741

Scheduling for just-in-time flexible manufacturing

using adaptive fuzzy logic

J D Tedford* and C Lowe

Department of Mechanical Engineering, The University of Auckland, New Zealand

Abstract: Production scheduling for flexible manufacturing needs to satisfy multiple and often
conflicting criteria. The multicriteria problem is generally difficult to solve using existing heuristics
without a means of combining rules and adapting them to changing circumstances.
A scheduling system incorporating an adaptive fuzzy logic system is proposed which can be
tailored to suit the characteristics and objectives of an organization and maintain its performance by
constantly adapting to changes in the production environment. The system developed indicates that
its application shows benefits in terms of the resulting schedule performance and with respect to the
simplicity of the scheduling algorithms required. The computational overhead of the fuzzy system is
also shown to be insignificant, thus permitting real-time operation in a manufacturing situation.

Keywords: flexible manufacturing, fuzzy logic, adaptive, scheduling, real-time

1 INTRODUCTION soned schedules which perform better than traditional

heuristics.
The advent of flexible manufacturing systems operat-
ing ‘just-in-time’ (JIT) has placed increasing demands
on the production scheduling process. The importance
of meeting traditional performance criteria such as 2 FUZZY SCHEDULING
makespan and flowtime, etc., need to be weighted to
reflect other important operating requirements. Spe- Fuzzy systems, although mathematical techniques,
cifically, lateness is unacceptable in situations where define their behavioural framework through a com-
an assembly line running under JIT relies on the pact linguistic rule base. They are capable of clearly
timely production of components to supply it. defining most manufacturing environments and are
Much of the research undertaken for flexible manu- able to simultaneously consider multiple criteria. Fur-
facturing systems has focused on the adaptation of thermore, fuzzy systems do not generally present
existing rules [1]. These rules, however, display signifi- severe computational overheads [3].
cant drawbacks due to the assumption that the sys- Consequently, a fuzzy system is able to make rea-
tems operate in steady state conditions and because soned decisions based on many inputs, thus lowering
the rules usually seek to optimize a single criterion. the decision risk and increasing the confidence in the
This generally results in schedules with inherently output. Fuzzy systems may infer a plausible result,
higher risks of failure.
even in the absence of complete input data, and the
Fuzzy logic has been shown to be ideally suited to
fuzzy membership functions and rules can be readily
solving the multicriteria problem through its facility
modified to allow adaptability.
to combine and weight multiple inputs. This enables
Several fuzzy logic based scheduling systems have
fuzzy systems to emulate much of the decision logic
recently been developed [4 – 7], although direct com-
demonstrated by human schedulers [2]. A hybrid
parisons between them are difficult due to their differ-
scheduling method is proposed here which incorpo-
ent implementations and objectives. An alternative
rates adaptive fuzzy logic to generate intelligently rea-
approach to the scheduling problem is presented here.
It replaces traditional scheduling rules and iterative
The MS was recei6ed on 6 April 1999 and was accepted after re6ision simulations with an artificial intelligence based ap-
for publication on 8 April 1999. proach. Key features of the system are:
*Corresponding author: Department of Mechanical Engineering, Uni-
6ersity of Auckland, Pri6ate Bag 92019, Auckland, New Zealand. (a) schedule from the task not from the resource,
SC00999 © IMechE 1999 Proc Instn Mech Engrs Vol 213 Part B

