency and poor productivity. Short-term changes are more frequent and can significantly reduce system performance. In this context, MPCS should provide sufficient flexibility at the operational levels and ensure consistency with the defined objectives at the upper levels. This approach considers two decision levels: tactical and operational levels. At the tactical level, the decision concerns the production quantities for every item within a product family and for each period on a planning time horizon. This problem is generally associated with the master production schedule (MPS) and is usually modeled using a lot-sizing model (Pochet and Wolsey, 2006). The objective is to minimize production costs by establishing
70
a set of parameters as marginal costs and considering the system capacity. During the first period of this planning horizon and once each quantity (by item) has been obtained, these quantities must be divided and sequenced to be incorporated in the manufacturing system. This lotstreaming problem, (Sarin and Jaiprakash, 2007) whose objectives are to reduce the total production time (Cmax ), is usually applied to manufacturing systems that contain parallel manufacturing processes. The decision at this level is comprised of a sequence of sub-lots that correspond to the weekly planning, which considers the production start time and the quantity of each product to be manufactured. One of the primary aspects of the lot-streaming problem is that it assumes constant production rates. These rates may be affected by various disturbances, such as machine blocking, machine breakdowns or accidents. Due to these occurrences, changes in the parameters of the model may affect planning efficiency. These changes reduce production capacity and increase the gap between planning and the launched production. This gap is named “system nervousness”. Our aim is to study the relationships among decision levels in the context of PDS. Then, products or sub-lots are modeled as holons with the capacity to modify their environment. Holons are assumed capable of making a single distributed decision. More specifically, a holon can decide to stop its production at a certain stage and heuristically reassign the remaining quantity. Re-assignment consists of assigning the quantity that has not been manufactured to another lot(s) (one or many), which modifies the planning. Until the new assignment is completed, a part of the sub-lot remains in an intermediate stock (buffer). The feasibility of this “decision of splitting” is dependent on the remaining production and stock capacities, and is also dependent on the existence of similar types of sub-lots (same reference) that were previously planned Once the feasibility of divisions has been determined, the sub-lot holon evaluates the variation in planning through a re-planning linear programming model. The model seeks to replace the subdivided lots and evaluate different alternatives that will minimize the increase of Cmax . These alternatives, which correspond to different sub-sets of the same reference that will increase their size, are placed in the queue sub-module. Subsequently, we describe the industrial study case and how the sub-lot holon renders the previously described decision. 2.2 Study Case The company selected for the case study is a subcontractor that manufactures turbochargers for the automotive industry. The facility produces a maximum of ten thousand products per day with hundreds of references. The plant is divided into production cells, which encompass all stages of production that are required to produce a finished product. Some production cells are dedicated to a specific customer. In this study, we consider one of these production cells. In a cell, the production process is divided into two stages. An initial set of operations are performed in the 978-3-902823-33-5/13/$20.00 © 2013 IFAC
Fig. 1. Production module first line (module A), generating semi-finished products. These products are assembled into three independent sub-assembly modules (module B). The production cell includes storage of raw materials, semi-finished (buffer), and finished products. 2.3 Distributed decision model Indexes l = 1, 2, . . . , L i ∈ Ωl j = 1, 2, . . . , J k = 1, 2, . . . , K Variables Cmax : xbijk : ST Aj
:
ST Bjk
:
: : : :
lots, sub-lots in lot l, sequence positions, cells at B.
makespan re-planned sub-lot quantity of item i in sequence position j assigned to module k of stage B, start time at stage A of sub-lot in position j. start time at module k of sub-lot in position j
Parameters xcut qi T P Ai
: : :
T P Bi
:
SAi SBi L nl = bQl /ql c
: : : :
Ωl = {1, 2, . . . , nl } PL I = i=1 |Ωl | K
: : :
xfijk yfijk tnew
: : :
quantity to be re-planned minimum sub-lot size of item i, marginal production time at A of item i, marginal production time at B of item i, setup time at A of item i, setup time at B of item i, number of items, maximum number of sub-lots in lot l set of sub-lots in lot l number of sub-lots, number of sub-modules,
fixed sub-lot quantities. fixed sequence. new start time at A for the first sub-lot in the planning after disturbance detection. (P0 ) min Cmax s.t. J X K X X
xbijk = xcut,
(1) (2)
i∈Ωw j=1 k=1
xbijk = 0, i ∈ Ωl , ∀l : l 6= w, ∀j, ∀k ST A1 = tnew . L X K XX ST Aj = ST Aj−1 + T P Al · xfi(j−1)k i∈Ωl l=1 k=1
(3) (4) (5)
+ SAl · yfi(j−1)k , ∀j : j > 1 ST Bjk ≥ ST Aj +
L X X
T P Al · yfijk , ∀j, ∀k
(6)
l=1 i∈Ωl
71
ST Bjk ≥ ST B(j−1)k L X X + T P Bl · (xfi(j−1)k + xbi(j−1)k ) (7) l=1 i∈Ωl
+ SBl · yfi(j−1)k , ∀j : j > 1, ∀k Cmax ≥ ST BJk +
L X X
T P Bl · (xfiJk + xbiJk )
l=1 i∈nl
(8)
+ SBl · yfiJk , ∀k xbijk ∈ Z+ , ST A, ST B, Cmax ≥ 0
(9)
