=Paper= {{Paper |id=Vol-2739/paper_6 |storemode=property |title=Prescriptive System for Reconfigurable Manufacturing Systems Considering Variable Demand and Production Rates |pdfUrl=https://ceur-ws.org/Vol-2739/paper_6.pdf |volume=Vol-2739 |authors=Catarina Baltazar,João Pedro Correia dos Reis,Gil Gonçalves |dblpUrl=https://dblp.org/rec/conf/sam-iot/BaltazarRG20 }} ==Prescriptive System for Reconfigurable Manufacturing Systems Considering Variable Demand and Production Rates== https://ceur-ws.org/Vol-2739/paper_6.pdf

Prescriptive System for Reconfigurable
Manufacturing Systems Considering Variable
Demand and Production Rates
Catarina Baltazar, João Reis, Gil Gonçalves
SYSTEC, Research Center for Systems and Technologies
Faculty of Engineering, University of Porto
Rua Dr. Roberto Frias, 4200-465 Porto, Portugal
Email: {up201406435, jpcreis, gil}@fe.up.pt

Abstract—The current market is dynamic and, consequently, (PHM) frameworks emerge as they allow improvements in
industries need to be able to meet unpredictable market changes reliability and reduction of costs associated with maintenance
in order to remain competitive. To address the change in actions [3]. Advances in the Information and Communica-
paradigm, from mass production to mass customization, manu-
facturing flexibility is key. Moreover, current digitalization of the tion Technologies domain enable the development of more
industry opens opportunities regarding real-time decision sup- sophisticated PHM tools, especially, based on Deep Learning
port systems allowing the companies to make strategic decisions, methods as they simplify the process of feature learning
and gain competitive advantage and business value. and have superior performance. Deep Learning approaches
The main contribution of this paper is a proof of concept represent a promising path towards a one-fits-all framework
Prescriptive System with a highly parameterizable simulation
environment catered to meet the needs of Reconfigurable Manu- [4]. An effective PHM system should be able to timely
facturing Systems allied with an optimization module that takes predict failures by constantly monitoring health status of the
into consideration productivity, market demand and equipment equipment and also isolate and identify the faults [5]. Addi-
degradation. With this system, the effects of different throughput tionally, it must support decision-making systems to take full
rates are monitored which results in better recommendations strategic advantage of the predictions provided by diagnosis
to mitigate production losses due to maintenance actions while
taking into consideration the health status of the remaining assets. and prognosis techniques [6]. While prognosis is related to
In the proposed solution the simulation module is modeled failure prediction and tries to answer the questions ”What will
based on Directed Acyclic Graphs and the optimization module happen?” and ”When will it happen?” [7], diagnosis consists
based on Genetic Algorithms. in identifying and isolating the faults. Despite the intuitive
The results were evaluated against two metrics, variation of relationship between predictions and prescriptions, and the
pieces referred as differential and availability of the system.
Analysis of the results show that productivity in all testing
undeniable benefits to gain competitive advantage, prescriptive
scenarios improves. Also, in some instances, availability slightly systems’ area is the field with less research [8]. These systems
increases which shows promising indicators. intend to recommend one or more courses of action based
Index Terms—Reconfigurable Manufacturing Systems, Indus- on predicted future and, therefore, allow to take proactive
try 4.0, Variable Throughput, Genetic Algorithm measures [7].
A thorough review of prescriptive systems is given by [8]
I. I NTRODUCTION where three categories were identified: production schedul-
Nowadays industries face constant changes as the result of ing, life cycle optimization, supply chain management and
unpredictable market trends. The challenge is to be flexible logistics. For example, regarding inventory management, in
enough in order to respond in a timely manner to clients both [9] and [10], spare parts are ordered based on equipment
demand while maintaining a sustainable cost structure to degradation. In the former, decisions regarding the purchase
remain competitive in a fierce business environment. For the of spare parts are decided based on the levels of degradation
purpose of attending markets needs, it is necessary to increase observed during irregular inspections. In the latter, long short-
the efficiency of manufacturing processes in which machinery term memory (LSTM) networks are employed to predict
plays a fundamental role. failure probability during different time windows. Then, based
Reconfigurable Manufacturing Systems (RMS) arise to deal on the information provided by the prediction model, the
with uncertainty and individualized demand [1] by combining appropriate options regarding maintenance and order of spare
advantages of both Dedicated Manufacturing Lines and Flex- parts are chosen.
ible Manufacturing Systems [2]. Moreover, during the current From the three categories identified, in an industrial context,
industrial revolution, also referred as Industry 4.0, significant maintenance scheduling is the more predominant one. In
interest in the upgrade of Prognostics and Health Management [11], a Genetic Algorithm (GA) is employed to optimize

