You are on page 1of 21

International Journal of Production Research

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tprs20

A deep learning approach for integrated


production planning and predictive maintenance

Hassan Dehghan Shoorkand, Mustapha Nourelfath & Adnène Hajji

To cite this article: Hassan Dehghan Shoorkand, Mustapha Nourelfath & Adnène Hajji
(2023) A deep learning approach for integrated production planning and predictive
maintenance, International Journal of Production Research, 61:23, 7972-7991, DOI:
10.1080/00207543.2022.2162618

To link to this article: https://doi.org/10.1080/00207543.2022.2162618

Published online: 24 Jan 2023.

Submit your article to this journal

Article views: 664

View related articles

View Crossmark data

Citing articles: 2 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tprs20
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
2023, VOL. 61, NO. 23, 7972–7991
https://doi.org/10.1080/00207543.2022.2162618

A deep learning approach for integrated production planning and predictive


maintenance
Hassan Dehghan Shoorkanda,c , Mustapha Nourelfatha,c and Adnène Hajjib,c
a Department of Mechanical Engineering, Laval University, Quebec, Canada; b Department of Operations and Decision Systems, Laval University,
Quebec, Canada; c Interuniversity Research Center on Enterprise Networks, Logistics and Transportation (CIRRELT), Quebec, Canada

ABSTRACT ARTICLE HISTORY


This paper considers a multi-period multi-product capacitated lot-sizing problem. It develops an Received 12 August 2021
integrated predictive maintenance and production planning framework using deep learning and Accepted 2 November 2022
mathematical programming. The objective is to minimise the sum of maintenance, setup, holding, KEYWORDS
backorder, and production costs, while satisfying the demand for all products over the horizon under Deep learning; production
consideration. Based on a rolling horizon approach, the model dynamically integrates data-driven planning; rolling horizon;
predictive maintenance and production planning. The used maintenance policy includes replace- predictive maintenance;
ments and minimal repairs that are considered as preventive and corrective maintenance, respec- data-driven approach
tively. To select preventive maintenance actions, a long short-term memory model is employed to
accurately predict the health condition of the machine. Each rolling horizon consists of ordinary and
forecast stages, and by collecting new sensor data, the maintenance and production decisions are
simultaneously updated. The resulting integrated framework is validated using a benchmarking data
set. The results are compared for different approaches to highlight the advantages of the proposed
framework.

Acronyms w0 Pre-determined time window


Wt Upper bound of Qi,t
LSTM Long Short-Term Memory
PF,tr Failure probability in period t r
PM Preventive Maintenance
PF,tr +k Failure probability in period t r + k
PdM Predictive Maintenance
Atr Available capacity after maintenance for period
CM Corrective Maintenance
tr
RUL Remaining Useful Life
Atr +k Available capacity after maintenance for period
RHP Rolling Horizon Planning
tr + k
TPM Total Productive Maintenance
Gtr Average production rate of the machine during
period t r
Parameters Gtr +k Average production rate of the machine during
i Number of products, {i = 1, 2, . . . , P} period t r + k
t Number of periods, {t = 1, 2, . . . , T} CR Cost of PM (replacing)
T Total planning horizon TR Time of PM (replacing)
L Length of each period CCM Cost of CM (minimal repair)
r Index of rolling horizons TCM Time of CM (minimal repair)
k Index of periods in forecast stage in rolling hori- πi Cost of production for product i
zon r hi Cost of inventory holding for product i
tr Index of the ordinary stage in rolling horizon r bi Cost of backorder for product i
tr + k Index of the forecast stage in rolling horizon r si Cost of setup for product i
u Number of remaining periods xl Input of LSTM at time l
di,t Demand of product i in period t hl Output of LSTM at time l
g Nominal production rate of the machine Cl Cell state of LSTM at time l
(items/month) fl Forget gate of LSTM at time l

CONTACT Hassan Dehghan Shoorkand hassan.dehghan-shoorkand.1@ulaval.ca Departement of Mechanical Engineering, Pavillon Adrien-Pouliot,
local 1314A, Université Laval, Quebec, Canada (G1V 0A6)

