Artificial Intelligence in Engineering 12 (1998)49-61 0 1997Elsevier Science Limited PII:
ELSEVIER
SO954-1810(96)00052-O
Printed in Great Britain. All rights reserved 0954-1810/98/$19.00
Hybrid production control approach for JIT scheduling D. Trentesaux,a* C. Tahon’ & P. Lad& “Equipe Gekie Kndustriel et Logiciel, Laboratoire dilutomatique et de Mtkanique Industrielles et Humaines, Universith de Valenciennes et du Hainaut-CambrPsis, Le Mont Houy, BP 311, Vaienciennes cedex, France bIPI, Avenue Felix Viallet, 38000 Grenoble, France
(Received 27 July 1995; accepted 2 December 1996)
A distributed production activity control system has been developed in the LAMIH laboratory to solve several problems inherent in the inflexibility of classic centralized production activity control systems. The distributed approach implies local decision making and real time control for task allocations. This forbids correct forecast on critical information such as time cycle for each manufacturing order, meaning a lack of viability for industrial application. The aim of our work is to provide a structure which takes advantage of both extrema and which is able to support just-in-time concepts: the hybrid production activity control structure. The hybrid approach uses the notion of bottleneck and non-bottleneck resources and is based on the distributed control structure developed in the laboratory. The bottleneck is the only resource scheduled. It forces the synchronicity of the whole production structure because of its characteristics. This paper evaluates the relative performances of the three control structures (centralized, distributed and hybrid) through theoretical discussions and practical examples from a simplified industrial case study. 0 1997 Elsevier Science Limited. Key words: production activity control structure, centralized control, distributed control, hybrid control, just-in-time scheduling, scheduling effectiveness.
Tchako et al3 have studied the structures solving this important drawback by providing a decentralization of the decision abilities (e.g. egalitarian structures, decentralized but co-ordinated structures). The conclusion of this study is that the proposed structures still do not provide sufficient solutions to the production scheduling, processing and control issues. Some major questions still arise: How can supervisors be effectively eliminated? What is the role of human operators in these systems? How can fault-tolerance be achieved? etc. Tchako et aL3 have answered such questions by providing a distributed structure based on the egalitarian principle where no theoretical scheduling is required before the effective production process. In fact, the distributed structure dynamically realizes the scheduling. The main drawbacks have been studied in Trentesaux and Tahon.4 All of them inherit from the dynamic aspect of the scheduling: no information about next task allocations is really available. The conclusion is that such distributed control may lead to unsatisfactory results or problems about industrial viability. The scope of this paper is to propose a new kind of production control for FMS able to take into account the advantages of both distributed and centralized
1 INTRODUCTION
A scheduling can roughly be defined as a subset of the Cartesian What*When*Where product.’ An allocated task (operation or job to perform) is an element of this scheduling. A manufacturing order is a set of successive tasks in relation with the customer orders and the routing operation list. The production activity control (PAC) level is responsible for the realization of a scheduling according to a set of constraints (resource, product and flow constraints). If the PAC structure is centralized, the input is a theoretical scheduling including slack times worked out by the upper level of the production management system (production planning). In that case, the aim of the PAC level is to suit as far as possible this scheduling to the production process, so as to avoid to question it, which would lead to a re-scheduling.’ If a re-calculation is required, a part of the production process is halted and waits for a new scheduling. This idle time may alter considerably the global efficiency of the production process. *To whom correspondence
should be addressed. 49
50
D. Trentesaux, C. Tahon, P. Ladet
structures to provide efficient and JIT schedulings. We show the viability of this new PAC structure through simplified industrial case studies from an existing FMS.
Here, C can not be explicit in a generic way. We can only define its domain as follows: “,. m.times C : Rm x P” x %??$
+ 10; l]
((r1,r2,.-.,rm),(P~1~2,...,~n),S(a~,p~),...,s(a,~,p,~))~
2 SCHEDULING In this part, we first formalize a scheduling in a generic way. A basic definition will be specified for each different PAC structure. Let R be the set of resources: R = {ri] i E [l,m]}, m is the number of resources. Let P be the set of jobs to process (tasks): p=
(&lj
E Il741, n is the number of jobs.
Let M be the set of manufacturing M={mklkE
z
[lyz])y
The n-uple (r, -p,S) complies with the constraints C(r,p,~) = 1.
Dejinition The matrix 2 is a satisfying scheduling if C( 1, p, 2) = 1. Let S* be the set of satisfying schedulings: S* = {S/S -- application A C(r,p,s)
orders:
is the number of manufacturing
orders. Let B, E and D be respectively the starting date for processing the manufacturing order mk E M, the finishing time for processing the manufacturing order mk E &i and the due date of the manufacturing order mk E M: B:M+l%
m= c
mk I-+B(mk) = b(k)
’ e(k) - d(k)
$ provides a more efficient JIT scheduling than s2 (noted 2 1 D 22) when:
D:M+&? mk ++D(mk) = d(k) These functions will be used for the relative evaluation of two schedulings. The task pj E P can be processed on resource ri E R. A function T gives the processing time of ri on pj: T:RxP+W
d(i, j)
with dji, j) > 0 if ri can perform the task Pi, else 0. Let R, and P, be subsets of R and P, n, = card(P,) and m, = card(R,). The elements of P, are jobs to be scheduled, the elements of R, are resources where jobs are scheduled. Let S be the scheduling function: S:R,xP,dW G 4&j)
and 2 the scheduling matrix associated with s(i, j). s(i, j) equals the starting date for ri to perform the task else S(i, j) = 0. Let C be the function that combines the three types of hard constraints:
pj,
-
Z
f-
mk t-i E(mk) = e(k)
Cri,Pj) ++S(ri,Pj)
= 1)
Dejinition be two satisfying schedulings. Let Let @l,tj2) E S*’ (ml, cl) (resp. (m2,02)) be the couple (mean tardiness, standard deviation of tardiness) of 2 f (resp. s2 ) defined as follows:
k=l
(ri,Pj) ++T(r,,Pj) =
if
disjunctive constraints, capacity constraints, constraints for production process potential.
(1) Iml I < I m2 (2) OI < g2.
I
and
If (1) or (2) is not true, we can not conclude. We note: 1Sl
= g2.
(S *, D, x) defines a complete pre-order. Of course, one can design his own definition of a JIT scheduling. This is a multicriteria problematic and the reader could find more details in Roy and Bouyssou5 or Zeleny.6 The previous definition is efficient enough to provide good relative evaluations. Of course, in the case S 1 zz S2, other criteria must be taken into account to &aluFte more precisely the two schedulings but this is beyond the scope of this paper. These definitions will help us to focus on the differences among the three mentioned PAC structures (centralized, distributed and hybrid). 3 PRODUCTION STRUCTURES
ACTIVITY CONTROL
A PAC structure aims at managing the production structure so as to realize a scheduling as best as possible. For each approach (centralized, distributed and hybrid), we define the responsibilities of each control.