Objective function (1) minimizes the Cmax represented by the end date of the last piece in the sequence. Constraint (2) ensures that the sum of the re-assignments (xbijk ) will be equal to the remaining quantity in the intermediate stock (xcut). Constraint (3) establishes that the reassignment can only be performed for the planned sub-lots that belong to the same lot that was previously divided. The start time of the sub-lot in position j is set to tnew by constraint (4). Constant tnew represents the new start time of the first sub-lot after disturbance detection. This sublot corresponds to the first sub-lot in the planned sequence (not yet in production). The recursive relationship in constraint (5) expresses that the start time of module A for the sub-lot in the j-th position must be equal to the start time of the previous sub-lot (sub-lot in position j − 1) plus its setup and production time, which is determined by considering a fixed sequence (xfijk and yfijk ). Constraint (6) ensures that the start time of module B will always be greater than the start time of module A plus the production time for module A (corresponding product). Constraint (7) considers that the production time for module B must be increased proportionally by the reassigned quantities. The makespan is defined by constraint (8). 2.4 Simulation and parameters Simulation considers two decision levels. The first level (tactical) is implemented using an integer programming model that defines quantities as produced by item and period in a rolling horizon. The details of this model are provided by Herrera and Thomas (2009). In the first level, the quantities are divided into sub-lots during the first period, and the sequence to be used in the manufacturing process must be defined using an integer programming model that solves the lot-streaming problem. The model is similar to P0 but the quantities and sequences are variable, which increases the execution time but are performed only at the beginning of the operation period. Once the system has been initialized with these results, the simulation begins. A platform is used at this step (Pannequin et al., 2009) to facilitate the discrete events simulation and the evaluation of different criteria. During the production period (week), variations in the production times are simulated for the different modules; disturbances are simulated as blocking and breakdowns, for example. To react to perturbations, P0 is solved to determine if a certain quantity of items is placed into stock and to determine whether this decision improves the planning 978-3-902823-33-5/13/$20.00 © 2013 IFAC
with respect to the initial situation. The decision to use this distributed decision process is dependent on the variation between the planned waiting time and the real time of a product in the queue of module B. The simulation is performed considering a horizon of one year, obtaining weekly operational results, of which some results consider the distributed decision and some results disregard the distributed decision. Stability is achieved at the tactical level using a nervousness measure proposed in (Herrera and Thomas, 2009). This measures, quantifies the variation in the planned quantities on a weekly basis. At the operational level, the obtained Cmax and work-inprocess (W IP ) are compared. Some of the parameters that were utilized are listed in Table 1. Table 1. Simulation parameters Parameters Items (m) Planning horizon (n) Simulation horizon (H) Demand (d) Production (p) Stocks (h) Backlog (b) Setup (q)
Values 12 8 52 η(1200, σ) U([5,15]) U([10,25]) U([20,40]) U([1900,2200])
3. RESULTS 3.1 Nervousness Fig. 2. Nervousness Fig. 3. Filtered nervousness Table 2. ANOVA of nervousness results Source Columns Error Total
SS 1.007 128.511 129.518
df 1 118 119
MS 1.0069 1.08907
F 18.79
Prob > F 0.3382
Figure 2 shows the results, considering nervousness, of both cases (centralized and hybrid) for a comparison. The cases represent situations in which the product is active (hybrid) and situations in which the product is inactive (centralized). The complete experiment is discussed in Herrera (2011). The centralized case considers a model that reduces the nervousness of the plan, thus, its shape in Figure 2 represents a “stable plan”. These results represent the difference between the launched production and the weekly planned production for a one-year operational horizon. The figures cover 60 periods according to a transient period of 8 weeks. Figure 3 shows the same results after a filter is applied (Savitzky-Golay). The reason for applying a filter is to distinctly capture the differences between the two decision systems. This particular filter was chosen because it preserves the characteristics of the initial distribution and the relative minimum and maximum, as well as the width of the peaks. Table 2 displays the results of a statistical hypothesis test that was employed to verify if differences exist between the series. The H0 hypothesis was described
72
as “significant differences exist between both cases with respect to the nervousness results”, and the H1 hypothesis was described as “significant differences do not exist between both cases respect to the nervousness results”. The results reveal no changes in stability for the plans in which the products are active. 3.2 Cmax
Fig. 5. Filtered Cmax Table 3. ANOVA of Cmax results SS 471729.2 29622734.4 3434463.6
df 1 118 119
MS 471729.2 25107.9
F 18.79
Prob > F 3.09027e-5
Figure 4 shows the results, considering Cmax , of both cases (centralized and hybrid) for a comparison. Figure 5 shows the results after filtering. Table 3 demonstrates that active products can affect the planning process and considerably improve the production completion time. 3.3 W IP This indicator corresponds to the average stock of all references. Its average is calculated at the end of the week. Fig. 6. WIP Fig. 7. Filtered WIP Table 4. ANOVA of WIP results Source Columns Error Total
SS 11787.8 403804.4 415592.1
A simulation of different decision levels for MPCS has been presented. The objective was to analyze the coordination of decisions at different levels using centralized and distributed methods. Local decisions represent decisions made by a “holon lot” in the context of PDCS. The results demonstrate the feasibility and efficiency of coordination between central and local decisions with different objectives using a PDS approach. The feasibility of obtaining a stable planning in the middle-term (tactical level) is assessed and a significant performance in reactivity in the short term (operational level) is achieved.