© 2023 Informa UK Limited, trading as Taylor & Francis Group


INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7973

ol Output gate of LSTM at time l production planning can be a cost-effective way to ensure
il Input gate of LSTM at time l more efficiency as claimed by many research studies in
N Number of nods in each LSTM unit recent years. In the last decades, many researches have
been done to take advantage of integrated models for
production problem and PM for manufacturing systems.
Decision variables
These studies generally integrated PM and production
Qi,tr Quantity of product i which should be pro- planning to evaluate the impact on the total cost. More
duced during ordinary stage t r recently, some studies have extended these integrated
Qi,tr +k Quantity of produced product i which should models to deal with the capacitated lot-sizing problem
be produced during forecast stages t r + k (Beheshti Fakher, Nourelfath, and Gendreau 2017).
Bi,tr Quantity of backorder during ordinary stage On the other hand, due to the recent develop-
tr ment of sensor technologies, increasing research efforts
Bi,tr +k Quantity of backorder during forecast stages have been made on predictive maintenance (PdM).
tr + k PdM approaches are recognised as an advanced type of
Ii,tr Quantity of inventory holding during ordi- planned maintenance that is preventing machine fail-
nary stage t r ure, by analysing the real-time condition and sensory
Ii,tr +k Quantity of inventory holding for remaining data to identify patterns and predict issues before they
periods in each rolling horizon for product i occur. Generally, PdM models are categorised into two
Seti,tr Binary variable equals 1 if machine is setup groups: model-based and data-driven approaches. To
to produce product i in ordinary stage t r , and estimate any damage caused by a certain failure mecha-
otherwise 0 nism, a model-based approach requires a combination of
Seti,tr +k Binary variable equals 1 if machine is setup to experiment, observation, geometry, and data condition
produce product i in stage t r + k, and other- monitoring (Okoh, Roy, and Mehnen 2017). As the per-
wise 0 formance of these approaches relies heavily on the prior
mR,tr Binary variable equals 1 if replacing is per- statistical knowledge of the system’s lifetime, formulating
formed at the beginning of the period t r , and the real deterioration mechanism of a complex system is
otherwise 0 not easy. On the other hand, a simplified model of the real
mR,tr +k Binary variable equals 1 if replacing is per- working conditions of a system, such as considering sim-
formed at the beginning of the period t r + k, ilar components with the same degradation behaviour,
and otherwise 0 could yield less efficient maintenance decisions (Nguyen
nR Binary variable that is 1 if replacing is per- and Medjaher 2019). Therefore, employing the statistical
formed during the current period, and other- method to estimate the failure probability might cause
wise 0 the real condition of the system to be ignored. Unlike
nCM Binary variable that is 1 if CM is performed model-based approaches, data-driven prognostic tech-
during the current period, and otherwise 0 niques use sensory data to represent the behaviour of
the system’s degradation (Tobon-Mejia et al. 2012). In
real manufacturing environments, the impact of faraway
1. Introduction
unknown events is normally low. As a result of having
Even though maintenance and production planning are access to sensory data after a specific working period,
often performed separately, there is a significant inter- data-driven approaches can be modified with new data.
dependency between these two manufacturing activities. The performance of these models can be determined by
Preventive maintenance (PM) actions are generally car- feature extracting and data processing methods. More
ried out at predetermined intervals to reduce the system recently, deep learning, as the latest research area of arti-
failure probability. Production planning addresses the ficial neural networks, has been used in fault detection.
issue of determining the production lot-sizes of various The conducted literature review shows that the prob-
items, while maintenance is carried out to restore the lem of integrating PdM and multi-period multi-product
production system so as to reach a certain level of reli- capacitated lot-sizing planning is still open. So, there is a
ability. A more reliable system would only be guaranteed need to propose an improved methodology to help deci-
if the equipment is kept in a proper condition. Reach- sion makers with integrated production and PdM plan-
ing this level of reliability requires applying maintenance ning based on the recent technological developments.
actions on the system. A temporary capacity decline can From this point of view, the present paper develops a
occur as a result, and production could be disrupted new method to integrate PdM and medium-term tacti-
(Bajestani 2014). In this case, integrated maintenance and cal production planning in a multi-product production
7974 H. DEHGHAN SHOORKAND ET AL.