38
maintenance scheduling for manufacturing systems with a According to [15], these configurations are defined as Class II
fixed structure. In this paper, it is assumed that the infor- RMS.
mation regarding failure probabilities is available. Similarly,
in [12], a GA is used to schedule maintenances based on
machine degradation. However, in this case, the variables that
are optimized are the throughputs of machines and possible
maintenance actions instead of discrete time moments. In
general, the proposed optimization procedure searches for the
best trade-off between maintenance actions and throughput
settings. Likewise, in [13] a continuous maintenance system
based on real-time monitoring is proposed. The optimization
module is also based on GA and assures production targets
by searching the best sequence of machine throughputs taking
into consideration equipment degradation. In contrast, in this
paper, a Predictive Maintenance module is integrated and
the GA helps in avoiding unexpected breakdowns based on
constant condition monitoring in real-time. Solving scheduling
problems is not limited to the application of GA but these Fig. 1. Generic Manufacturing Layout
algorithms represent the majority of the proposed solutions
[14]. Accordingly, DAGs were chosen to model the system. This
Few Prescriptive Systems are applied to RMS. In this approach allows the rapid response in changing layouts con-
context, the mitigation of production losses due to machines figuration and the control of pieces flow in the manufacturing
downtime can be achieved not only by tuning throughputs system. In order to implement it, the package Networkx, only
of different machines, but also by routing pieces to healthy available in Python, was chosen.
assets. Accordingly, the main contributions of this paper are an Each node of the graph represents a machine and the edges
optimization approach that shows good indicators in finding connections between machines that might be, for instance,
throughput sequences that balance productivity and mainte- conveyor belts. The edges are weighted and represent path
nance actions in a RMS context, as well as a straightforward priority. The lowest the weight the higher the priority. This
simulation module based on Directed Acyclic Graphs (DAG) approach allows to favour, for example, the shortest path when
that allows quick layout changes and easy parametrization deciding to which machine should the piece be sent.
of the shop-floor namely, scheduling of maintenance shifts, The machines are represented by the class Machine and each
different types of failures and types of equipment. instance represents a node of the graph. This approach allows
The remainder of this paper is organized as follows. In high parameterization of the equipment and the parameters can
section II both simulation module and optimization module be separated in three main groups:
are discussed. Then, in section III, the scenarios that are tested • Identifying Parameters: relate to the identification of the
in order to validate the solution were presented. Additionally, equipment
some preliminary results are discussed. A more in depth – machine id;
analysis of the results presented in the previous section can – type of machine;
be found in section IV and finally, in section V conclusions – age;
and future work are discussed. – line;
– stage;
II. I MPLEMENTATION • Operations Parameters: relate to the machine operation
The proposed Prescriptive System is mainly composed of – available operations;
two modules: simulation module and optimization module. In – current throughput;
the following subsections each module is further described and • Reliability-related parameters: relate to degradation of the
this current section concludes with the interactions between the equipment
two. – mean time to repair (MTTR);
– mean time between failures (MTBF);
A. Simulation – types of failures.
The goal is to model manufacturing layouts such as the one Concerning to identifying parameters, line and stage cor-
presented in Fig. 1 so it allows easy changes in configurations respond to the position of the machine in the layout, Fig. 1,
in order to respond to different demands in the future. These while the remaining parameters in this category are related to
configurations possess crossovers and all machines within the specifications of the equipment. In respect of operation param-
same stage execute the same tasks. Consequently, pieces in eters, available operations relate to the range of operations that
stage i can be transferred to any machine at the stage i+1. the machine can perform and current throughput identifies the