Fig. 4. Cmax
Source Columns Error Total
5. CONCLUSION AND FUTURE WORK
df 1 118 119
MS 11787.8 3422.1
F 3.44
Prob > F 0.066
Figure 6 displays the results considering the work-inprocess, of both cases (centralized and hybrid) for a comparison. Figure 7 shows the after filtering. Table 4 demonstrates that the intermediate stock level is utilized more frequently when the products are active. 4. DISCUSSION The results of Cmax are particularly interesting because they reveal a net gain without any deterioration in stability (see Figure 5). This finding indicates that it is possible to achieve “robustness” in the final results. Figure 7 demonstrates that we “must pay” in stock (which is intuitively expected) the gains in nervousness and Cmax . The Cmax efficiency is directly related to the increase and even saturation of the intermediate stock. Thus, it is possible to conclude that the proposed approach enables a better utilization of this resource. 978-3-902823-33-5/13/$20.00 © 2013 IFAC
The use of models and methods that are based on mathematical programming is justified because these models enable acceptable approximations to the problem and provide a comparison with other approaches as, for example, collaborative strategies. This stage of our work represents a starting point for further research developments. The usefulness of the proposed system should be validated in large-production environments or for situations in which the consideration of centralized decisions is not feasible at all. REFERENCES Fischer, K. (1999). Agent-based design of holonic manufacturing systems. Robotics and Autonomous Systems, 27(1-2), 3–13. Herrera, C. (2011). Cadre g´en´erique de planification logistique dans un contexte de d´ecisions centralis´ees et distribu´ees. Ph.D. thesis, Universit´e Henri Poincar´e Nancy I. Herrera, C., Belmokhtar, S., and Thomas, A. (2012). “Viable System Model approach for holonic productdriven manufacturing systems”, volume 402 of Studies in Computational Intelligence (SCI), 169–181. Springer. Herrera, C. and Thomas, A. (2009). Simulation of less Master Production Schedule nervousness model. In Proceedings of the 13th IFAC Symposium on Information Control Problems in Manufacturing, 1585–1590. Marik, V., Pˇechouˇcek, M., Stepankova, O., and Lazansky, J. (2000). Proplant: Multiagent system for production planning. Applied Artificial Intelligence, 14(7), 727–762. McFarlane, D., Sarma, S., Chirn, J., Wong, C., and Ashton, K. (2002). The intelligent product in manufacturing control. Journal of EAIA, 5464. Morel, G., Panetto, H., Zaremba, M., and Mayer, F. (2003). Manufacturing Enterprise Control and Management System Engineering: paradigms and open issues. Annual Reviews in Control, 27, 199–209. Morel, G., Valckenaers, P., Faure, J.M., Pereira, C.E., and Diedrich, C. (2007). Manufacturing plant control challenges and issues. Control Engineering Practice, 15, 1321–1331. Pannequin, R., Morel, G., and Thomas, A. (2007). Benchmarking issues for product-driven decision-making. 9th International Conference on the Modern Information Technology in the Innovation Processes of the Industrial Enterprise, MITIP’2007. Pannequin, R., Morel, G., and Thomas, A. (2009). The performance of product-driven manufacturing control:
73
An emulation-based benchmarking study. Computers in Industry, 60(3), 195–203. Pˇechouˇcek, M., Reh´ ak, M., Charv´ at, P., Vlˇcek, T., and Kol´ aˇr, M. (2007). Agent-Based Approach to MassOriented Production Planning: Case Study. IEEE Transactions on Systems, Man, and Cybernetics, Part C, 37(3), 386–395. Pochet, Y. and Wolsey, L. (2006). Production planning by mixed integer programming. Springer New York, New York. Sarin, S. and Jaiprakash, P. (2007). Flow Shop Lot Streaming. Springer, New York. Tang, D., Gu, W., Wang, L., and Zheng, K. (2011). A neuroendocrine-inspired approach for adaptive manufacturing system control. International Journal of Production Research, 49(5), 1255–1268. doi: 10.1080/00207543.2010.518734. Valckenaers, P., Brussel, H.V., Verstraete, P., Germain, B.S., and Hadeli (2007). Schedule execution in autonomic manufacturing execution systems. Journal of manufacturing systems, 26(2), 75–84. Valckenaers, P., Cavalieri, S., Germain, B., Verstraete, P., Hadeli, Bandinelli, R., Terzi, S., and Brussel, H. (2006). A benchmarking service for the manufacturing control research community. Journal of Intelligent Manufacturing, 17(6), 667–679. Van Brussel, H., Wyns, J., Valckenaers, P., Bongaerts, L., and Peeters, P. (1998). Reference architecture for holonic manufacturing systems: PROSA. Computers in industry, 37(3), 255–274. Wooldridge, M. (2000). Reasoning about rational agents. MIT press. Wunsch, D. and Bratukhin, A. (2007). Multilevel order decomposition in distributed production. In Emerging Technologies and Factory Automation, 2007. ETFA. IEEE Conference on, 872–879. IEEE.