Downloaded from pib.sagepub.com at FLORIDA ATLANTIC UNIV on May 23, 2015

742 J D TEDFORD AND C LOWE

(b) generate a schedule through reasoned decision i.e. a resource (machine) with good reliability and fast
making, performance.
(c) fully utilize available modern manufacturing infor- Components are then backwards scheduled and
mation systems, queued for the selected resource as long as the lateness
(d) build real-time capability into the system, criterion can be met. Order within the queue for a
(e) build adaptability into the system. resource is determined by the priority index. If there is
A fuzzy system incorporating these features was de- insufficient capacity for the component at the preferred
veloped. The attributes of each job in the system await- work centre, the part with the lowest priority index will
ing processing were assigned both a numerical task be transferred from this work centre to an alternate
index and a priority index [0..100], based on the follow- resource with the next closest resource index. This
ing factors: promotes clustered production of related components
and reduces the likelihood of a subsequent assembly
Value the cost of the component line product missing one or more key components.
Run length standard cycle time×run quantity The fuzzy logic functions were implemented as run-
Slack lead time time libraries which were able to be called by the
Type whether the component was produced scheduling software. Design of the fuzzy system was
for inventory, consignment or as an facilitated by the use of FuzzyTech® fuzzy logic devel-
expedited order opment software and later compiled to C+ + code.
These resulting libraries perform only the fuzzy process-
In a similar manner, the performance and reliability of ing functions. A diagram outlining the structure of the
the available resources were assigned a resource index scheduling system is shown in Fig. 1.
[0..100], based on: The membership functions and rule base of the fuzzy
system are stored externally in a relational database
Cost the cost of processing
and dynamically read into the fuzzy system at run-
Run length actual cycle time×run quantity
time. This allows the definition of the fuzzy system to
MFT mean failure time
be readily accessed and changed in real-time, thus
MTBF mean time between failures
Quality typical reject rate enabling adaptability. The interface between the sched-
ule input/outputs, database and fuzzy system is handled
The system then attempts to match a job’s task index by front-end software written using Microsoft Visual
with a resource that has a complementary resource Basic®.
index. For example: The computational overhead of the system is mini-
mal, as tested on a standard Pentium® class personal
If the production run is lengthy and urgent and computer operating under Microsoft Windows NT®.
difficult and the task index evaluates to 90, The fuzzy system described can be loaded and 100 000
Then select a resource with a resource index as close fuzzy calculations performed on each of the operating
as possible to 90. and performance indexes in less than 15 s.

Fig. 1 Scheduling system information flow

Proc Instn Mech Engrs Vol 213 Part B SC00999 © IMechE 1999

Downloaded from pib.sagepub.com at FLORIDA ATLANTIC UNIV on May 23, 2015

 SCHEDULING FOR JIT FLEXIBLE MANUFACTURING USING ADAPTIVE FUZZY LOGIC 743

Fig. 2 Comparison of results after 50 replications

3 SIMULATION MODEL AND RESULTS performance evaluation between the different schedul-
ing methods.
An initial simulation model was constructed using Fifty replications were run equating to one year of
Arena® simulation modelling software. The model for production. The performance criteria included the num-
analysis was based on a subset of the injection mould- ber of orders completed late, average lateness,
ing shop of a medium-sized manufacturing organiza- makespan and flowtime. Figure 2 presents comparative
tion. The simulated model consisted of five injection histograms of the results, where the worst performing
moulders with 25 different components to mould. A dispatching rule is given a normalized value of 100 per
cent. It can be seen that flowtime, using the fuzzy
weekly master schedule was created for the assembly
method, was reduced by 28.8 per cent, while makespan
line and individual component orders derived from it.
was reduced by an average of 11.5 per cent. The
Within the simulator a module varied set-up times,
average lateness and the number of orders completed
cycle times, failures, etc., using distributions derived
late criteria benefited the most from the fuzzy method.
from acquired factory data. Each simulated week pro- The number of orders completed late fell to 45.6 per
cessed 90–150 orders. cent using the fuzzy method, while the lateness of
The simulator is also able to schedule production individual orders was, on average, 62.1 per cent less
using the FIFO, SPT and SLACK dispatching rules in when compared with the other rules—a significant
addition to the fuzzy method proposed here. Each reduction in average lateness.
production scenario was simulated using each of these Following on from these results, a more complex
rules in turn. The randomness of work centre failures, simulation model was developed using Microsoft Visual
etc., was repeatable, ensuring as far as possible that a Basic. Although lacking the graphical displays afforded
similar scenario was being modelled for comparable by Arena, the development of a more flexible simulator
SC00999 © IMechE 1999 Proc Instn Mech Engrs Vol 213 Part B

Downloaded from pib.sagepub.com at FLORIDA ATLANTIC UNIV on May 23, 2015

744 J D TEDFORD AND C LOWE

Fig. 3 Comparison of results using real-time adaptivity

was possible, having superior simulation performance The fixed parameters were:
and with virtually limitless model variations.
(a) constant velocity for transfer between machines,
As configured for the subsequent analysis of the
(b) invariant set-up costs for machines,
fuzzy scheduling technique, the simulator featured: (c) fixed batch sizes, i.e. components were only moved
(a) 100 machines arranged in 20 groups of 5 machines, to the next stage of processing after all components
(b) 1000 tasks comprising a mixture of 200 compo- were completed.
nents with process plans requiring up to 5 different This model featured real-time adaptive fuzzy schedul-
machines per component, ing. Each task when moving from one process stage to
(c) variable cycle and set-up times for each part and another was re-evaluated with respect to its task index
machine combination, in real-time and the resource index of each machine was
(d) variable cost of processing for each component at re-evaluated at the end of each week to account for
different machines, trends in machine performance.
(e) variable reject rate for each machine–tool The results from this model are shown in Fig. 3.
combination, Overall the results reflect a similar pattern to the earlier
(f) variable placement of machines to reflect a real model, with a clear improvement through the use of
factory layout, adaptive fuzzy scheduling. However, the increase in
(g) variable time between failure and failure time pro- task variety and process plans requiring up to five
files, for each machine, which could be set for con- machines had a slightly adverse effect on the degree of
stant, random or trending (gradual wearing out of superiority of the fuzzy scheduling technique. This was
tooling for example), due to some tasks being moved to less than optimal
(h) run-time was for one simulated month of production. machines for processing in order to maintain reason-
Proc Instn Mech Engrs Vol 213 Part B SC00999 © IMechE 1999