51
JIT scheduling
The cooperation is defined by a cooperation degree, which ranges from fully cooperative to antagonistic agents. Communication between agents depends on the selected protocol, that is the set of rules that specifies the way to synthesize messages to make them significant and correct. Each agent controls a set of production resources (objects as a set of production tools such as mill, lathe, robot, etc.). Thus, tasks are operations to be performed by these production tools. Two kinds of cooperation level can be defined: horizontal (between agents who cooperate) and vertical (between each agent and the human operator associated with this agent as a supervisor). The human operator has the responsibility for task resolution and allocation (vertical cooperation through a decision support system: DSS). The task allocation consists in allocating responsibility for process on products. The integrated management system (IMS) represents an agent responsible for operation processes (objects are production facilities such as mill or lathe) and operation allocations.3>7 It integrates an operational-level decision support system for the task allocation.12 The DSS model for task allocation is based on Sprague’s concept for a DSS defined as a set of three sub-systems (data, modelling and dialog systems). l3 An IMS includes:
3.1 Centralized PAC structure Tchako7 sums up the main characteristics of a centralized PAC structure: each resource needs to be linked to a single manager to hold a dialogue among them (no direct cooperation is allowed). The decisions come from this upper level, and resources only return information on the consequences of these decisions. To calculate an optimum scheduling, a function F is first defined to create an ordered set (S’, +): F:S*--+B &+ F(g) q%,s2) ---
with F(SI)*-(SZ)@$ =
E s*z,
=
cc
>S2
e.g. F should be the function that minimizes the average lead time of the set of the manufacturing orders, or a heuristic if the problem is too large. Let s” be defined by the relation: t/z E S” - {g},F(r) F.
2 F(g),
g is the maximum by
The PAC structure is responsible for the realization of the planned scheduling S”. If the production process is altered enough (malfunction of resource, stockout, etc.), some constraints may no longer be complied (C( 1, p, -- S) = O), then the production process is stopped and waits for a new planned scheduling based on the new state of the production system. Thus, the centralized PAC structure is responsible for suiting as much as possible the planned scheduling s”.
-
3.2 Distributed control of a production structure
-
Basic concepts of a distributed PAC structure were first developed by Duffie et aL8 and Bakker.’ Tchako and Tahon’O extended and applied this basic structure to the managing of a packaging line. We sum up the main characteristics. The distributed production activity control structure (cf. Fig. 1) stems from the agent concepts: an agent” is defined as an abstract or a physical entity able to act on itself and on its environment and to communicate with other agents. It aims at performing a set of tusks, parts of a global problem. To execute these tasks, the agent may use a set of objects. Control defines the cooperation between agents, the group organization, and its evolution.
-
-
-
A decision system: it is based on a representation of the different agents and supports the algorithms for decision making. This sub-system supervises its queue of jobs. It participates in cooperation with other agents to attempt to process tasks too. A control system: it is responsible for the command orders for the handling automation (e.g. command orders for a conveyor, or a control signal to motor). A communication system: it is responsible for the information exchanges through a local area network. An information system: it supports the required information for the other sub-systems (local database). An interface system: it insures the dialogue with the human operator, and the interactions between the different sub-systems.
Figure 2 presents the integration of the dynamic task
Operator level Ve*$Zperat
ion
Agentlevel Reduction resourcelevel
I
I
1
Fig. 1. The distributed production structure.
52
D. Trentesaux, C. Tahon, P. Ladet
Fig. 2. The production management structure.
allocation in the global plan of production. Task and allocations are dynamically realized. A global production planning is first created regarding the whole set of constraints of the manufacturing orders (workshop capacity smoothing, etc.). This planning does not describe the sequence of jobs on resources, but only provides to the production system a set of feasible manufacturing orders according to major constraints such as: lead times, average workload, etc. The scheduling function is incorporated in the real time production control part. Hence, it is established dynamically through the task allocation process. The distributed PAC structure is responsible for the work process without any planned scheduling (that is, if S’ is the set of potential planned schedulings, S’ = 0). A set of manufacturing orders established by the production planning level is sent to the distributed PAC structure that manages this whole set as a sum of subtasks (operations, parts of a manufacturing order) to be allocated and executed. The vertical control is managed by a specific operational-level DSS responsible for task allocation support and queue management. The horizontal control supports fully cooperative data exchanges and cooperation through negotiation. To be as satisfactory as possible, the operation allocation must be performed by a set of cooperative IMS that has to support global (static) and local (dynamic) constraints. The basic principle of the communication protocol (negotiation paradigm) is the following: when an operation of a particular manufacturing order has been performed, the IMS sends a request about the next operation to process, according to the routing list of this manufacturing order. Each of the IMS able to perform this operation returns an acceptance. The requesting IMS selects one of the proposed IMS through cooperation with human-operator, dialogue and decision making (DSS) and sends to this IMS a reservation and a release to the others. A discharge from the selected IMS concludes the protocol. The complete protocol has been
modeled using coloured and temporized Petri nets and is presented in Trentesaux and Tahon.2 A particular IMS (‘InOut IMS’) is added to the agent structure. Once a manufacturing order is provided to the FMS, it is first managed by this IMS which has the responsibility for introducing it in the production processing through the previous negotiation protocol. Thus the InOut IMS dispatches the manufacturing orders as and when they are provided to the FMS. Once a manufacturing order has been completed, it is automatically sent by the IMS to the InOut IMS for final storage. A study has been done on the integration of multicriteria algorithms into the decision system of an IMS.14 The conclusion of this study is that multicriteria algorithms suit well the problem of task allocation and queue management in a discrete production environment (the pertinence of the set of criteria is a key for the pertinence of the selection). A complete multicriteria DSS has been developed to help the human operator for decision making. Three main cooperation levels have been defined (manual, semi-automatic, automatic) to provide help when the operator requires some. Thus, according to particular wishes and information, he may force himself a task allocation, indeed, the databases of each IMS will never own all the data required. Hence, the responsibility level of the human operator is highly increased: his role has been moved from production process to production control. 3.3 Hybrid control The hybrid structure is based on the distributed approach previously presented, but a resource is scheduled (not all of them, as the centralized approach does). This resource is called the bottleneck of the production system and is dynamically detected as and when new manufacturing orders are to be processed. Thus, to present the hybrid control, it is necessary to introduce the concept of bottleneck. This notion has been first exploited by Goldratt and Cox15 to develop the OPT method. Marris16 has pointed out some problems and forgetting that gave the birth to strong critical and proposes a more scientific approach to the industrial aspects of bottlenecks developing the notions of unbalanced production systems and dual views to achieve a JIT scheduling. According to Marris, a bottleneck is a resource in which average capacity is equal to or below the needs. Thus, the bottleneck can be detected after a certain period of tries (long term study), where mean workloads of each resource are estimated. Each resource which rate is below the needs is considered as a bottleneck. The strategy of the managers is to own only one bottleneck. It is easy to synchronize production if one bottleneck temporizes the production process. In that case, synchronicity is a way to realize just-in-time production. The bottleneck resource is first scheduled, remaining resources are then roughly scheduled according to this planning.