978-3-902823-33-5/13/$20.00 © 2013 IFAC
74
WeFT1.3
Simulation of a Hybrid Product-Driven System for Manufacturing Planning and Control ? Carlos Herrera ∗ Andr´ e Thomas ∗∗ V´ıctor Vera ∗∗∗ ∗
Departamento de Ingenier´ıa Industrial, Universidad de Concepci´ on, Chile (e-mail: [email protected]). ∗∗ Centre de Recherche en Automatique de Nancy (CRAN), CNRS UMR 7039, France (e-mail: [email protected]) ∗∗∗ Departamento de Ingenier´ıa Industrial, Universidad de Concepci´ on, Chile (e-mail: [email protected]). Abstract: Manufacturing planning and control systems (MPCS) incorporate processes that consider several levels of product aggregation and different time horizons for decision making. The decisions rendered on each level do not always have similar objectives. In the context of intelligent manufacturing systems (IMS), the coordination of decisions on different levels is a fundamental problem. Extensive research on IMS, specifically regarding coordination among decision levels in product-driven control systems (PDCS), is nonexistent. Therefore, simulations of the planning and control processes are proposed to analyze the coordination of multilevel objectives. The proposed implementation simulates the coordination between tactical and operational levels. At the tactical level, production plans are obtained through a system based on advanced planning and scheduling (APS). At the operational level, a decentralized system, which is based on distributed decision rules, is implemented. The simulation considers decentralized decisions that are managed by production lots, which are modeled as holons, and based on an industrial study case. The results indicate that coordination is feasible and highlight the importance of the reactivity caused by the distributed decisions made by the active lots. The proposed simulation schema can also be used to compare conventional and holonic collaborative approaches. Keywords: Intelligent manufacturing systems, industrial production systems, manufacturing systems, mathematical programming, production control. 1. INTRODUCTION Currently, holonic manufacturing systems (HMS) are a feasible alternative for improving the flexibility and adaptability of manufacturing. These characteristics are fundamental because of the complexity and dynamism of the current systems (Valckenaers et al., 2007). In a holonic system, entities (machines, robots, AGVs or workers) are modeled as holons, which consist of a physical component and an information processing component. Holons can be individual entities or can be composed of other holons. A set of organized holons is named a holarchy. The objective is to achieve collaborative behavior among these holons, in real or simulated systems, which enables them to make decisions that reflect their environment. A considerable amount of research has been devoted to the application of this concept for scheduling decisions ? The authors gratefully acknowledge the financial support of the CPER 2007-2013. “Structuration du Pˆ ole de Comp´ etitivit´ e Fibres Grand’Est” (Competitiveness Fiber Cluster), through local (Conseil G´ en´ eral des Vosges), regional (R´ egion Lorraine), national (DRRT and FNADT) and European (FEDER) funds.