system, by considering a rolling horizon approach to feasibility constraint. They showed that maintenance had
take advantage of dynamic decisions. Deep learning algo- an impact on production capacity and machine availabil-
rithms are used to predict the health class of the machine, ity. Yildirim and Nezami (2014) analysed an integrated
and the probability of each class membership based on production and PM planning for a single-component sys-
different time windows. Each time window reflects the tem where system degradation had an effect on product
length of the next period since the system is consid- processing times and energy consumption with aim of
ered operating during the whole period. The integrated minimising the total incurred production cost. Hnaien
predictive maintenance and production planning frame- et al. (2016) integrated production planning and PM for
work can provide accurate information to avoid impend- a single machine by minimising the total production, PM,
ing failures and the reduction of system availability. The minimal repair, set-up, and backorder costs. Beheshti
developed mixed-integer programming model is solved Fakher, Nourelfath, and Gendreau (2017) suggested an
by CPLEX. integrated model of the production capacitated lot-sizing
The remainder of the paper is arranged as follows. problem with numerous products and PM planning for
Section 2 includes a literature review to better high- a multi-machine system, where the objective function
light the paper’s contribution in comparison to exist- minimises the total cost. They linked the machine age
ing papers. Section 3 presents the proposed dynamic to the production capacity and system reliability. Arani
integrated predictive maintenance and production plan- et al. (2020) considered the problem of the integrated pro-
ning. Section 4 describes the data set used to validate duction process and maintenance planning on a machine
the model. Numerical examples are given in Section 5 which is subject to random failure. The purpose of this
to illustrate the advantages of the proposed framework. paper was to minimise the cost of production accom-
Section 6 concludes the paper. panying PM, the expected CM, and production cost.
To solve the problem, a mixed-integer linear program-
ming model was developed. Bahria et al. (2020) proposed
2. Literature review
new production and maintenance strategies for a produc-
The advantages of joint planning have been widely tion system composed of a single machine, which must
reported in the literature. Two main literature groups continuously satisfy demand and may randomly shift to
are discussed in this section. The first one, which con- an ‘out-of-control’ state due to equipment degradation.
stitutes a major part of the existing integrated models, Alimian, Ghezavati, and Tavakkoli-Moghaddam (2020)
considers model-based approaches for PM and PdM. The presented an integrated production and PM planning for
second group corresponds to the growing research using a dynamic cellular manufacturing system to reduce the
data-driven approaches for PdM. number of sudden failures which leads to a lower total
cost. Kaddachi, Gharbi, and Kenné (2022) addressed a
production planning and maintenance control problem
2.1. Model-based approaches
for unreliable manufacturing systems, which combines
Cassady and Kutanoglu (2005) integrated PM with a production control policy of a feedback nature with
single-machine scheduling planning by minimising the multiple hedging levels and, an age-based replacement
total expected weighted completion time of jobs. They policy for maintenance control. They aimed to develop an
used a linear relationship between the PM cost and integrated production and maintenance control policy to
the improvement of the machine condition. Aghezzaf, minimise the total production cost. Pourghader Chobar
Jamali, and Ait-Kadi (2007) introduced an integrated (2022) developed a mathematical model that integrates
model for the capacitated lot-sizing problem and an production, PM, and human resources. They considered
interval-based PM plan for a single-component system. personnel skills, equipment usage rate and equipment
They have shown the economic effectiveness of such an failure rate as the model parameters. For a two-stage dete-
integrated model. Wang and Liu (2013) included the riorating manufacturing systems, Shao, Chen, and Sarker
concept of economic production quantity in the inte- (2022) considered a joint decision-making issue of pro-
grated production and maintenance planning, where duction and maintenance plan with imperfect and perfect
the optimal number of inspections, the economic pro- maintenances. They employed a model-based approach
duction quantity, and the PM level should be dynami- to analyse the costs of the perfect maintenance as well
cally determined. They demonstrated that the integrated as the risk of random failures for each stage. The results
model has a significant impact on economic produc- indicated that the proposed policy and the model are
tion quantity results. Géhan, Castanier, and Lemoine feasible and effective.
(2014) developed an integrated production and mainte- Recently, due to the development of sensor technolo-
nance tactical planning model for a single system under a gies, PdM has become even more attractive in several
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7975