53
JIT scheduling If no bottleneck is detected, the managers must create one to synchronize production. If more than one are detected, the managers should modify their characteristics to turn all of them but one into non-bottleneck resources. This raises some questions: how to detect the potential bottleneck(s)? How to own only one bottleneck? How to know if another resource should be the bottleneck? etc. These questions deal with the subject of the best bottleneck (production design), which is not the scope of this paper. The reader should find complete answers in Marrn? where a methodology compatible with our approach is presented. Our definition of the bottleneck is quite different from Marris’ one: because of the dynamic control of the production, it is useful to integrate the notion of dynamic workload of the production system instead of the Marris’ definition for the workload. Let R be the function that stipulates on the state of job pj on resource ri: R:RxP-+W
1 if pj has been processed on ri or will never be processed on ri else the realized percentage time for ri to perform pi if pi is being processed by ri else 0 if pj can be processed by ri Definition The dynamic remaining workload Wi of a resource ri E R is defined as follows: wi = CyZl d(i, j) * (1 - w(i, j)) and is calculated each time a new set of jobs is allocated to the production structure. Dejinition A dynamic bottleneck i is a resource ri which workload is the maximum workload for the set of resources. The bottleneck is calculated each time a new set of manufacturing orders is allocated to the production structure, according to the dynamic workload estimations. The bottleneck is the only resource scheduled: R, = {i}, m, = card(&) = 1. The elements of P, are jobs to be scheduled that will be processed on i. The function S is simplified: S:(k)xP,+9 and S the vector (ilPj) I-b SCi,Pj) 3 4j)
associated with s(j). S is the vector of the scheduled jobs on resource i. The global control structure is based upon the distributed production activity control structure: each agent
(including the b o ttl eneck) manages its own set of tasks as described in the previous part. The bottleneck i is the only one agent that has to process its jobs at scheduled dates. The algorithm for a JIT scheduling for the resource t is the following: If newmanufacturing ordertobeprocessed Begin Let Wi be resource workloads Findbottlenecki For each job Pj of the manufacturing order mk E M do Begin If pj will have to be realized on i Let reqj =requirement date of Pj theoreticalremainingproductionlead mk after being processed by i End Let sch = array of recordcdate, count) Order dates reQj by crescent values in sch(cpt) .date=reqj and sch(cpt) .count=j For cpt=l to n, = card(P,) Begin Schedule i for job psch(cpt),countat date sch(cpt).date - Wy~sCh(Cpt).Co,t). If it is not possible Begin schedule r as soon as possible after date End End End
of
this
The previous algorithm estimates first f according to the remaining tasks to process. Then it schedules the bottleneck i by the use of a priority criterion expressed in terms of minimized production lead time from the bottleneck to requirement date. The workload of each resource is estimated each time a new set of manufacturing orders is to be processed. If the bottleneck has changed, then the previous bottleneck is to cancel its scheduling and the new bottleneck is scheduled according to the new state of the production process. It is obvious that such scheduling can be improved through: -
-
The improvement of the workload evaluation. One can modify the workload definition by adding new modules. The improvement of the scheduling part of the algorithm. One can optimize the scheduled dates by use of more complicated heuristics or artificial intelligence tools and provide like this more efficient schedulings, not necessary JIT schedulings, which is the aim of this present study.
The InOut IMS is responsible for the detection and the scheduling of the dynamic bottleneck. Through a set of exchanged messages, it informs selected IMS on the
54
D. Trentesaux, C. Tahon, P. Ladet
scheduling part. Some messages are for example: InformNewBottleNeck, SetBottleNeckScheduling, CancelBottleNeckScheduling. Hence, the InOut IMS holds the management of the scheduling part as a supervisor besides its responsibilities about the input of the manufacturing orders and the output of finished goods (cf. distributed control). This method can easily be extended to the case where the bottleneck belongs to a family of resources. In that case, the whole family is scheduled and the dynamic workload must be reduced in relation to the number of resources belonging to this family.
It is shown through examples and discussion how hybrid control takes advantage of both structures to provide an efficient control. 5 ABILITY OF A PAC STRUCTURE TO PROVIDE A JIT SCHEDULING This part deals with the way we can evaluate the effectiveness of different PAC systems through the realization of their schedulings. Effectiveness can be evaluated through the resistance analysis on planned and nonplanned disturbances (regulation) and on the static evaluation (normal operating conditions): -
4 INDUSTRIAL VIABILITY The three PAC architectures have been specified and characterized. We can now focus on the industrial viability of each approach to justify in another way the hybrid PAC architecture. 4.1 Centralized control The industrial viability of such a kind of control is very high since the scheduling entirely explicits the needs. Since it is calculated before the effective work, it is possible to anticipate the needs of raw material and to forecast the finishing times. A classical examples is the MRP II planning control system for requirements linked to a centralized scheduler. 4.2 Distributed control The distributed control manages the set of manufacturing tasks dynamically without any forecast. Hence it is impossible to detect the date of raw material needs, date of tool requirements, of human resources needs, date for customer deliveries, etc. Thus, the industrial viability is very low since we have no answer to vital questions such as: ‘when will this manufacturing order set be completed?’ or ‘when will we need human resources?’ or even ‘when do we have to order raw materials or tools? etc. 4.3 Hybrid control Hybrid control takes advantage from the centralized approach, which roughly allows us to anticipate production and requirements: the algorithm presented previously can estimate from the scheduling of the bottleneck, the needs of raw material and tools for example. That is difficult to manage in a complete distributed architecture where no forecast is really available. This is why the hybrid approach presents a good level of industrial viability. The second main part of the paper focuses on the ability of the PAC structure to provide a JIT scheduling.
normal operating conditions: does the PAC structure propose a more efficient JIT scheduling than another PAC, when no disturbance occurs? regulation: does a PAC structure propose a more reactive or flexible JIT scheduling than another PAC structure, when disturbances occur?
Relative effectiveness in normal operating condition can be evaluated for the three different PAC structures since they aim at providing schedulings. Hence, comparisons of these schedulings are possible. If planned or non-planned disturbances occur, the way the PAC structure reacts differs conceptually: for centralized structure scheduling may be re-calculated and part of the production process may be stopped, which is not the case for distributed and hybrid structures, which integrate the disturbance dynamically (cf. following part for detailed discussion). Thus, although comparisons of realized scheduling are always possible, it would not be judicious to compare them since they are obtained through too different conceptual solutions. Of course, relative comparison of hybrid vs distributed control on regulation is possible. It is important to note that relative performances evaluated on normal operating conditions are influenced by the typology of processed tasks and the workload, which may condition the relative effectiveness of control systems. Hence several sets of manufacturing orders will be studied to minimize such an influence. Table 1 sums up the possible relative evaluations. 5.1 Influence of the reactivity of a PAC structure on the JIT scheduling (regulation) Reactivity can be defined as the ability of a structure to support random events that may occur during the production process.4 We define a random event as a new real time constraint that was not forewarned. Three axes have been defined: 5.1.1 Resource viewpoint This axis includes (but is not limited to): -
Breakdown or unavailabilities of human operators or production mean for resources,
55
JIT scheduling
Table 1. Relativeevaluationaccordingto tbe structwe typology Relative evaluation through . . . Distributed control Hybrid control Centralized control
Hybrid control
Centralized control
Regulation + normal operating conditions
Normal operating conditions
Distributed control
VS
Normal operating conditions
Regulation + normal operating conditions Normal operating conditions
Normal operating conditions
Breakdown of components of the production structure (wearing parts, etc.).
5.2.3 Product viewpoint Modification, integration or suppression of new or old products according to the life cycle, etc.