978-3-902823-33-5/13/$20.00 © 2013 IFAC
and manufacturing execution systems (MES). However, fewer studies have focused on the relationship between the operation and execution levels and the upper levels of decision making or enterprise resources planning (ERP) systems. For example, InteRRaP (Fischer, 1999) defines an agent as a set of functional layers linked by a control structure, which is based on communication. This architecture is based on agents using the social model BDI (Belief, Desires and Intentions) (Wooldridge, 2000). InteRRaP was proposed for flexible manufacturing systems (FMS). Functional layers correspond to the three basic activities that agents must perform in a FMS. Activities include coordination, problem solving and implementation of local plans. This architecture considers planning as the highest level in the decision tree. Decisions at this level are implemented offline without the potential for subsequent change. The architecture PROSA (Product-Resource-Order-Staff Architecture) (Van Brussel et al., 1998) is an architecture used for modeling and implementing a holonic manufacturing system. PROSA defines three basic types of holons: orders, products and resources. In addition, a holon staff can be defined as decision procedures or knowledge. All
69
holons in this holarchy are organized to conduct manufacturing activities. An interesting aspect of PROSA is its interpretation of a product as an active entity in the production process, which is a widespread practice in studies that address smart products (Pannequin et al., 2007). Nevertheless, the majority of the applications using PROSA as modeling framework for holonic systems concerns operational and control levels and, more specifically, in the context of modeling and implementation of MES systems. ExPlanTech is an agent-based technology for planning and production control (Pˇechouˇcek et al., 2007). This technology is based on Proplant (Marik et al., 2000), which is an architecture that was developed as a multi-agent system for project-based production systems. ExplanTech represents a generalization of Proplant for mass-production companies. The system functions by using a community of autonomous agents that represent entities or production information. A central feature of this technology is based on the premise that no centralized decision mechanism is utilized. PABADIS Promise is an architecture for production control based on a pyramid with three levels of automation (Wunsch and Bratukhin, 2007). One of its main objectives is to avoid the centralization of decisions by positioning the decision levels closer to the work-flow levels. Decision levels correspond to the ERP level (tactical level), MES level (operational level) and control level. Communication between ERP and MES is based on “web services” using ACL (agent communication language). The architecture is based on a manufacturing-order operation decomposition that is obtained from ERP. Although these approaches consider planning and control levels, they do not consider the inclusion of products as central entities in the decision-making process (except PROSA). On the contrary, PDCS (McFarlane et al., 2002; Morel et al., 2003, 2007) convert the role of products to active agents in the decision-making process, in which products can be also modeled as holons. On the other hand, decisions made at the planning level require the inclusion of medium-term horizons to prevent “myopia”. For that, the resources required for production (personnel, labor, raw materials, and machinery maintenance) should be necessarily planned in advance. Conversely, operational-level decisions are concerned with short-term horizons, which are inherently “myopic”. Finally, operational-level decisions must respond quickly and efficiently to disturbances (production blocking, machine breakdowns, and demand changes). Thus, planning and control systems should be robust, flexible and reactive with respect to short, medium, and long term decisions. In this context, Herrera (2011); Herrera et al. (2012) proposed an architecture that models a holarchy from products and sets of products at each level. This approach allows coordination among decision levels and their associated decision horizons, while focusing on the main objective of a production system (its products). Similar to other approaches in manufacturing (Tang et al., 2011), this architecture is based on the Viable System Model. The main advantage of this modeling framework is recursion, 978-3-902823-33-5/13/$20.00 © 2013 IFAC
that is, each composition level of the holarchy organization exhibits the same structure and organization on each of its levels. Recursion enables the replication of the same functions at each level, with only modifications to the objectives and decision methods. Therefore, the aim of this work is to analyze the coordination between centralized planning and decentralized control decisions in PDCS. Decentralized decisions are assumed to be performed by numerous holons. These holons detect disturbances in planning and trigger local changes that affect central planning. To accomplish this task, we developed an agent-based simulation of decisions that is based on an industrial case study. At the planning level, the goal is to preserve the stability of the plans. At the operational level, the goal is to minimize makespan degradation by satisfying buffer-stock constraints. From the viewpoint of PDCS, the objective of this approach is to demonstrate the advantages of a distributed decision and analyze its relationship with the objectives of other decision levels (tactical). The results are obtained through simulation. This approach aims to obtain quantitative comparisons and validation measures for suitable benchmarking regarding traditional techniques for production planning and control. This last point has been defined as a main objective of newer challenges proposed by the HMS community (Valckenaers et al., 2006). This paper is organized as follows: Section 2 is dedicated to all elements considered in the simulation experiment, and Section 3 presents the simulation results. Section 4 presents an analysis of the main results, and Section 5 states the conclusion and provides some research perspectives. 