other fields where the safety, reliability, and availability dispatching rules, the results indicate improvements in
of systems need to be critically considered. Few stud- key performance indicators. Due to the current evolu-
ies have been conducted to integrate PdM and pro- tion of technologies and maintenance strategies in the
duction planning. Pan, Liao, and Xi (2012) proposed industrial field, Arena et al. (2022) developed an inte-
a mathematical model to minimise the maximum tar- grated model for maintenance and production problems
diness of a single machine where scheduling produc- by combining the knowledge of the data collected from
tion and PdM based on machine remaining useful life physical assets for PdM management. They took advan-
(RUL) prediction would be scheduled at the same time. tage of the model-based approach to estimate the system’s
Fitouri et al. (2016) proposed a heuristic model to solve a RUL, while maintenance is performed based on this RUL.
job shop production and PdM problem while minimis- Most of the current existing models are based on
ing the make span and the total cost of maintenance. stochastically modelling the system degradation process
They used prognostic information to deal with this prob- using model-based approaches, which take advantage of
lem. Liu, Dong, and Chen (2018) proposed an inte- statistical methods.
grated model that considers PdM planning for a single Nevertheless, the prior statistical knowledge of the
machine, based on the health status and dummy age with system’s lifetime has a significant impact on the effective-
respect to machine degradation, in order to minimise the ness of model-based approaches. Therefore, developing
total expected cost. Bougacha et al. (2019) represented a a theoretical model for the real deterioration process
health management technique in an integrated approach, of a complex system can be highly difficult. In addi-
based on the prognostic information of the component to tion, admitting a theoretical model is developed, it can-
reduce the conflict between production and maintenance not be systematically used in other industrial contexts
services. Khatab et al. (2019) proposed an integrated under different operation condition changes. For exam-
model for production and condition-based maintenance ple, variations in loads over time can make the proposed
in a stochastically deteriorating single-product, single- model no longer valid. However, simplifying a system’s
machine production system to determine the optimal real working conditions can result in poor maintenance
inspection cycle, degradation threshold level, and which decisions (Hesabi, Nourelfath, and Hajji 2022). Recently,
maintenance should be carried out, while minimising data-driven approaches have been developed to over-
the total cost. Wang, Lu, and Ren (2020) presented an come such problems of model-based approaches.
integrated model for production planning and condition-
based maintenance under uncertainty of demand and of
2.2. Data-driven approaches
failure by considering the limited capacity for the produc-
tion system. Ghaleb et al. (2020) tackled an integrated More recently, due to the rapid development of criti-
production scheduling and PdM for a single-machine cal technologies, machine learning and deep learning
production environment. They considered that machine algorithms have attracted more attention to be used as
deterioration and failures affect processing times and a fault detection method through real-time condition
energy consumption. Paprocka, Kempa, and Skołud monitoring. A substantial literature has recently grown
(2021) developed an integrated model of production around data-driven maintenance approaches employed
scheduling and PdM to make the production system on production systems to improve system availability and
insensitive to disruption. Their goal was to find the opti- reliability while reducing production losses. The papers
mal predictive schedule for multiple factors, including reviewed in this part have considered PdM based on
makespan, total tardiness, flow time, and machine idle data-driven approaches.
time. Han et al. (2021) proposed a RUL prediction model To investigate the impact of the imperfect prognostics
based on PdM to improve the mission reliability of the on maintenance decision, Nguyen and Medjaher (2019)
manufacturing system by considering the components’ proposed a dynamic data-driven approach by consider-
dependence for product quality requirements. El Cadi ing long short-term memory (LSTM) method as a PdM
et al. (2021) proposed an integrated model for production model for a manufacturing system. They highlighted
and PdM control of manufacturing systems by consider- the efficiency and performance of their proposed model
ing operation-dependency for both reliability and quality. comapred to the classical periodic PM policy. Morariu
The main purpose was to minimise the total cost by et al. (2020) considered a hybrid control solution that uses
optimising the production and maintenance control set- Big Data techniques and machine learning algorithms to
tings simultaneously. Triska et al. (2021) suggested an process in real-time information streams for a produc-
adaptive method for integrating production and mainte- tion system. Chen et al. (2021) developed a novel main-
nance planning by taking advantage of simulation-based tenance strategy for a repairable complex system by tak-
optimisation. When compared to scenarios with simple ing advantage of data-driven PdM. Appropriate time to
7976 H. DEHGHAN SHOORKAND ET AL.