Flow viewpoint
Stockout (sub-contracting or supplier problem, etc.). Priority manufacturing order: the manufacturing order priorities may be modified (e.g. some new high-priority manufacturing orders may appear), etc. Product viewpoint
Product modifications: the manager may modify some production parameters according to customers’ wishes (e.g. routing), etc. Supplier mistake, etc. 5.2 Influence of the flexibility of a PAC structure on the JIT scheduling (regulation)
The more a production structure is flexible, the easier it is to re-modify the production structure to support new planned objectives or constraints. That is, the easier it is to integrate dynamically new constraints or objectives that have been scheduled, avoiding as much as possible to question the production process, including the notion of self-configurability.4 Again, three axes are defined: Resource viewpoint
Modification of the structure: such as addition or suppression of resources (global workload modification), etc. Modification of the production facilities: production facility change or improvement, etc. Flow viewpoint
Routing or production flow optimizations, (e.g. depending on the product modifications), etc. wax
5.3 Performance evaluation of PAC structure on JIT schedulings (normal operating conditions) Once the scheduling realized, it would be interesting to evaluate it vs other schedulings realized by other production control structures in the case where no disturbance occurs. In our case, to evaluate the centralized vs the distributed and hybrid architectures in normal operating conditions, we need to create a static set of tasks. Hence, we do not evaluate the reactivity nor the flexibility.
6 CASE STUDY In this part, we will first sum up the well-known performances of centralized control in terms of theoretical reactivity, and flexibility through JIT scheduling effectiveness. Then we will focus more precisely on the effectiveness of the two other previously presented architectures. Some examples are provided when required to illustrate the discussion. Since the effectiveness of production control depends on the nature of tasks in normal operating conditions, the performance of the realized schedulings (centralized, distributed and hybrid control systems) will be evaluated through a complete study on different sets of manufacturing orders. The production structure studied in this paper is a finishing FMS in a foundry factory. The whole production process can be described through two main steps as shown in Fig. 3. The first step is considered as a flow shop production
covering
. flow shop
\
Fig. 3. A synthesis of the production flow.
Y job-shop
I
D. Trentesaux, C. Tahon, P. Ladet
56
system composed of four sections: the first section (prototyping section) realizes some moulds of the product to realize. Resources are classical mill, lathe or electrophoresis. The second section (wax section) produces as many wax-compounded products as required (including potential loss according to scrap rates) from the moulds and operators fix them on clusters. Resources of the third section (ceramic coating section) coat these clusters with a fireproof ceramic and bake them (wax is destroyed). We obtain then an empty ceramic mould which is filled by melting steel. After cooling, the mould is destroyed (pre-finishing section) and the steel clusters are inputted into a finishing workshop. Second step: a finishing workshop receives sets of semifinished products grouped in clusters. The characteristic of this workshop belongs to the job-shop typology and is actually structured as a FMS. The responsibility of this FMS is to control, to remove the clusters, to saw to eliminate connections, to sand-blast, to heat-treat, to calibrate and to pack the products. Visual control is used several times during the process. A centralized PAC structure establishes each week the scheduling when required. The scope of our case study is to evaluate the relative effectiveness of the three control structures through JIT scheduling performance evaluations. The manufacturing orders and resource characteristics have been simplified for the testbeds. The global workload of the FMS has been overvalued to focus on the influence of the control, else decision makings for allocations should not influence enough the effectiveness of a scheduling. Important lead times allow one to cancel bad-quality decisions. The following figures are screen printings from simulated hybrid and distributed control structures on a single 486-PC-DX2-66Mhz. The prototype has been developed in C++ programming language. The environment is Microsoft Windows 3.1. Each detailed scheduling (hybrid and distributed control structures) is established dynamically as and when task allocations are performed, the time scaling is automatic. When the centralized control is evaluated, the schedulings are calculated by the SavePlan SoftwareI in a regressive mode (tasks are scheduled in a JIT way, that is as late as possible). 6.1 Effectiveness of centralized control in terms of flexibility
the production system is strongly dependent on production control. Hence, flexibility and reactivity are costly. Thus, centralized control structures are very powerful when the production system is seldom disrupted (reactivity: steady supply and sell forecast periods, low failure rates of resources, etc.) or weakly coupled with the variations in customers’ demand (flexibility). 6.2 Reactivity of distributed and hybrid PAC structures (regulation) 6.2.1 Introduction
Reactivity will be evaluated through three examples: (1) A testbed on reactivity - resource viewpoint for distributed and hybrid PAC structures. Since the hybrid structure is based upon the distributed approach, the characteristics for reactivity are the same. When a disturbance occurs, the bottleneck may realize a scheduled task later than expected and the other resources from both distributed and hybrid structures act alike. Yet we can not provide the same conditions for resource failures and so for consequent evaluations. Indeed, the schedulings realized by both structures are different. Hence, it is difficult to evaluate the influence of reactivity - resource viewpoint - on hybrid vs distributed JIT schedulings through failure testbeds. We only provide examples to show the integration of a real reactivity - resource viewpoint - support. No comparison is feasible. (2) A testbed on reactivity - flow viewpoint - to compare distributed vs hybrid structures. The behaviours of both control structures differ in the way they are able to manage the new priorities of manufacture. In that case (and unlike the previous testbed), it will be possible to create the same production disturbances by introducing new manufacturing orders at the same date for both structures. (3) A testbed on reactivity - product viewpoint for distributed and hybrid PAC structures. The DSS will be dynamically parametred in a different way to show the influence on the JIT schedulings. Thus, comparisons will be feasible. The study will focus on task allocation decision making. Thus only the two last testbeds will allow result comparisons on JIT schedulings. The first will provide illustrations on the support of resource failures by the two structures.
and reactivity (regulation)
6.2.2 First example: reactivity testbed (resource The effectiveness of centralized control in terms of reactivity and flexibility is well known:2’8 reactivity is highly reduced once a disturbance implies C( r, p, 2) = 0. In that case, the whole production system is stopped and a new scheduling must be calculated. Most of the industrial implementations of such NP-hard algorithms require non-negligible calculation times, which may imply the collapse of the production effectiveness when unconsidered events occur. The production activity control system is strongly coupled with the production system and each element of
viewpoint) for the distributed and hybrid PAC
Here we study the ability of the structures to support malfunctions of components. A product that can not be finished may stay on the deficient IMS until repair or can be re-allocated to other IMS. This aspect needs particular rules or decision help to make a correct decision. We assume here that no re-allocation is realized, although it can easily be managed through the DSS. Distribution of knowledge implies a better isolation of each entity. Failure or malfunction of a component will not make the global production system collapse, but will
JIT scheduling
57
MS 11, Sawing IMS 10, Calibr. MS 3, Saud-blst. IMS 8, Packing It& 7. Calibr. It& 6, T-irttat. IMS 5. Retouch. MS 4, Sand blst. IMS 3. Sawing IMS 2, Cutting IMS 1, v ctr1
Fig. 4. A
complete detailed scheduling (no breakdown).
reduce only the global capacity as it will be shown through the testbed. For both studies (hybrid and distributed structures), nine manufacturing orders (MO), with an average of eight jobs per order, are to be processed by the production system. Failing IMS can neither work on the product during the breakdown duration nor communication with other IMS (total breakdown symbolized by a black and white striped rectangle included in the working time). For both structures, we present the complete detailed scheduling obtained without any extra random event nor attempt at flexibility evaluation. We consider this detailed scheduling as the main scheduling to be compared with the oncoming two different schedulings. The I- character in front of each IMS aims at indicating the completion date for an IMS to perform a task and will be used to evaluate flexibility in the next part. 6.2.2.1 Distributed PAC structure. Figure 4 shows a scheduling realized without any breakdown, Figs 5 and
6 show two different detailed schedulings with 1 and 3 breakdowns, respectively. The results are shown in Table 2. We can not evaluate relative effectiveness of the three schedulings since breakdowns have not the same influence and the same importance. The study could only focus on the evaluation of such influence, but it is beyond the scope of this paper. In fact, this part aims at proving that real-time control provides high-level support for reactivity resource viewpoint -. 6.2.2.2 Hybrid PAC structure. Figure 7 shows a scheduling realized without any breakdown. Black triangles point out the scheduled dates for each job on bottleneck resource. White triangles point out the real starting date for the job. A possible extension could be the integration of specific heuristics that could anticipate the resolution of the breakdown and could locally optimize the scheduling of the bottleneck by providing re-calculations of the bottleneck scheduling. Some studies are being made on this aspect.