2. MATERIALS AND METHODS 2.1 Centralized and distributed decisions At different levels of MPCS, decisions are obtained by considering a rolling horizon. Levels are associated with different levels of an aggregation of products, such as product families, production orders, lots, and finished products or components. A major challenge is to preserve the coherence of decisions among the levels. But, in practice, when disturbances occur, the objectives for each level are not easily achieved, and disturbances may cause major planning changes. Note that frequent changes can be the source of considerable instability. In addition, these effects often cause reduced efficiency and poor productivity. Short-term changes are more frequent and can significantly reduce system performance. In this context, MPCS should provide sufficient flexibility at the operational levels and ensure consistency with the defined objectives at the upper levels. This approach considers two decision levels: tactical and operational levels. At the tactical level, the decision concerns the production quantities for every item within a product family and for each period on a planning time horizon. This problem is generally associated with the master production schedule (MPS) and is usually modeled using a lot-sizing model (Pochet and Wolsey, 2006). The objective is to minimize production costs by establishing
70
a set of parameters as marginal costs and considering the system capacity. During the first period of this planning horizon and once each quantity (by item) has been obtained, these quantities must be divided and sequenced to be incorporated in the manufacturing system. This lotstreaming problem, (Sarin and Jaiprakash, 2007) whose objectives are to reduce the total production time (Cmax ), is usually applied to manufacturing systems that contain parallel manufacturing processes. The decision at this level is comprised of a sequence of sub-lots that correspond to the weekly planning, which considers the production start time and the quantity of each product to be manufactured. One of the primary aspects of the lot-streaming problem is that it assumes constant production rates. These rates may be affected by various disturbances, such as machine blocking, machine breakdowns or accidents. Due to these occurrences, changes in the parameters of the model may affect planning efficiency. These changes reduce production capacity and increase the gap between planning and the launched production. This gap is named “system nervousness”. Our aim is to study the relationships among decision levels in the context of PDS. Then, products or sub-lots are modeled as holons with the capacity to modify their environment. Holons are assumed capable of making a single distributed decision. More specifically, a holon can decide to stop its production at a certain stage and heuristically reassign the remaining quantity. Re-assignment consists of assigning the quantity that has not been manufactured to another lot(s) (one or many), which modifies the planning. Until the new assignment is completed, a part of the sub-lot remains in an intermediate stock (buffer). The feasibility of this “decision of splitting” is dependent on the remaining production and stock capacities, and is also dependent on the existence of similar types of sub-lots (same reference) that were previously planned Once the feasibility of divisions has been determined, the sub-lot holon evaluates the variation in planning through a re-planning linear programming model. The model seeks to replace the subdivided lots and evaluate different alternatives that will minimize the increase of Cmax . These alternatives, which correspond to different sub-sets of the same reference that will increase their size, are placed in the queue sub-module. Subsequently, we describe the industrial study case and how the sub-lot holon renders the previously described decision. 2.2 Study Case The company selected for the case study is a subcontractor that manufactures turbochargers for the automotive industry. The facility produces a maximum of ten thousand products per day with hundreds of references. The plant is divided into production cells, which encompass all stages of production that are required to produce a finished product. Some production cells are dedicated to a specific customer. In this study, we consider one of these production cells. In a cell, the production process is divided into two stages. An initial set of operations are performed in the 978-3-902823-33-5/13/$20.00 © 2013 IFAC
Fig. 1. Production module first line (module A), generating semi-finished products. These products are assembled into three independent sub-assembly modules (module B). The production cell includes storage of raw materials, semi-finished (buffer), and finished products. 2.3 Distributed decision model Indexes l = 1, 2, . . . , L i ∈ Ωl j = 1, 2, . . . , J k = 1, 2, . . . , K Variables Cmax : xbijk : ST Aj
:
ST Bjk
:
: : : :
lots, sub-lots in lot l, sequence positions, cells at B.
makespan re-planned sub-lot quantity of item i in sequence position j assigned to module k of stage B, start time at stage A of sub-lot in position j. start time at module k of sub-lot in position j
Parameters xcut qi T P Ai
: : :
T P Bi
:
SAi SBi L nl = bQl /ql c
: : : :
Ωl = {1, 2, . . . , nl } PL I = i=1 |Ωl | K
: : :
xfijk yfijk tnew
: : :
quantity to be re-planned minimum sub-lot size of item i, marginal production time at A of item i, marginal production time at B of item i, setup time at A of item i, setup time at B of item i, number of items, maximum number of sub-lots in lot l set of sub-lots in lot l number of sub-lots, number of sub-modules,
fixed sub-lot quantities. fixed sequence. new start time at A for the first sub-lot in the planning after disturbance detection. (P0 ) min Cmax s.t. J X K X X
xbijk = xcut,
(1) (2)
i∈Ωw j=1 k=1
xbijk = 0, i ∈ Ωl , ∀l : l 6= w, ∀j, ∀k ST A1 = tnew . L X K XX ST Aj = ST Aj−1 + T P Al · xfi(j−1)k i∈Ωl l=1 k=1
(3) (4) (5)
+ SAl · yfi(j−1)k , ∀j : j > 1 ST Bjk ≥ ST Aj +
L X X
T P Al · yfijk , ∀j, ∀k
(6)
l=1 i∈Ωl
71
ST Bjk ≥ ST B(j−1)k L X X + T P Bl · (xfi(j−1)k + xbi(j−1)k ) (7) l=1 i∈Ωl
+ SBl · yfi(j−1)k , ∀j : j > 1, ∀k Cmax ≥ ST BJk +
L X X
T P Bl · (xfiJk + xbiJk )
l=1 i∈nl
(8)
+ SBl · yfiJk , ∀k xbijk ∈ Z+ , ST A, ST B, Cmax ≥ 0
(9)