39
production rate at which the equipment is operating. Lastly, working days and working hours are considered. In regards to
regarding reliability-related parameters, this are of the utmost maintenance shifts, if one decides to integrate them in the
importance to simulate the degradation of the equipment. In simulation, the starting times and duration of said shifts can
terms of different types of failures, each machine can have be defined. The only thing, which in some cases might be
associated different ones which will correspond to different considered a limitation, is the fact that the maintenance shifts,
MTTR and, as a result, maintenance actions will have different by default, are periodic. Simply put, in every working day
periods of time. Also, MTBF will be used as a mean to predict the shift starts at the same time and has the same duration.
the failure. Additionally, different sequences of operations can be applied
In addition, in this case, the machines are also responsible to to the pieces in order to achieve different final products as long
control the flow of production in the shop-floor. Each machine as the needed operations are available in the current machines
has a state machine associated as the one represented in Fig. and as long as the operations can be performed in a sequential
2. manner as represented in Fig. 1. All these features allows the
simulation of a wide variety of scenarios not only on time
domain but also specification wise.
In this paper, it is assumed that the information regarding
probability failures is known, as no predictive model is pro-
posed. Recalling the parameters associated to each machine,
namely, reliability-related ones, both MTBF and MTTR are
known. In a simplified manner, MTTR refers to the average
time to repair certain component and MTBF the forecasted
time between failures [16]. Both these terms will allow to
simulate degradation of the equipment as well as management
of maintenance actions in order to implement the present
system. As a result, the prediction of a pending failure will be
calculated based on the difference between MTBF and current
simulation time. If that difference is below a certain threshold,
the failure will be signaled and maintenance scheduling takes
place. Both Fig. 3 and Fig. 4 exemplify how the maintenance
scheduling is handled. The difference between MTBF and
Fig. 2. State Machine associated with each machine current simulation time corresponds to a certain time window.
This time window is the time to failure and is represented
The machine has four states. It starts in its IDLE state and if by the yellow area. If during that time window a shift
the machine is not going to start any maintenance, maintenance takes place, blue area, then the maintenance of the respective
= 0, and is available, the machine can receive pieces. Once the equipment will occur when the shift starts (Fig. 4). Otherwise,
pieces are received they are processed. When the processing an emergency maintenance is triggered (Fig. 3).
time ends, three things might happen: if the next machine
is available the piece is dispatched and then the machine
can return to its IDLE state or IN MAINTENANCE state.
Otherwise, it will transition to WAITING state. This transition
happens when there are no available machines and the current
machine behaves as a buffer until a possible machine becomes
available. While in the WAITING state, the machine cannot
receive any pieces. In the case that the machine does not
have the respective tool, the piece experiences the same cycle,
however, processing times are equal to zero. In short, the Fig. 3. Pending Failure that will result into an emergency maintenance
edges of the graph provides the different connections between
machines and each connection is only admissible if green-
lighted by the destination machine state.
In addition, not only machines can be parameterized but also
other parts of the manufacturing environment. The simulation
module developed in this paper takes into consideration,
different simulation times, maintenance shifts and different
sequences of operations to apply to different raw materials.
Simulation times are related to how many seconds each tick
(time unit in the simulation environment) worth and how many
working weeks are being simulated. Also, it defines how many Fig. 4. Pending Failure that will result into a scheduled maintenance

40
Furthermore, different machines’ throughputs have different
impacts in degradation of the equipment. As stated in [12],
when a machine decelerates it is expected that its degradation Fig. 5. Chromosome Structure. Source: [13]
slowdown, and vice-versa if a machine increases its through-
put. To simulate the degradation effects influenced by the
chosen production rates, the MTBF will be inversely propor- Ti,j is an integer between -2 and 2 and corresponds to the
tional to production rate. Similar to [13], five throughputs are machine i operation mode at the day j. Thus, the size of the
available where mode 2 increases production rate two times chromosome is variable and equal to i × j.
in regards to baseline production, mode 1 production rate is Companies’ main goal is to attend customer’s needs while
1.5 times higher, mode 0 corresponds to baseline throughput, remaining competitive and profitable. Therefore, it is crucial to
mode -1 production rate decreases in 1.5 times and, lastly, meet production targets in the most efficient way. Accordingly,
mode -2 where production rate decreases 2 times. the fitness function (1) not only takes into consideration
production targets but also machines’ degradation.
B. Optimization
" N N
The optimization module is key to the implementation of the X X
F = min Kp (W − P )2 + Ksm Fsmi + Kem Femi +
Prescriptive System as it is responsible for the compensation of
i i
production losses due to machines’ downtime. A standard GA N
X N
X N
X
#
approach was chosen as its employment is well documented Knw Fnwi + Kch Cchi + Ksd Si
and produces near-optimal solutions [17]. GAs can be under- i i i