IMS 11, Sawing IMS 10. Calibr. MS 3, Ssnd-blst. IMS 8, Packing IMS 7. Calibr. IMS 6. T-Treat. IMS 5. Retouch. MS 4, Sand-blst. IMS 3. Sawing IMS 2. Cuting IMS 1, V-Ctl
Fig. 5. A
complete detailed scheduling (1 breakdown (IMS No. 4)).
58
D. Trentesaux,
C. Tahon, P. Ladet
MS 9, Sand-bist
IMS MS
6. T-Treat. 5, Retouch.
MS 4. Sand-blat.
Fig. 6. Another detailed scheduling (three breakdowns (IMS Nos. 3, 6, 7)). Table 2. Scheduling effectiveness for distriboted and hybrid controls Distributed control No breakdown
0% 0)
(3.1;158*5)
Hybrid control
1 breakdown
3 breakdowns
No breakdown
1 breakdown
5 breakdowns
(6.4;164.8)
(84$244.9)
(42.2;41.2)
(93.2;69.7)
(91.6;69.7)
What is important to note is the relative importance of breakdowns (cf. previous part). Two breakdowns occurring resnectivelv on two resources should influence less or more ;he effectiveness of the schedulings, cf. Figs 8 and 9. It is not mainly due to the workload of those resources: this is illustrated in Table 2, hybrid control: a 5 breakdown-scheduling is more JIT than a single breakdown scheduling, although one of the five breakdowns occurs on the bottleneck resource.
6.2.2.3 Conclusion. This testbed shows how resource failures are automatically supported by the distributed or hybrid PAC structure, breakdowns only alter the global process for the allocation. Thus, reactivity for resource viewpoint is well-supported by the distributed and hybrid control structures.
_I
It is important not to try to estimate how a breakdown could alter the relative effectiveness of JIT schedulings since: this assumes that the same breakdown should occur on the same resource for the same job. This is hard to realize: the real-time schedulings differ; this also assumes that the breakdown should have the same influence on both controls, which is false. 6.2.3 Second example: reactivity viewpoint) for the distributed
In this case study, five manufacturing orders, with an average of 6-8 operations each, are to be processed by the production system from the date 0. The mean due
BI
MS 9, Sand-blsi.
IMS
testbed (jlow vs hybrid structure
5. Retouch.
MS 4, Sand-blst.
Fig. 7. A complete detailed scheduling (no breakdown).
JIT scheduling
59
MS 11. sawing IMS 10, Calibr. IMS 9, Sand-blst. IMS 8. Packing MS 7, Calibr. IMS 6, T-Treat. MS 5, Rclouch. MS 4, Sand-blet IMS 3, Sawing MS 2, Cutting IMS 1*V-Cl11 It
292.975
585.95
878.925
1171.9
Fig. 8. A complete detailed scheduling (one breakdown (IMS No. 5)).
date is 500. At date 200, two other manufacturing orders are to be processed (due dates: 900 and 1000) and at date 320, two other manufacturing orders (due dates: 1000 and 1100). Thus the distributed and hybrid structures will have to process a final set of nine manufacturing orders. The production structure is the same for both control architectures and no breakdown is allowed for the legibility and the relevance of the results. 6.2.3.1 Distributed control. The global scheduling shown in Fig. 10.
is
6.2.3.2 Hybrid control. The global scheduling is presented in Fig. 11. From date 0 to date 200, IMS 6 is the bottleneck. The real-time workload at date 200 is maximum for IMS 8 that becomes the bottleneck from this date to the date 320. Then, due to the two last manufacturing orders, IMS 6 again becomes the bottleneck from this date till the end. 6.2.3.3 Scheduling eflectiveness study for reactivity flow viewpoint - evaluation. To compare the effectiveness
of both structures, it is interesting to evaluate the schedulings in terms of just in time production. The values (m, a) are: (md, odd)= (6.1; 101.4) for the distributed structure, (Q, ~7~)= (26.7; 33.4) for the hybrid structure. According to the definition of the relation between two JIT schedulings, 84 = $. Nevertheless, relative to the average due date, the average tardiness (or lead) is negligible: in both cases, the mean tardiness is no more than 5% of the average due date, but ad is too high to be negligible and alters considerably the effectiveness of the scheduling in JIT terms. The conclusion of this testbed illustrates the previous discussion: the difference holds in the ability for the hybrid structure to re-evaluate the bottleneck (real-time workload evaluation) and to take into account the new priorities of the manufacturing orders through the re-scheduling of the bottleneck. The distributed structure will only consider the new set of manufacturing orders as a set in which the earliest starting time is
It& 10, Csltbr.
MS 6, l’ Treat. MS 5, Retouch. IMS 4, Sand-blst.
Fig. 9. Another detailed scheduling (five breakdowns (IMS Nos. 3, 4, 6, 7, 10)).
D. Trentesaux, C. Tahon, P. Ladet
60
IMS 11. Sawing MS 10, Cslibr. IMS 9. Sand blsl.
MS 6, i Ireat. IMS5, Iktouch. IMS 4, Sand blst.
Fig. 10.
Distributed control.
the date when this set is available. That may lead to unsatisfactory results. 6.2.4 Third example: reactivity testbed (product viewpoint) for the distributed vs hybrid structure 6.2.4.1 Testbed through the utilization of the DSS of both distributed and hybrid structures. The behaviours for the
product viewpoint (e.g. manufacturing order priorities to be modified, tasks to be allocated on particular resources according to users’ wishes, modification of the weights of criteria, etc.) can be dynamically studied to evaluate the ability of the two PAC structures to support modifications on global or local objectives (interactive mode of the DSS). We only focus on task allocation problems. Other case studies should focus on queue management, re-allocation decision making, and so on.
IMS
1 I,
The nine previous manufacturing orders are available at date 0. The criteria favoured have been modified at date 200. Table 3 shows the influence of the relative weights of some criteria used to allocate tasks. The study focuses on two different criteria and the mix of them: 1. Favour the IMS that proposes the fastest process time (FPT). 2. Favour the IMS that proposes the shortest next available time (SNAT) for processing the job. 6.2.4.2 Scheduling efectiveness study for reactivity product viewpoint - evaluation. The results are shown in Table 3. The graph (S”, D, M) is described in Fig. 12. The arrow represents the relation ‘is better than’ (that is: D). No arrow means incomparability between the schedulings (that is: x). F rom this it can be deduced
Sawing
It& IQ. Calihr. MS 9, Sand blst.