Objective function (1) minimizes the Cmax represented by the end date of the last piece in the sequence. Constraint (2) ensures that the sum of the re-assignments (xbijk ) will be equal to the remaining quantity in the intermediate stock (xcut). Constraint (3) establishes that the reassignment can only be performed for the planned sub-lots that belong to the same lot that was previously divided. The start time of the sub-lot in position j is set to tnew by constraint (4). Constant tnew represents the new start time of the first sub-lot after disturbance detection. This sublot corresponds to the first sub-lot in the planned sequence (not yet in production). The recursive relationship in constraint (5) expresses that the start time of module A for the sub-lot in the j-th position must be equal to the start time of the previous sub-lot (sub-lot in position j − 1) plus its setup and production time, which is determined by considering a fixed sequence (xfijk and yfijk ). Constraint (6) ensures that the start time of module B will always be greater than the start time of module A plus the production time for module A (corresponding product). Constraint (7) considers that the production time for module B must be increased proportionally by the reassigned quantities. The makespan is defined by constraint (8). 2.4 Simulation and parameters Simulation considers two decision levels. The first level (tactical) is implemented using an integer programming model that defines quantities as produced by item and period in a rolling horizon. The details of this model are provided by Herrera and Thomas (2009). In the first level, the quantities are divided into sub-lots during the first period, and the sequence to be used in the manufacturing process must be defined using an integer programming model that solves the lot-streaming problem. The model is similar to P0 but the quantities and sequences are variable, which increases the execution time but are performed only at the beginning of the operation period. Once the system has been initialized with these results, the simulation begins. A platform is used at this step (Pannequin et al., 2009) to facilitate the discrete events simulation and the evaluation of different criteria. During the production period (week), variations in the production times are simulated for the different modules; disturbances are simulated as blocking and breakdowns, for example. To react to perturbations, P0 is solved to determine if a certain quantity of items is placed into stock and to determine whether this decision improves the planning 978-3-902823-33-5/13/$20.00 © 2013 IFAC
with respect to the initial situation. The decision to use this distributed decision process is dependent on the variation between the planned waiting time and the real time of a product in the queue of module B. The simulation is performed considering a horizon of one year, obtaining weekly operational results, of which some results consider the distributed decision and some results disregard the distributed decision. Stability is achieved at the tactical level using a nervousness measure proposed in (Herrera and Thomas, 2009). This measures, quantifies the variation in the planned quantities on a weekly basis. At the operational level, the obtained Cmax and work-inprocess (W IP ) are compared. Some of the parameters that were utilized are listed in Table 1. Table 1. Simulation parameters Parameters Items (m) Planning horizon (n) Simulation horizon (H) Demand (d) Production (p) Stocks (h) Backlog (b) Setup (q)
Values 12 8 52 η(1200, σ) U([5,15]) U([10,25]) U([20,40]) U([1900,2200])
3. RESULTS 3.1 Nervousness Fig. 2. Nervousness Fig. 3. Filtered nervousness Table 2. ANOVA of nervousness results Source Columns Error Total
SS 1.007 128.511 129.518
df 1 118 119
MS 1.0069 1.08907
F 18.79
Prob > F 0.3382
Figure 2 shows the results, considering nervousness, of both cases (centralized and hybrid) for a comparison. The cases represent situations in which the product is active (hybrid) and situations in which the product is inactive (centralized). The complete experiment is discussed in Herrera (2011). The centralized case considers a model that reduces the nervousness of the plan, thus, its shape in Figure 2 represents a “stable plan”. These results represent the difference between the launched production and the weekly planned production for a one-year operational horizon. The figures cover 60 periods according to a transient period of 8 weeks. Figure 3 shows the same results after a filter is applied (Savitzky-Golay). The reason for applying a filter is to distinctly capture the differences between the two decision systems. This particular filter was chosen because it preserves the characteristics of the initial distribution and the relative minimum and maximum, as well as the width of the peaks. Table 2 displays the results of a statistical hypothesis test that was employed to verify if differences exist between the series. The H0 hypothesis was described
72
as “significant differences exist between both cases with respect to the nervousness results”, and the H1 hypothesis was described as “significant differences do not exist between both cases respect to the nervousness results”. The results reveal no changes in stability for the plans in which the products are active. 3.2 Cmax
Fig. 5. Filtered Cmax Table 3. ANOVA of Cmax results SS 471729.2 29622734.4 3434463.6
df 1 118 119
MS 471729.2 25107.9
F 18.79
Prob > F 3.09027e-5
Figure 4 shows the results, considering Cmax , of both cases (centralized and hybrid) for a comparison. Figure 5 shows the results after filtering. Table 3 demonstrates that active products can affect the planning process and considerably improve the production completion time. 3.3 W IP This indicator corresponds to the average stock of all references. Its average is calculated at the end of the week. Fig. 6. WIP Fig. 7. Filtered WIP Table 4. ANOVA of WIP results Source Columns Error Total
SS 11787.8 403804.4 415592.1
A simulation of different decision levels for MPCS has been presented. The objective was to analyze the coordination of decisions at different levels using centralized and distributed methods. Local decisions represent decisions made by a “holon lot” in the context of PDCS. The results demonstrate the feasibility and efficiency of coordination between central and local decisions with different objectives using a PDS approach. The feasibility of obtaining a stable planning in the middle-term (tactical level) is assessed and a significant performance in reactivity in the short term (operational level) is achieved.