stood as an abstraction of the theory of evolution by natural (1)
selection by Darwin and are suitable to solve multi-objective subject to: Fsmi , Femi , Fnwi = {0, 1, ..., N } ∀i
problems [18]. The genetic variability within a population is Cchi = {0, 1, ..., d} ∀i
simulated through mutation and crossover operators and the Si ≥ 0 ∀i
selection is done based on the survival of the fittest [19]. The first term is the difference between production weekly
The optimization module can be triggered in two instances: target, W , and number of pieces produced, P , by the sys-
when an emergency maintenance takes place, or when a tem, squared. In essence, it evaluates how far the system
maintenance does not finish during a maintenance shift. As production is from the target and the square ensures that the
a result, two types of maintenance can be identified: algorithm does not favour solutions that exceedingly surpass
• Emergency Maintenance - a maintenance that occurs the target, and the non-negativity of the values. The following
outside a maintenance shift; three terms are regarding the different maintenances. Each
• Scheduled Maintenance - a maintenance that is allocated type of maintenance is different and, as a result, also their
to a maintenance shift. weight in the fitness function. The second and third term is
Emergency maintenances are more costly not only because scheduled maintenance, sm, and emergency maintenance, em,
of resources allocation, but also their impact in production. respectively, and their different impacts were already stated.
Even if a scheduled maintenance continues beyond the shift Throughout the formulation of the fitness function, initially
duration, the losses in production are lower because the there was no distinction between those two maintenances and
downtime during maintenance shift is expected, which does the results were good so if a more broad approach is desired
not happen in a context of an emergency maintenance. When the maintenance might not be distinguished. However, the
formulating the optimization problem, both types of mainte- prescriptive system proposed has scheduled maintenance shifts
nance are taken into consideration with different weights, as integrated and the distinction between the two makes sense
their impact is also different on production weekly goals. since they have different impacts in the production system.
The used approach follows very closely the one presented in The fourth term is also related to maintenance, but it is
[13]. The proposed formulation was applied to three parallel regarding the first three days of the next week, nw. To increase
machines and can easily be applied to N parallel machines. production the throughput of some machines has to inevitably
However, other configurations require some fine-tuning in their increase, which accelerates the degradation of those machines.
weights and the addition of some terms depending on the So, this term is to prevent new failures in the beginning of the
problem. In summary, the goal is to extend the mentioned next week as it will affect the production goals of the next
formulation to a more broad spectrum of layouts and adapt it week.
to the RMS system considered in this Prescriptive System. The constant change of throughputs in a real production line
Every week, the production should comply with the cus- is not practical. As a result, the last two terms are introduced
tomers orders so the GA optimizes a maximum of one week to promote homogeneous solutions. The first term of the two,
and once the current week ends, the throughputs return to their ch, corresponds to the number of changes in relation to the
baseline unless new optimization takes place in that week and baseline, mode 0, and the second is the standard deviation, S,
the process repeats itself once again. In this regard, each gene of the suggested throughputs to machine i.
of the chromosome will represent the throughput of machine Initially, the weights considered were the same as the ones
i at the day j as represented in Fig. 5. presented in [13]. After several simulations, it was observed

41
that the convergence of the solutions was not quite as desired.
At the boundary of solutions that achieve the weekly targets
and solutions with deficits, sometimes close to 2%, but with
throughput rates more homogeneous, the latter were given
priority (i.e., better fitness values). This behaviour was further
proved by the conduction of a sensitivity analysis where
the contributions from the different types of maintenance
were considered constant and the remaining terms of the
Equation (1) variable. As a term of comparison, margins of
1% in relation to production in regards to the desired targets
were considered acceptable. So, the weights needed to be
refined. Accordingly, based on the previous sensitivity analysis
and additional simulations, the finals weights are as follows:
K p = 10, K sm = 900, K em = 1000, K nw = 300, K ch = 300
and K sd = 400.