MS 6, T-Treat. IMS 5, Retouch, MS 4. Sand-blst.
Fig. 11.
Hybrid control.
61
JIT scheduling Table 3. Influence of criteria on the effectiveness of the scheduliugs Distributed structure
Hybrid structure Scheduling Manufacturing order No. 1 2 3 4 5 6 7 8 9
$h/FFT
~~~SNAT
gh/MIX
sd/FPT
gd/SNAT
Ed/MIX
Tardiness or lead FPT
Tardiness or lead SNAT
Tardiness or lead MIX
Due date
Tardiness or lead FFT
Tardiness or lead SNAT
Tardiness or lead MIX
250 700 450 300 500 880 850 950 1000
-0.4 f73.3 +35.4 +35.4 +16.6 f101.9 +27.6 f36.9 +124.7
+3>3 +85.3 +47.4 +55.8 +32.5 +113.3 +39.6 +48.9 +1326
+11.6 +23%6 +63.3 +55-8 +32.5 t-162.9 -5.3 +97*4 +132.6
f282 1-281 -82,8 +318 -101.6 -96.6 +37.7 +54.7 +164
+32.3 +232.2 f47.4 $55.8 +32.5 -146 -11.7 -7.6 +101.7
f41.7 $132 +433.4 +47.4 -67.7 -246.2 -111.9 -7.6 +101.7
(50.2;41.2)
(65,3;36.6)
(87.7;78+)
(95.2;171.7)
(37.4;100.4)
(35.9;189)
64 4
identical for the resource viewpoint. The difference is set at the flow and product viewpoints where hybrid control seems to perform better than the distributed approach, but this preference is not categorical.
6.3 Flexibility (regulation)
Fig. 12. The graph for relative evaluation of JIT schedulings.
that no scheduling outclasses the others. The only conclusion is that the hybrid PAC structure seems to provide schedulings globally better than the distribution approach, as the second testbed shows. To conclude on the reactivity evaluation through scheduling effectiveness, the performances are quite
of distributed and hybrid PAC structures
6.3.1 Introduction In this part we evaluate the ability of the distributed and hybrid controls to support some new constraints that are planned. One example on the resource viewpoint is provided. In both control structures, self-configurability is by nature automatically integrated. This is due to the sharing of knowledge and information and to the realtime aspect of the control. The example will show how planned modifications (such as addition of an off-line resource) can be easily supported by both control structures, improving or not the performances of the FMS in JIT terms.
IMS 1 U. Calibr. MS Y. Sand blst.
IMS
7. Caiibr.
IMS 6, 1 Treat Il.6
5, Retouch.
MS 4, Send-blst.
Fig. 13. No modification
of the production
structure (distributed control structure).
D. Trentesaux, C. Tahon, P. Ladet
62
IMS 11,Sawing MS
10.Cslibr.
IMS 9,Sand-blst. MS &Packing MS 7,Cslibr. IMS fi, T-Treat. MS 5,Retouch. IMS 4.Sand-blst. MS 3.Sawing IMS 2,Cutting IMS 1,V-Cttl
Fig. 14. From date 0 to 240, 11 IMS were working. Table 4. Comparison of JIT scheduling effectiveuess for flexibility - resource viewpoint Distributed control 11 IMS (fi, o)
(854;2459)
Hybrid control
12 IMS
11 IMS
12 IMS
(79;288.3)
(208.9;96.4)
(112.6;54.6)
6.3.2 Fourth example: flexibility testbed (resource viewpoint) for the distributed vs hybrid structure 6.3.2.1 Distributed control. The following example (10 manufacturing orders) shows how the distributed PAC structure integrates dynamically the addition of a new IMS. This addition concerns a low-performance resource (IMS No. 12) which is only used to prevent extra local workload. This addition is scheduled at date 200 where it is observed from the beginning that the workload of IMS No. 5 will alter considerably the effectiveness of the scheduling. Figure 13 will allow us to rule on the effectiveness of the addition of IMS No. 12 (Fig. 13 is to be compared with Figs 14 and 15). The results are detailed in Table 4. The effectiveness
of the distributed control is there evaluated regarding the hybrid control for the same condition: IMS No. 12 is available at date 200. 6.3.2.2 Hybrid control. Since a period of time lost on the bottleneck should alter considerably the effectiveness of a JIT scheduling, it is important to support addition or suppression of resources that could help for a correct effectiveness. This is a major aspect of the Goldratt (and Marris) thesis: avoid possible lost time on bottleneck, since they make the effectiveness of the scheduling collapse (that is, an hour lost on the bottleneck resource is an hour lost for the whole production system). The results are detailed in Figs 16-18. 6.3.2.3 Scheduling effectiveness study for Jlexibility resource viewpoint - evaluation. Since the new resource is provided at the same time for both structures, it would be interesting to evaluate the JIT schedulings, which is summed up in Table 4. The distributed control presents several difficulties in supporting high workload, although the average lead
MS 12,Retouch. IMSll,Sawing MS 10,Calibr. MS S,Sand-blsL MS 8,Packtng IMS 7,Calibr. IMS 6, r rrcst. IMS 5,Retouch. MS 4.Sand-blst. IMS 3,Sawing IMS 2,Cutting IMS l.V-Ctrl 0
__-_.-__._-____+___-_-_ 351.5
103
11354.5
Fig. 15. IMS 12 is provided to increase efficiency.
14%
63
JIT scheduling
MS 10. Calibr. MS 9. Sand-blst.
IMS 6, T-Treat. IMS 5. Retouch. MS 4. Sand-blst.
0
403.7
Fig. 16. No modification of production
It&
11, sawing
MS
10, Cellbr.
806.4
1?1l!l.6
1611
structure (hybrid PAC structure).
. II
J
)
IMS 9, Sand-blst. IMS g. Packing IMS I, Calibr. IMS 6. T Treat. IMS 5, Retouch. IMS 4, Sand blst. IMS 3. Sawing IMS 2, Cutting IMS 1. v CWI 0
60
1WU
74u
Fig. 17. From date 0 to 240, 11 IMS were working.
IM:;
11. Cietnuch.
IMS
11.
Sawing
IMS 10, Calibr. IMS If, Smd blst.
. sj
i
IMS: 8. Packing IMS7. Cnlibr. IMS 6, I
1rest.
IMS 5, Retowh. IMS
4, Sand-blst.
IM:; 3. Sawing IMS 2. CuRiltg IMS
1. v ari
i
-e0
.1/?.3
744.5
11 16.9
Fig. 18. IMS 12 is used to increase efficiency.
14119.7
64
D. Trentesaux,
C. Tahon, P. Ladet
Table 5. Results for scheduling comparisons Distributed control
normal operating conditions
Hybrid control
Centralized control
MO set No.