perform maintenance is determined based on the failure model for a production system that has the ability to be
probability, which has been calculated using the LSTM implemented and integrated to manufacturing decisions,
method. Cui, Wang, and Li (2021) developed a model even in different dynamic loading conditions. The pro-
for a manufacturing system by considering data-driven posed methodology makes it possible to integrate pro-
PdM to improve the product quality and reduce the num- duction and PdM actions with regards to the health con-
ber of unplanned machine failure. They have ignored the dition of the system by monitoring sensor data. The used
production planning part and just considered the main- maintenance policy includes replacements and minimal
tenance part. Bampoula et al. (2021) represented a data- repairs. A long short-term memory model is advanta-
driven approach for maintenance actions to be planned geously used to accurately predict the health condition
according to the actual machine’s condition using the of the machine. The results are compared with different
LSTM model. Ayvaz and Alpay (2021) developed a data- approaches to highlight the advantages of the proposed
driven PdM to monitor manufacturing production lines framework.
by using machine learning methods to take advantage of
real-time sensor data. Yousefi, Tsianikas, and Coit (2021)
took advantage of a deep learning algorithm to provide 3. Dynamic integrated predictive maintenance
a dynamic PdM model for a degrading manufacturing and production planning
system subject to random failures.
3.1. Assumptions
A few studies considered a data-driven approach for
integrated models. Regler (2020) suggested an integrated The following assumptions are considered.
production and data-driven PdM planning model with
random machine failures. The result revealed cost savings • The manufacturing system consists of a single machine.
as compared to the classical method. Zhai, Kandemir, • The classical assumptions of the multi-product capac-
and Reinhart (2022) proposed an integrated production itated lot-sizing problem are satisfied.
scheduling and PdM using the deep learning method as • PM actions restore the machine to an ‘as-good-as-
a data-driven approach. Zonta et al. (2022) suggested a new’ state and are performed at the beginning of some
model to optimise maintenance and production sched- production periods.
ules predictively based on the available data, which gener- • CM actions restore the machine to an ‘as-bad-as-old’
ate the information that allows decision-makers to min- state (minimal repair) and should be carried out as
imise the production cost. To predict the system’s RUL, soon as the failure occurs.
they used a deep neural network method as a data-driven • The planning time horizon initially begins with a new
approach. or as good as new machine.
The above literature review shows that only very few • The manufacturing system is a resource-sharing pro-
papers take advantage of data-driven models to integrate duction system. Due to high set-up costs, when the
maintenance and production planning. The majority of system operates only for a portion of the production
the existing papers ignored the production side of the period, it is assumed that the machine is used for other
problem and is rather concentrated on the maintenance production purposes until the end of the period. For
part. example, the machine can be used by the other depart-
ments of the company for the remaining time. When
this assumption does not hold, it is worth mention-
2.3. The proposed integrated PdM and production
ing that, especially for relatively high demands, the
planning framework
portion of the production period when the system
The above literature review shows that there is no existing is not producing is negligible and will not affect the
paper integrating tactical production planning and pre- availability computation.
dictive maintenance (PdM). To fill this gap, the present • It is assumed that the state probabilities follow a
paper develops an integrated dynamic PdM and pro- steady-state distribution. Regarding this assumption,
duction planning framework based on Rolling Horizon the stationary state will be obtained for each period
Planning (RHP), and using deep learning and mathemat- (Nourelfath and Châtelet 2012). This assumption is
ical programming. Considering a multi-period multi- made regarding a tactical production planning prob-
product capacitated lot-sizing problem, tactical produc- lem since determining items’ quantity is typically con-
tion planning decisions aim to minimise the total cost sidered a tactical planning decision. Tactical planning
associated with integrated production and maintenance. which is a medium-term activity bridges the transition
Deep learning is used to predict the machine health state. from the strategic planning level (long-term) to the
The proposed approach introduces a data-driven PdM operational planning level (short-term). Depending
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7977