C. Prescriptive System
The proposed Prescriptive System involves the two modules Fig. 6. Overview of the proposed Prescriptive System
explained above and an overview can be found in Fig. 6.
Once a failure is detected and if the requirements regarding
the conditions in which the maintenance will occur are met, by the ratio of total real operation time of all machines by
the optimization module is triggered. As shown in Fig. 6, the total theoretical operation time of all machines. Taking
represented by blue rectangles, two instances of the simulation into consideration Fig. 1, the configurations will be referred
module are present: Manufacturing Environment Simulation as nxm, where n corresponds to the amount of stages and m the
and Simulation Module. The former corresponds to the simu- amount of production lines. The GA parameters were selected
lation of the shop-floor of interest and the latter is an image of after several runs and set to:
the former. However, in this case, its purpose is solely to feed • Population size = 100;
the optimization module with the needed variables to evaluate • Maximum generations = 100;
the candidate solutions: pieces produced and number of main- • Mutation Rate = 0.2;
tenances during current week and the following one. These • Crossover Rate = 1.0;
outputs are what allows the calculation of the solution fitness • Crossover Method: Single-point crossover;
value represented by Equation (1). Additionally, in both these • Selection Method: Elitism.
modules, a model to predict failures can be easily integrated. All tests were performed in a personal computer with the
This cycle between optimization module and simulation model specifications: Intel core i5-3750 CPU @ 3.40GHz and 8.00
stops once the termination criteria is met. In this paper, the GB RAM.
optimization stops when the maximum number of generations
is exceeded. When the optimization module finishes, the best A. First Set Scenarios
solution is recommended (white rectangle) and applied to the All tests were performed using a 3x2 configuration. In the
manufacturing environment simulation if the operator decides first test, one of the machines is down a whole working day
to. and another machine is on the verge of failing in the following
week. In the second one, the same machine is down, however
III. S YSTEM VALIDATION AND V ERIFICATION there is a second machine that fails in the middle of the
To evaluate the proposed system the testing was divided week, during half-day. In the third and last test of this set,
into two phases. Firstly, a set of tests are applied in order there are no broken machines but the connections from one
to analyze and validate the results provided by the GA as of the machines are interrupted which isolates the equipment
well as to prove that this system might be easily applied to and, consequently, pieces processed by it have nowhere to
configurations not fully connected or easily upgraded to handle flow to. The main goal of all tests is to understand, under
failure in transport equipment. Secondly, scenarios that are different conditions, if the weekly target is achieved and how
more complex are investigated in order to check scalability. the algorithm deals with the different maintenance moments.
The simulation time in all tests is one working week. Also, However, the Test 3 is performed not only as a mean to study
there will be two shift changes per working day, where the previous statements but also as a tool to prove that this
maintenance actions can be performed. One in the beginning of system might be applied to layouts different than the one
the day and other in the middle. The metrics used to assess the presented in Fig. 1 where all stages are fully-connected. It
performance of the system are the variation of pieces produced may be applied, for example, to layouts where the stages
in relation to target, named as differential, and an extension have different number of machines. Also, it demonstrates that
of availability per machine [20] to the whole system defined failures related to transportation equipment can be considered