Mean
Deviation
Mean
Deviation
Mean
Deviation
1 2 3 4 5 6
50.25 56.87 58.57 -87.86 -157.62 -159.49
244.58 212.17 263.35 288.17 273.73 348.3 1
81.91 86.72 77.28 68.35 76.63 23.10
3437 71.12 48.20 56.67 67.14 11.56
117.00 138.00 128.00 23.00 55.00 -8.00
241.25 282.56 278.52 155.28 175.52 7.89
time is lower than the one realized by the production system controlled by the hybrid structure. This is due to the fact that the distributed PAC structure has caused the production system to finish many manufacturing orders a long time before their due dates and others a long time after. This is confirmed by the important level of standard deviation. On the other hand, the hybrid PAC structure manages the realization of the manufacturing orders according to the JIT approach. Thus, the manufacturing orders should be finished at their due dates. But since the workload is high, they are effectively finished after this date. This explains the high average for lead times and the low level of standard deviations. On the other hand, it is important to note that none of the four schedulings outclasses another one. Hence, it is not obvious that the addition of a resource implies a better JIT scheduling, in both PAC structures. It is very important to evaluate the need for new resource and a strategy should be adopted. This paper only focuses on the ability of the hybrid and distributed control structures to support random or planned constraints for the resource, flow and product viewpoints, not on the way they should use these abilities. Other testbeds on both flow and product viewpoints should be held to evaluate flexibility. Some studies have shown that none of the two structures obviously outclasses the other. The global conclusion on regulation testbed is that hybrid PAC seems to provide more JIT schedulings than the distributed PAC structure. But it would be a mistake to declare that hybrid JIT schedulings outclass distributed JIT schedulings each time. The last part deals with the effectiveness of the three
control structures on the scheduling realized. Thus, this study will focus on the way a PAC structure manages its manufacturing order set regarding the two others. This will help us to answer the question: ‘we know that distributed and hybrid control structures provide regulation supports, but in normal operating conditions, do they provide efficient schedulings?’ The comparison with a calculated scheduling from a centralized structure will help to answer this question. 6.4 Fifth example: Performance evaluation of the schedulings realized (normal operating conditions) Table 5 sums up the testbed for average sets of 10 manufacturing orders. To focus on the JIT support, we provided a decreasing average global workload with the number of sets: that is, global workload for set No. 1 is the higher and global workload for set No. 6 is the lower. The outclassed schedulings are shown in Fig. 19. An outclassed scheduling is greyed. One can note that: -
-
the lower the global workload, the better the centralized control, and worse the distributed control (cf. MO sets Nos. 4, 5 and 6). the higher the global workload, the worse the centralized control and more and more difficult it is for one of the three architectures to outclass the others (cf. manufacturing order set No. 1).
The conclusions are: (1) The centralized control is not as efficient as expected. This is due to the criteria selected. SavePlan prefers to provide JIT with a maximum of manufacturing order sets, at the risk of having huge tardiness on
0
Sdisl
Fig. 19. Outclassed JIT schedulings.
65
JIT scheduling
some manufacturing order sets. This is the case for manufacturing order sets Nos. 1 and 2 where the number of JIT manufacturing orders reaches 60%. But the two last MO are processed with an average tardiness of 600, while the average due date for these two MOs is date 1000. In that case, it would be interesting to extend the definition of the relation ‘is more JIT than’. Of course, this would lead to reconsideration of the relative rankings, and would provide a better ranking for the centralized approach (this is a difficult multicriteria decision making, and it is not the scope of this paper to describe a complete multicriteria methodology to evaluate JIT schedulings, the reader could refer to Roy and Bouyssou5 for more details and studies). On the other hand, when workload is low, the centralized control is very efficient, cf. MO set No. 6: the mean and standard deviation of the tardiness are very low (no more than 0.7% of the mean due date), since it is no more difficult to provide a very efficient JIT scheduling. In fact, such a JIT scheduling is not only due to the strong couplings, but to both centralization and aggregation of data too: optimum decisions are possible and when calculation times are too important, some heuristics may give a solution very close to the optimum one.‘” (2) The effectiveness of the distributed control is not proved. The distributed control is outclassed three times and never outclassed the two other controls. This is due to the bad integration of JIT objectives: if the workload is high, the distributed control has no better solution than the two other structures to solve the task allocation. If the workload is low, no control is provided on the realization of the tasks to move them forward. Thus, the MO are processed in advance (cf. mean finishing times of MO sets Nos. 4, 5 and 6). In fact, the decomposition of global to local information leads to redundancy and restriction of knowledge reducing the performance of the algorithms used to solve problems (non-optimum solutions).19320 (3) The hybrid control is never outclassed and never outclasses any other PAC structure. That is, it is never the best nor the worse control of the three in every case (low through high global workloads). The mean and standard deviation of the tardiness are roughly constant. The control is robust enough to provide good support for the JIT concept and needs.
7 DISCUSSION The relative evaluation of schedulings is still an open problem. One can not provide generic tools for such an evaluation. That is why in this paper we focused only on the JIT support of the three architectures, through the couple (mean tardiness, standard deviation of the tardiness) as criteria. Of course one may argue that it is not enough, and would propose the integration to the relation ‘is more JIT than’ of the number of lead manufacturing orders, for example. So, according to the difficulty in comparing scheduling effectiveness, we use subjective notions to describe such an evaluation. In that case it is more easy to describe the JIT effectiveness through relative dynamic and static evaluations, that is testbeds on regulation and normal operating conditions. Thus we sum up here the results of the five previous testbeds in terms of: -
-
Regulation (hybrid and distributed control): does a control that supports flexibility or reactivity still provide efficient schedulings (in JIT terms) vs the other control when planned and non-planned disturbances occur? Normal operating conditions (centralized, hybrid and distributed control structures): does a control provide efficient schedulings (in JIT terms) vs one another when no disturbance occurs?
We then provide a sum up of the industrial viability of the different architectures. The results are presented in Tables 6-8. Table 6 presents a sum up of examples Nos. 1, 2, 3 and 4 about regulation effectiveness. Table 7 sums up example No. 5 (normal operating conditions). Table 8 gives a summary of the industrial viability level for each of the three introduced control structures. Despite that it is indicated that distributed approach does not present a good level of industrial viability (this paper has demonstrated the limit of such structures in complex production structures such as job-shop), it has been shown that the distributed approach can be considered in special industrial environments, such as flow-shop (packaging, etc.).2’
Table 6. Regulation effectiveness, hybrid vs distributed control
Regulation Reactivity
Does the hybrid control seem to provide more JIT schedulings than the distributed control?
Flexibility
Resource viewpoint (study No. 1)
Flow viewpoint (study No. 2)
Product viewpoint (study No. 3)
Resource viewpoint (study No. 4)
0
+
+
+
Legend for Tables 6-8: --, the answer is no; -, the answer seems to be no; 0, no answer: we can not state; +, the answer seems to be yes; ++, the answer is yes.
66
D. Trentesaux, C. Tahon, P. Ladet
Table 7. Normal operating conditions effectiveness, hybrid vs distributed and centralized controls
the InOut IMS should be able to modify the scheduling if required when a disturbance occurs. For example, the InOut IMS should modify the scheduling to make the bottleneck process a task without waiting for the deficient IMS (if the disturbance occurs on this IMS expected by the bottleneck). This implies forecast modules for the InOut IMS on potential disturbances. Maintenance methods should be strongly implicated in the design of such a system.
Normal operating conditions Global workload regarding global capacity of the FMS
Does the hybrid control seem to provide more JIT schedulings than the centralized control? Does the hybrid control seem to provide more JIT schedulings than the distributed control?
Low
Medium
High
-
0
0
++
+
0
8.2 Evaluating the different PAC structures Table 8. Industrial viability of hybrid, distributed and centralized controls
Industrial viability Distributed control Does the PAC provide a good level of industrial viability?