on the industry, the time horizons for each planning stage, all information is available, and the LSTM network
level can be various. Typical values for operational utilises this information to predict the health condition of
planning, tactical planning, and strategic planning are the machine. Then, the maintenance plan is made based
one week (or less), one month (or more), and one on the current degradation level. Therefore, the produc-
year (or more), respectively (Nourelfath, Fitouhi, and tion plan in this stage is solved after updating the sys-
Machani 2010). The components of many modern tem availability by subtracting the maintenance time. In
production systems are frequently dependable, and practice, the decision made at this stage is implemented.
PM tasks must be integrated at the tactical level. The second stage is the forecast stage and includes the
remaining periods. Contrary to the ordinary stage, due
to the lack of sensory data, information is expected at
3.2. System description
this stage. A model-based approach is used to determine
Let consider a machine that produces a set of products i ∈ the results, which are not implemented in practice. In this
{1, 2, . . . , P} over a defined finite planning horizon T with case, a tentative maintenance plan is made based on the
t periods, t = {1, 2, . . . , T}. Each period has the same expected degradation level. Then, after updating the sys-
fixed length L. Depending on the production environ- tem availability in the remaining periods by subtracting
ments, the time horizon of tactical planning may vary and the maintenance time. These solutions are only used to
a typical value is one month or more. In the numerical evaluate the performance of the ordinary stage. The pro-
example, the period length is one month corresponding posed model to determine the production plan is solved
to what is typically observed for tactical planning in many by CPLEX. The dynamic RHP approach is performed
industrial contexts. For each product i, during period t, a from one planning horizon to the next one in order to
pre-determined demand di, t should be satisfied. We con- update the information for each period. Figure 1 repre-
sider a multi-period multi-product capacitated lot-sizing sents the detail of implementation of the RHP approach
problem. Maintenance actions can make the machine used in our integrated framework.
‘as good as new’, and they are executed only when the Before beginning the rolling horizon, the LSTM net-
total cost is less than the total cost when ‘doing nothing’. work design is necessary to accurately predict the RUL of
Maintenance actions can be carried out depending on the the machine and the failure probabilities in the different
machine’s health condition. The manufacturing machine time windows in future operations. To this purpose, the
can only be preventively maintained at the beginning of RUL classification considers two classes. The first class
the next period. We also consider that minimal repair indicates that the system’s RUL is greater than a pre-
(as a corrective maintenance) is conducted whenever an determined time window, while the second class shows
unpredicted failure takes place. The objective is to min- that the system’s RUL is less than this pre-determined
imise the sum of maintenance, setup, holding, backorder, time window. The pre-determined time window which is
and production costs, while the demand for all products indicated by w0 is equal to the period length L. As a result,
is satisfied over the entire horizon. by taking advantage of the LSTM method, the probability
To integrate data-driven PdM and production plan- of each class membership can be calculated. The prob-
ning, a long short-term memory (LSTM) model is con- ability for the second class, in contrast to the first one,
sidered. LSTM is a specific recurrent neural network returns the system’s failure probability in the next period.
architecture that is widely used in sequence prediction, The LSTM takes time-series sensor data as inputs to
machine health monitoring, speech recognition, and extract the representations of the system’s health con-
fault detection (Dargan et al. 2019; Namuduri et al. 2020). dition. This data is imported to the LSTM network in
In this paper, LSTM is used in the context of rolling hori- two steps. In the first step, which is the training one, his-
zon planning (RHP). Using RHP is common in dynamic torical data are used to construct a model to learn RUL
environments (Chand, Hsu, and Sethi 2002), and it is classification. In the second one, that is the test step, the
widely investigated and used in both academia and indus- constructed model is used to determine the probabil-
try (de Sampaio, Wollmann, and Vieira 2017). There are ity of each class member. Eventually, after evaluating the
T rolling horizons in the total RHP that each rolling accuracy of the model, it can predict the health condi-
horizon includes t periods. At the beginning of each tion of the system by using new data. This point is the
rolling horizon, new information of the system is avail- beginning of the next period in the ordinary stage. The
able, including quantity of demand, inventory, backorder, new collected data feeds the LSTM, and the RUL classi-
and level of system’s degradation. fication and probability of each class member are com-
There are two stages in each rolling horizon planning puted. At this point, based on the obtained results of the
except of the last rolling horizon. The first stage is an ordi- LSTM, the decision-maker can determine what decisions
nary stage that includes only the current period. At this should be performed. As soon as this decision is made,
7978 H. DEHGHAN SHOORKAND ET AL.

Figure 1. An example of rolling horizon planning.