Fig. 4. Cmax
Source Columns Error Total
5. CONCLUSION AND FUTURE WORK
df 1 118 119
MS 11787.8 3422.1
F 3.44
Prob > F 0.066
Figure 6 displays the results considering the work-inprocess, of both cases (centralized and hybrid) for a comparison. Figure 7 shows the after filtering. Table 4 demonstrates that the intermediate stock level is utilized more frequently when the products are active. 4. DISCUSSION The results of Cmax are particularly interesting because they reveal a net gain without any deterioration in stability (see Figure 5). This finding indicates that it is possible to achieve “robustness” in the final results. Figure 7 demonstrates that we “must pay” in stock (which is intuitively expected) the gains in nervousness and Cmax . The Cmax efficiency is directly related to the increase and even saturation of the intermediate stock. Thus, it is possible to conclude that the proposed approach enables a better utilization of this resource. 978-3-902823-33-5/13/$20.00 © 2013 IFAC
The use of models and methods that are based on mathematical programming is justified because these models enable acceptable approximations to the problem and provide a comparison with other approaches as, for example, collaborative strategies. This stage of our work represents a starting point for further research developments. The usefulness of the proposed system should be validated in large-production environments or for situations in which the consideration of centralized decisions is not feasible at all. REFERENCES Fischer, K. (1999). Agent-based design of holonic manufacturing systems. Robotics and Autonomous Systems, 27(1-2), 3–13. Herrera, C. (2011). Cadre g´en´erique de planification logistique dans un contexte de d´ecisions centralis´ees et distribu´ees. Ph.D. thesis, Universit´e Henri Poincar´e Nancy I. Herrera, C., Belmokhtar, S., and Thomas, A. (2012). “Viable System Model approach for holonic productdriven manufacturing systems”, volume 402 of Studies in Computational Intelligence (SCI), 169–181. Springer. Herrera, C. and Thomas, A. (2009). Simulation of less Master Production Schedule nervousness model. In Proceedings of the 13th IFAC Symposium on Information Control Problems in Manufacturing, 1585–1590. Marik, V., Pˇechouˇcek, M., Stepankova, O., and Lazansky, J. (2000). Proplant: Multiagent system for production planning. Applied Artificial Intelligence, 14(7), 727–762. McFarlane, D., Sarma, S., Chirn, J., Wong, C., and Ashton, K. (2002). The intelligent product in manufacturing control. Journal of EAIA, 5464. Morel, G., Panetto, H., Zaremba, M., and Mayer, F. (2003). Manufacturing Enterprise Control and Management System Engineering: paradigms and open issues. Annual Reviews in Control, 27, 199–209. Morel, G., Valckenaers, P., Faure, J.M., Pereira, C.E., and Diedrich, C. (2007). Manufacturing plant control challenges and issues. Control Engineering Practice, 15, 1321–1331. Pannequin, R., Morel, G., and Thomas, A. (2007). Benchmarking issues for product-driven decision-making. 9th International Conference on the Modern Information Technology in the Innovation Processes of the Industrial Enterprise, MITIP’2007. Pannequin, R., Morel, G., and Thomas, A. (2009). The performance of product-driven manufacturing control:
73
An emulation-based benchmarking study. Computers in Industry, 60(3), 195–203. Pˇechouˇcek, M., Reh´ ak, M., Charv´ at, P., Vlˇcek, T., and Kol´ aˇr, M. (2007). Agent-Based Approach to MassOriented Production Planning: Case Study. IEEE Transactions on Systems, Man, and Cybernetics, Part C, 37(3), 386–395. Pochet, Y. and Wolsey, L. (2006). Production planning by mixed integer programming. Springer New York, New York. Sarin, S. and Jaiprakash, P. (2007). Flow Shop Lot Streaming. Springer, New York. Tang, D., Gu, W., Wang, L., and Zheng, K. (2011). A neuroendocrine-inspired approach for adaptive manufacturing system control. International Journal of Production Research, 49(5), 1255–1268. doi: 10.1080/00207543.2010.518734. Valckenaers, P., Brussel, H.V., Verstraete, P., Germain, B.S., and Hadeli (2007). Schedule execution in autonomic manufacturing execution systems. Journal of manufacturing systems, 26(2), 75–84. Valckenaers, P., Cavalieri, S., Germain, B., Verstraete, P., Hadeli, Bandinelli, R., Terzi, S., and Brussel, H. (2006). A benchmarking service for the manufacturing control research community. Journal of Intelligent Manufacturing, 17(6), 667–679. Van Brussel, H., Wyns, J., Valckenaers, P., Bongaerts, L., and Peeters, P. (1998). Reference architecture for holonic manufacturing systems: PROSA. Computers in industry, 37(3), 255–274. Wooldridge, M. (2000). Reasoning about rational agents. MIT press. Wunsch, D. and Bratukhin, A. (2007). Multilevel order decomposition in distributed production. In Emerging Technologies and Factory Automation, 2007. ETFA. IEEE Conference on, 872–879. IEEE.
978-3-902823-33-5/13/$20.00 © 2013 IFAC
74