42
as long as the failure predictions are fed to the algorithm in system can still comply in those situations. Four different
order to trigger the optimization module. configurations were tested and Table III summarizes all the
In Table I, the effects of maintenances and connection inter- scenarios as well the effects of number of maintenances in
ruptions during normal operation without optimization module the system without the optimization module. It was decided
are represented and summarized. Expected Production is the to increase the number of maintenances as the configurations
number of pieces produced by the system if no disturbances in increase in size in order to test similar levels of stress. This
the system occur. Pieces produced are the pieces that system increase, in Table III, is referred as ”Number of maintenances”.
manufactured under the conditions explained previously for Expected Production and Pieces Produced, as well as, differ-
each test without the intervention of the Prescriptive System. ential and availability, have the same meaning as the presented
Also, differential and availability are the metrics previously in Table I. Each configuration has two different targets as they
explained taking into consideration that no optimization took correspond two each type as stated before.
place.
TABLE III
TABLE I S CENARIO DEFINITION OF THE SECOND TESTING SET
E FFECTS OF FAILURES IN THE SYSTEM WITHOUT OPTIMIZATION MODULE
Number Pieces
Expected
Expected Pieces of Produced Availa- Test
Differential Availability Config. Produ- Target
Production Produced mainte- (Diffe- bility name
ction
Test 1 796 731 -8,16% 96,3% nances rential)
Test 2 796 698 -12,31% 94,4% 1113 97,5% 1194 Test1a
3x3 1 1194
Test 3 796 607 -23,74% 92,0% (-6,78%) 97,5% 1433 Test1b
1412 97,9% 1532 Test2a
4x4 2 1532
(-7,83%) 97,9% 1838 Test2b
1490 97,5% 1554 Test3a
B. First Set Results 7x7 5 1554
(-4,12%) 97,5% 1865 Test3b
Each test was executed three times. In Table II, the averages 10x10 8 2030
1954 98,3% 2030 Test4a
of these three runs are presented, together with standard (-3,14%) 98,3% 2436 Test4b
deviation, σ, of differentials.
D. Second Set Results
TABLE II
R ESULTS OF FIRST TESTING SET WITH OPTIMIZATION MODULE
In all tests the target was achieved within 1% margin and,
in some cases, the availability slightly increased. Those cases
Pieces Processing are marked in bold in Tables IV and V. In these instances, the
Differential σ Availability
Produced Times
Test 1 796 -0,044% 0,259% 96,3% 4,27h
increase in availability was because the algorithm “pushed”
Test 2 795 -0,084% 0,258% 94,4% 7,9h some failures to next week as a result of a reduction in
Test 3 796 0% 0,000% 92,0% 2,87h the throughputs of the respective machines. In addition, this
happened in higher order configurations, which indicates that
Recalling the conditions the test 1 was under, one of the is likely due to the higher redundancy in these systems.
possible outcomes could be the advancement of the failure
that was scheduled to the beginning of the following week. TABLE IV
However, this did not happen. In the second test, two optimiza- R ESULTS FOR TESTS TYPE A
tion moments occurred, one per each failure. This is further Pieces Processing
supported by the fact that in both cases the availability did not Differential σ Availability
Produced Times
change, which means that the downtime neither increased or Test1a 1193 0% 0,181% 97,5% 3,0h
Test2a 1533 0,13% 0,134% 97,9% 8,7h
decreased. In the third test, it is confirmed that the system can
Test3a 1554 0% 0,273% 98,0% 30,9h
handle other types of situations and/or layouts. In this case, Test4a 2024 -0,279% 0,203% 98,7% 71,3h
both differential and standard deviation are 0% because in all
three runs the weekly target was scrupulously achieved.
TABLE V
C. Second Set Scenarios R ESULTS FOR TESTS TYPE B
Previous tests showed that the system behaves as expected
Pieces Processing
so scenarios that are more complex were tested in order to in- Produced
Differential σ Availability
Times
vestigate the scalability of the system. For each configurations Test1b 1434 0,07% 0,057% 97,5% 3,1h
tested, two types of situations were considered: Test2b 1838 0,108% 0,112% 97,9% 9,7h
Test3b 1864 -0,018% 0,241% 97,8% 29,5h
• Type A - weekly production target equal to expected Test4b 2438 0,096% 0,102% 98,5% 77,3h
production;
• Type B - weekly production target 1,2 times higher than Still, in respect to the increase in availability, the com-
expected production. parison between Fig. 7 with Fig. 8 gives an insight of how
The purpose of type B tests is to explore situations where the algorithm dealt with the different maintenance actions.
market demand increases and verify if the manufacturing These figures are related to Run 1 of test4b and its results

43
can be found in Table VI. In Fig. 8, maintenance regarding To evaluate how the results vary from configuration to
machines J5 and G7 disappeared from the current week and configuration in order to draw some conclusions, the averages
the throughputs in those machines are, in general, lower than of the differential were plotted and the graphs are presented in
baseline. This is consistent with Equation (1) as maintenances Fig. 9 and Fig. 10, tests type A and tests type B, respectively.
in next week, F nw are less penalizing than current week and
the algorithm found a way of decreasing the fitness value by
pushing the maintenance to next week without jeopardizing
the achievement of the weekly target. In addition, considering
once more Fig. 8, maintenance regarding machine G8 was
advanced in relation to Fig. 7 however, this advancement trans-
lated into a scheduled maintenance instead of an emergency
maintenance which is also consistent with the fitness function
as emergency maintenances, F em , are more penalizing than
scheduled maintenances, F sm .

Fig. 9. Differential Averages per Configuration in tests type A

Fig. 7. Part of layout of configuration 10x10. Simulation correspondent to
Run1 of test4b where no optimization took place. The red vertical bands
represent the time that a machine is under maintenance and the blue horizontal Fig. 10. Differential Averages per Configuration in tests type B
lines are the throughput rates in place during certain day.