Downloaded from pib.sagepub.com at FLORIDA ATLANTIC UNIV on May 23, 2015

 SCHEDULING FOR JIT FLEXIBLE MANUFACTURING USING ADAPTIVE FUZZY LOGIC 745

ably balanced machine utilization at the first stage of modify its decision making in response to changes in
component production. This was compensated for, the behaviour of the system in real-time, thus maintain-
however, by the better resource – task matching that ing its performance superiority over other commonly
was possible to achieve at each subsequent step in the used heuristics.
production of a component. Although adaptability was implemented, it is felt that
The percentage of orders late under adaptive fuzzy under the simulation model tested here the run length
scheduling rose from 50 to 65 per cent relative to the was not sufficiently long to realize its full benefits.
worst performing rule and average lateness of orders Additional simulation runs spanning much longer time
similarly increased from 33 to 47 per cent. This was periods would probably demonstrate these benefits
largely due to the flow-on effects of the lateness in- more fully, since they would enable greater long-term
curred in the early stages of processing for individual trends in the system performance to be observed.
components. Further research will determine whether a Future developments will centre on more variations
change in scheduling strategy to curtail or abandon in models over a longer time period to further refine the
production of some components may be beneficial. adaptive strategy used along with the implementation
Overall, the fuzzy scheduling technique proposed was of a genetic algorithm to assist in a more rapid conver-
able to make, on average, schedules with at least as gence on the best schedule.
good, if not better, performance than the other com-
monly used scheduling rules in terms of flowtime and
makespan. The performance of the fuzzy scheduling
technique, however, was consistently superior with re- REFERENCES
spect to lateness, which is an important criterion for
just-in-time production. 1 Montazeri, M. and Van Wassenhove, L. N. Analysis of
scheduling rules for an FMS. Int. J. Prod. Res., 1990,
The overall computational overhead was more than
28(4), 785 – 802.
acceptable, requiring 1 – 2 min to build and derive a 2 Bartos, F. J. Fuzzy logic reaches adulthood. Control
series of initial schedules on a Pentium class computer. Engng, July 1996, 50 – 56.
3 Shin, Y. C. and Vishnupad, P. Neuro-fuzzy control of
complex manufacturing processes. Int. J. Prod. Res., 1996,
4 CONCLUSIONS 34(12), 3291 – 3309.
4 Caprihan, R., Kumar, S. and Wadhwa, S. Fuzzy systems for
Through the inherent tolerance in fuzzy systems, vary- control of flexible machines operating under information
ing levels of detail can be incorporated as required by delays. Int. J. Prod. Res., 1997, 35(5), 1331 – 1348.
any particular manufacturing organization. This can 5 Grabot, B. and Geneste, L. Dispatching rules in scheduling:
a fuzzy approach. Int. J. Prod. Res., 1994, 32(4), 903–915.
aid the practical implementation of reliable scheduling
6 Somlo, J. and Watanabe, T. Intelligent scheduling of FMS:
where there is missing or incomplete information avail- the optimisation approach. In JAPAN/USA Symposium
able to the scheduler. on Flexible Automation, 1992, Vol. 1, pp. 447 –455.
The results to date indicate that the fuzzy scheduling 7 Watanabe, T., Tokumaru, H., Nakajima, Y. and Hashi-
technique presented is able to make schedules of im- moto, Y. Job-shop scheduling using fuzzy inference to take
proved performance over other commonly used profit into account. In JAPAN/USA Symposium on Flex-
scheduling rules. Its adaptivity provides the ability to ible Automation, 1992, Vol. 1, pp. 423 – 427.

SC00999 © IMechE 1999 Proc Instn Mech Engrs Vol 213 Part B

Downloaded from pib.sagepub.com at FLORIDA ATLANTIC UNIV on May 23, 2015