--
Hybrid control
Centralized control
+
++
8 CONCLUSION This paper has described how JIT concepts can be integrated to a distributed structure through the development of a hybrid structure, which takes advantages from both extremes: -
-
From the distributed structure: it inherits the ability to provide an efficient regulation through flexibility or reactivity, From the centralized structure: it inherits the ability to provide efficient schedulings according to the JIT approach. It also inherits the industrial viability, since the scheduling of the bottleneck allows to forecast the requirements (tools, raw material, and so on) and the potential finishing times.
Some improvements can be made. 8.1 Scheduling the bottleneck The calculation of the workload can be extended through anticipation modules that can be added to prevent potential extra workload (flexibility or reactivity) of a resource. That is, it should not be necessary to declare the bottleneck as the resource with maximum workload. On the other hand, the scheduling of the bottleneck should be modified: -
~
If the support of JIT concepts is not required. For example, it is not necessary to schedule the bottleneck for tasks to be processed as late as possible, but to schedule it to process them in order to minimize the global inventory level. To take into account dynamic reactivity. That is,
Our evaluation holds for the ability for a PAC structure to provide a more JIT scheduling than another through testbeds on regulation and normal operating conditions. Other evaluations could focus on the ability of a PAC architecture to provide for example: -
a maximum number of lead manufacturing sets; a minimum makespan, etc.
order
That is, to evaluate schedulings regarding more criteria than mean tardiness, standard deviation of tardiness. This open problem has not been solved yet. The multicriteria concepts present tools integrating the human way of thinking, uncertainty, miss of knowledge, etc., and could bring efficient solutions to such classical problems in both the design and evaluation phases of schedulings.22>23 If a complete study on the efficiency of a scheduling is then possible, it would be interesting to link such a simulator to other databases from CAD and CAM functions. Thus, the evaluation would not only focus on the time dimension, but on the quality or cost dimensions, which would entirely delimit the problem of global effectiveness of such distributed-based control structures. 8.3 Designing a hybrid structure This paper has presented a new PAC structure that is able to support both JIT concepts and regulation constraints, but it is not described how to design and to implement such a structure. That is, we still do not provide answers to questions such as: For a specific production structure, is it necessary to distribute the control? If a hybrid structure is adopted, when is it useful to temporary add new resources? (cf. fourth example) How to rule shared responsibilities between computers and human operators? Moray et a1.24and Johannsen et a1.25have studied the problematic of decision making by the human operator for task allocations and have focused on the need of balanced workload between the human operator and the computer.
JIT scheduling
Such questions raise the problem of hybrid structure design. Some answers are available in O’Hare.26 This is the next step before a real industrial application.
European Workshop on Integrated Manufacturing Systems Engineering, IMSE’94, INRIA, Grenoble, France, 1994,
pp. 383-93. 13. Sprague, R. H. Jr, DSS in context,
Decision Support
System, 1987, 3, 197-202.
REFERENCES 1. Parunak, V. H. I)., Characterizing the manufacturing scheduling problem. Journal of Manufacturing Systems, 1991, 10(3), 241-52. 2. Trentesaux, D. & Tahon, C., Modele de communication inter-agents pour une structure de pilotage temps reel distribuee. Revue d’Automatique et de Productique Appliq&es, 1994, 7(6), 703-27. 3. Tchako, J. F. N., Beldjilali, B., Trentesaux,
D. & Tahon, C., Modeling with coloured Petri nets and simulation of a dynamic and distributed management system for a manufacturing cell. International Journal of Computer Zntegrated
Manufacturing, 1994, 7(6), 323-39. 4. Trentesaux, D. & Tahon, C., DPACS: a self-adaptative production activity control structure. ZNZZUA/ZEEE Conference on Emerging Technologies and Factory Automation,
Paris, 1995. 5. Roy, B. & Bouyssou, D., Aide Multicritere a la Decision: Methodes et Cas. Economica, Paris, 1993. 6. Zeleny, M., Multiple Criteria Decision Making, McGrawHill, New York, 1982. 7. Tchako, J. F. N., Contribution a la conception d’un Systeme de Pilotage Distribue pour les systemes Automatises de Production. Thesis, LAMIH, Valenciennes, France, 1994. 8. Duffie, N. A., Piper, R. S., Humphrey, B. J. & Hartwick, J. P. Jr, Hierarchical and non-hierarchical cell control with dynamic part oriented scheduling. In Proceedings of NAMRC-XIV, Minneapolis, Minesota, 1986, pp. 504-7. 9. Bakker, H., DFMS: A new control structure for FMS. Computers In Industries, 1988, 10, 1-9. 10. Tchako, J. F. N. & Tahon, C., Distributed management system for a packaging line. In Proceedings of International Conference on Industrial Engineering and Production Management. Fucam, Mons, Belgium, 1993, pp. 833-45.
Il. Ferber, J. Les systemes multiagents. Inter Editions, Paris, France, 1995. 12. Trentesaux, D., Dindeleux, R. & Tahon, C., A multicriteria decision support system for dynamic task allocation in a distributed production activity control structure.
14. Trentesaux, D. & Tahon, C., Dynamic and distributed production activity control: a multicriteria approach for task allocation problematic. International Conference on Zndustrial Engineering and Production Management, Vol. 1, Marrakech, Morroco. Fucam, Mons, 1995, pp. 137-154. 15. Goldratt, E. & Cox, J., The Goal: Excellence in Manufacturing. North River Press, Croton-on-Hudson, New York, 1984. 16. Marris, P., Le Management par les Contraintes en Gestion Zndustrielle. Les Editions d’organisation, Paris, 1994. 17. Saveplan 1984, Software User’s Guide, Sligos company, Paris. 18. Giard, V., Gestion de la Production, 2nd ed. EconomicaCollection Gestion, Paris, 1988. 19. Duffie, N. A., Chitturi, R. & Mou, J., Fault tolerant heterarchical control of heterogeneous manufacturing system entities. Journal Of Manufacturing Systems, 1988, 7(4), 3 15-27. 20. Matsuura, H. & Tsubone, H., A comparison of centralized and decentralized control rules in push-type production ordering systems. European Journal of Operational Research, 1993,70,
199-211.
21. Trentesaux, D., Tchako, J. F. N. & Tahon, C., Distributed and multicriteria management tools for integrated manufacturing. In Lijie Cycle Modelling for Znnovative Products and Processes. Chapman and Hall, London, 1995, pp. 576-88. 22. Belton, V. & Elder, M. D., Can multipie criteria methods help production scheduling? 10th International Conference on M C D M TAIPEI’, Taiwan, July, 1992, pp. 171-8. 23. Tabucanon, M. T., Multiple criteria decision making in industry. Studies in Production and Engineering Economics, Elsevier, Amsterdam, 1988, 8. 24. Moray, N., Dessouky, M. I., Kijowski, B. A. & Adapathya, R., Strategic behavior, workload, and performance in task scheduling, Human Factors, 1991, 33(6), 607-29. 25. Johannsen, G., Levis, A. H. & Stassen, H. G., Theoretical problems in man-machine systems and their experimental validation. Automatica, 1994, 30(2), 217-3 1, 26. O’Hare, G. M. P., Designing intelligent manufacturing systems: a distributed artificial intelligence approach. Computers in Industry, 1990, 15, 17-25.