IV. D ISCUSSION
The results show large improvements in the pieces differ-
ential and, in some instances, a slight increase in availability.
Despite the decrease in differential, in some instances, the
target value was not fully met, presenting low deficits (<1%),
but always by far better than the results without optimization.
The parameters of the GA are problem dependent. In the GA
implementation employed in this system, both generations and
population size are fixed. However the size of each chromo-
some is not. Remembering previous sections, the chromosome
size is equal to N ×d where N is the total number of machines
and d, the days from the point the optimizer was triggered
until the end of the week. So, not only between different
configurations but also within configurations, the chromosome
Fig. 8. Part of layout of configuration 10x10. Simulation correspondent to size varies but the parameters are not recalculated. This could
Run1 of test4b where the measures recommended by the Prescriptive System
were adopted.The red vertical bands represent the time that a machine is under led to believe that the algorithm when applied to bigger
maintenance and the blue horizontal lines are the throughput rates in place configurations would generate worse solutions.
during certain day. When comparing the averages of each configuration, the
desired results are that they gravitate towards zero with low
deviations. The solutions seem to follow this behaviour, how-
TABLE VI ever, there is a visible increase in deviation from configuration
R ESULTS OF RUN 1 OF TEST 4 B
3 to configuration 4, Fig. 9, in tests of type A but it did
Target Pieces Produced Differential Availability not go beyond 1%. In fact, this corresponds to an average
2436 2441 (+5) 0,205 % 98,8 % deviation of 0,279% as can be observed in Table IV. Therefore,