production capacity is determined for both ordinary and The next sub-section develops the LSTM network
forecast stages. This decision specifies the amount of pro- used to predict the RUL classification and the probability
duction, backorder, and inventory, and the whole process of failure in the next period. Then, Subsection 3.4 inte-
is repeated up to the end of the last period in the rolling grates the PdM and production planning within an opti-
horizon. misation problem so as to minimise the total incurred
The present paper deals with tactical aggregate pro- cost.
duction planning decisions that aim to develop an inte-
grated production and PdM planning model. Produced 3.3. Predictive maintenance model based on the
items at the tactical planning level often relate to a prod- LSTM network
uct family which is defined as a group of finished items
that could be produced by a common manufacturing set- The prognostic information is required for the prediction
up. The whole process of preparing the equipment to of the RUL classification and the probability of failure in
correspond to the next product family is called set-up. To the next period. Fault prognostics can be divided into two
accomplish this, it could be necessary to adjust the equip- main approaches: model-based prognostics and data-
ment or non-adjustable pieces may need to be changed driven prognostics. The model-based approaches use
to accommodate the product family. To have an effec- different probability distributions and design the degra-
tive maintenance plan according to the Total Produc- dation mechanism of the system to predict the RUL.
tive Maintenance approach (TPM), maintenance tasks Designing a model-based approach depends on the avail-
should be considered as components of the production ability of a degradation model. Problems associated with
plan rather than as interruptions to that plan. Regard- cracks by fatigue, corrosion and wear are among the
ing this issue, we assume that machine operators who instances of degradation models (Medjaher and Zer-
are in charge of set-up tasks also carry out PM activities. houni 2013). In contrast to model-based approaches,
The set-up activities are performed at the end of a plan- data-driven prognostics techniques use the collected data
ning period (before the beginning of the next period). As by sensors to extract features and learn the health pat-
a result of sharing common labour and time resources terns to predict the RUL. In recent years, deep learn-
under TPM, it enables the successful integration of PM ing methods have become widely used as a data-driven
tasks into these set-up tasks at the beginning of planning approaches for a time-series data. Accurate RUL pre-
periods. In this case, the time and the cost of PM actions diction plays a critical role in Prognostics and Health
will be significantly lower than interrupting production Management. There are some popular methods in deep
to perform PM tasks during a production cycle due to learning context for prediction. In this context, Zheng
executing PM tasks by machine operators responsible for et al. (2017) compared the accuracy of the convolutional
set-up activities (Nourelfath and Châtelet 2012). neural network, recurrent neural network, and LSTM for
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7979

RUL prediction on time-series data. They show that the data is used to construct a model and to learn RUL clas-
LSTM model’s prediction accuracy is higher than the sification. In the second one, which is the test step, the
other models and gives the best performance in RUL constructed model is used to determine the probability of
prediction. In addition, the result of RUL prediction for each class member. After evaluating the accuracy of the
LSTM and neural network showed that the LSTM model model, it is used to make forecasts for the new collected
outperforms other existing models. The LSTM model data.
can retain important or remove unnecessary information The fundamental idea of LSTM lies in the fact that at
during long-term data sequences due to its memories. each time step, only a handful of gates control the passage
This ability makes LSTM more accurate than the other of information over the sequences, as a result of which
methods where the data is time-series. Since the model long-range dependencies can be captured more accu-
proposed in this paper is a tactical model, the data will rately (Zhao et al. 2016). Figure 2 represents the structure
be long-term sequences. Therefore, taking advantage of of the LSTM unit at time l. A cell state and three gates (a
the ‘sequence-to-sequence’ LSTM method may improve forget, an input, and an output gates) are included in a
the ability of the studied multi-period model to predict typical LSTM unit. The flow of information regulates and
the machine’s condition. manages into and out of the cell state. In the LSTM unit,
LSTM is in fact a specific recurrent neural network xl , hl , and Cl are assumed to present the input, output,
architecture, especially designed to overcome exploding and cell state. Let us assume that there are N nodes in each
and vanishing gradient problems in classical recurrent LSTM unit, therefore, hl ∈ RN×1 and Cl ∈ RN×1 .
neural networks, which typically occur while learning Input gate (il ∈ RN×1 ) determines important informa-
long-term dependencies, even in case of very long min- tion in the current input (xl ) and the output of previous
imal time lags (Hochreiter and Schmidhuber 1996). LSTM unit (hl−1 ) which should be used in computing cell
Instead of neurons, LSTM networks have memory blocks state (Cl ). Then, forget gate (fl ∈ RN×1 ) removes unnec-
that are connected through layers. A block has compo- essary information from the previous cell state (Cl−1 ) and
nents that make it smarter than a classical neuron and a updates the current cell state (Cl ) based on the retained
memory for recent sequences. A block contains gates that information in Cl−1 . In the last step, the information
manage the block’s state and output. A block operates from cell state (Cl ) that should become the output of the
upon an input sequence and each gate within a block uses present LSTM unit (hl ) is determined by the output gate
the sigmoid activation units to control whether they are (ol ∈ RN×1 ). The formulas below are used to calculate the
triggered or not, making the change of state and addition aforementioned gates.
of information flowing through the block conditional.
fl = σ (Wf xl + Rf hl−1 + bf ) (1)
The cell remembers values over arbitrary time intervals
and the three gates regulate the flow of information into
and out of the cell. LSTM networks are well-suited to clas- il = σ (Wi xl + Ri hl−1 + bi ) (2)
sifying, processing and making predictions based on a
time-series. It is well-established that LSTM has a great ol = σ (Wo xl + Ro hl−1 + bo ) (3)
capacity to retain memory and learn from data sequences
where σ is a nonlinear activation function.
(Aydin and Guldamlasioglu 2017).
Usually, the sigmoid function is employed as an activa-
In this study, we use the LSTM model in ordinary
tion function for the gates in LSTM; because it produces a
stages to predict the RUL classification and the proba-
value between 0 and 1, there can be either no flow or com-
bility of failure in the next period. In the first ordinary
plete flow of information throughout the gates. Equation
stage, as the sensory data is not available, a model-
(4) represents it:
based approach is used. In the LSTM model, the machine
performs the next period only when the RUL of the 1
sigmoid(x) = (4)
machine is greater than the pre-determined threshold 1 + e−x
w0 (Ptr (RUL > w0 )). The incurred total cost consists of l is generated
Inside the LSTM, the intermediate state C
the total production, inventory, backorder, set-up, and as:
maintenance costs. The maintenance decision made is
based on the comparison between the total costs of all l = ϕ(WC xl + RC hl−1 + bC )
C (5)
alternatives, to select the lower cost alternative.
where ϕ is a tanh which is used as a nonlinear activation
The LSTM takes time-series sensors data as inputs to
function
extract the representations of the machine’s health con-
dition. This data feeds the LSTM network in two steps. ex − e−x
tanh(x) = (6)
Over the first step, which is the training step, historical ex + e−x
7980 H. DEHGHAN SHOORKAND ET AL.