44
this increase does not seem enough to jeopardize the results [2] Y. Koren, X. Gu, and W. Guo, “Reconfigurable manufacturing systems:
regarding the tested configurations, and can be attributed to Principles, design, and future trends,” Front. Mech. Eng., vol. 13, no. 2,
pp. 121–136, 2018.
the search strategy and convergence of the GA. However, [3] G. W. Vogl, B. A. Weiss, and M. Helu, “A review of diagnostic and
more testing should be conducted. Despite the increase in prognostic capabilities and best practices for manufacturing,” J. Intell.
complexity of the system, the GA model was always able to Manuf., vol. 30, no. 1, pp. 79–95, 2019.
[4] L. Zhang, J. Lin, B. Liu, Z. Zhang, X. Yan, and M. Wei, “A Review on
find solutions with 1% margin. As a matter of fact, the biggest Deep Learning Applications in Prognostics and Health Management,”
differential was a deficit of 0,542% that occurred during Run IEEE Access, vol. 7, pp. 162415–162438, 2019.
1 of Test4a. Additionally, note that the Test 4 refers to a [5] G. Xu et al., “Data-driven fault diagnostics and prognostics for predictive
maintenance: A brief overview,” IEEE Int. Conf. Autom. Sci. Eng., vol.
configuration 10×10 meaning that 100 machines are operating 2019-Augus, no. 1, pp. 103–108, 2019.
which is already a considerable amount of equipment. [6] F. Ansari, R. Glawar, and T. Nemeth, “PriMa: a prescriptive maintenance
model for cyber-physical production systems,” Int. J. Comput. Integr.
V. C ONCLUSION Manuf., vol. 32, no. 4–5, pp. 482–503, 2019.
[7] K. Lepenioti, A. Bousdekis, D. Apostolou, and G. Mentzas, “Prescrip-
A Prescriptive System capable of adapting machines’ tive analytics: Literature review and research challenges,” Int. J. Inf.
Manage., vol. 50, no. April 2019, pp. 57–70, 2020.
throughput depending on variable demand and taking into [8] A. Diez-Olivan, J. Del Ser, D. Galar, and B. Sierra, “Data fusion
consideration pending machine failures was presented. The and machine learning for industrial prognosis: Trends and perspectives
factory is modelled based on graphs theory which allows a towards Industry 4.0,” Inf. Fusion, vol. 50, pp. 92–111, 2019.
[9] F. Zhao, X. Liu, R. Peng, and J. Kang, “Joint optimization of inspection
quick response in layout changes and the throughput sequences and spare ordering policy with multi-level defect information,” Comput.
are managed by a simulation-based GA. The proposed system Ind. Eng., vol. 139, no. 3, p. 106205, 2020.
was evaluated to different layouts and showed consistent [10] K. T. P. Nguyen and K. Medjaher, “A new dynamic predictive mainte-
nance framework using deep learning for failure prognostics,” Reliab.
results among them – the Differential decreased and had a Eng. Syst. Saf., vol. 188, pp. 251–262, 2019.
positive influence in the availability of the system as previously [11] Z. (Max) Yang, D. Djurdjanovic, and J. Ni, “Maintenance scheduling in
stated. manufacturing systems based on predicted machine degradation,” vol.
19, no. 1, pp. 87–98, 2008.
There is no denying that Prescriptive Systems rely heavily [12] Z. Yang, D. Djurdjanovic, and J. Ni, “Maintenance scheduling for a
on each company’s goals and specifications, which leads to manufacturing system of machines with adjustable throughput,” IIE
one of the main reasons for the lack of prescriptive systems in Trans. (Institute Ind. Eng., vol. 39, no. 12, pp. 1111–1125, 2007.
[13] L. Antao, J. Reis, and G. Goncalves, “Continuous Maintenance System
the current literature. In this respect, this paper tries to tackle for Optimal Scheduling Based on Real-Time Machine Monitoring,” in
this gap by implementing both manufacturing simulation en- IEEE International Conference on Emerging Technologies and Factory
vironment and optimization module in a considerably generic Automation, ETFA, 2018, vol. 2018-Septe, pp. 410–417.
[14] B. Çaliş and S. Bulkan, “A research survey: review of AI solution
manner. Thus, regarding the optimization module, despite the strategies of job shop scheduling problem,” J. Intell. Manuf., vol. 26,
objectives being already established they were chosen in order no. 5, pp. 961–973, 2015.
to be suitable to any manufacturing industry. [15] Y. Koren and M. Shpitalni, “Design of reconfigurable manufacturing
systems,” J. Manuf. Syst., vol. 29, no. 4, pp. 130–141, 2010.
The proposed system was developed with a future inte- [16] S. Aguiar, R. Pinto, and G. Gonc, “Life-cycle Approach to Extend
gration with a Predictive Maintenance Module in mind and Equipment Re-use in Flexible Manufacturing,” INTELLI 2016 Fifth
that would be one of the immediate improvements that could Int. Conf. Intell. Syst. Appl. (includes InManEnt 2016) Life-cycle, no.
November, pp. 148–153, 2016.
be done to this system in order to offer a whole cohesive [17] C. Renzi, F. Leali, M. Cavazzuti, and A. O. Andrisano, “A review on
framework that assists in the process of making decisions artificial intelligence applications to the optimal design of dedicated and
based on constant monitorization of the machines health reconfigurable manufacturing systems,” Int. J. Adv. Manuf. Technol.,
vol. 72, no. 1–4, pp. 403–418, 2014.
status. Also, further research should be conducted not only [18] S. N. Mirabedini and H. Iranmanesh, “A scheduling model for serial
by increasing the number of runs per tests but also explore jobs on parallel machines with different preventive maintenance (PM),”
how the system performs with real-data. In addition, other Int. J. Adv. Manuf. Technol., vol. 70, no. 9–12, pp. 1579–1589, 2014.
[19] K. Jebari and M. Madiafi, “Selection Methods for Genetic Algorithms,”
limitation that needs to be addressed are the long processing Int. J. Emerg. Sci., vol. 3, no. 4, pp. 333–344, 2013.
times needed which are a huge restriction when applied to real- [20] S. H. Huang et al., “Manufacturing productivity improvement using
life scenarios. In this case, the exploration of distributed or effectiveness metrics and simulation analysis,” Int. J. Prod. Res., vol.
41, no. 3, pp. 513–527, 2003.
parallel GA approaches can help to overcame this constraint.

ACKNOWLEDGMENT
This paper is integrated in the project INDTECH 4.0 –
New technologies for intelligent manufacturing. Support on
behalf of IS for Technological Research and Development
(SI à Investigação e Desenvolvimento Tecnológico). POCI-
01-0247-FEDER-026653.

R EFERENCES
[1] Z. M. Bi, S. Y. T. Lang, W. Shen, and L. Wang, “Reconfigurable
manufacturing systems: The state of the art,” Int. J. Prod. Res., vol.
46, no. 4, pp. 967–992, 2008.