Figure 2. LSTM cell.

To overcome the problem of vanishing gradient, we Table 1. Confusion matrix for binary classification.
need a method whose second derivative may last for a Predicted RUL
long-range before vanishing. tanh is a useful function that Class 1 Class 0
possesses all of the aforementioned characteristics.
Actual RUL Class 1 True Positive False Negative
Then, the memory cell and hidden state of this LSTM Class 0 False Positive True Negative
are updated as:

l
Cl = fl ∗ Cl−1 + il ∗ C (7) The testing data set is used with the aim of validation
to yield an unbiased estimate of accuracy over the learn-
hl = ol ∗ ϕCl (8) ing process. The most common performance metrics to
evaluate the binary classification are confusion matrix
where ∗ denotes the element-wise product.
and accuracy. These metrics are explained as follows.
During the training process, the bias vector parame-
ters bf , bi , bo , bC ∈ RN×1 as well as the weights matrices
3.3.1. Confusion matrix
Wf , Wi , Wo , WC ∈ RN×k should be computed. As the This is a table that illustrates the performance of classi-
LSTM model is responsible to predict the machine’s state, fication on a set of test data where the true values are
its error should be computed throughout this process known. Table 1 shows this concept.
using a loss function. Regarding our binary classifica-
tion model in this paper, the average difference between 3.3.2. Accuracy
the actual and predicted classes is calculated by using a This is the percentage of correctly classified records to
binary cross-entropy loss function as follows: total records. If the false positives and false negatives have
similar costs, the best accuracy will be achieved.
1 
N
J= − yi . log(p(yi )) + (1 − yi ).log(1 − p(yi )) TP + TN
N Accuracy = (10)
i=1
TP + TN + FP + FN
(9)
where TP and FP represent the number of true posi-
To minimise the error of the next evaluation, the esti- tives and false positives, respectively. On the other hand,
mated weights and biases are updated as the LSTM model FN and TN show the number of false negatives and true
is responsible to make a prediction. Based on the best negatives.
score for cross-entropy which is 0, the error is minimised The input layer is used to import data to be processed
to achieve this number. in the network, while the hidden layer attempts to build
For each observation i, we use yi as the real machine’s the relation between the input and output layers. In this
state, and p(yi ) as the probability of the predicted paper, two hidden layers are utilised. Eventually, the out-
machine’s state. put layer is used as a prototype between the network and
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7981

the output, allowing transforming multiple multivariate Therefore, at the end of the planning horizon, all imple-
data at the hidden layer to a one-dimensional array at the mented decisions will be based on health state informa-
LSTM classifier output (Nguyen and Medjaher 2019). tion RUL predictions, based on a real sequence of sensory
data collected from the past period as the rolling hori-
zon moves. As indicated before, the LSTM network is
3.4. Integrated model
a sequence-to-sequence model, which is well-suited for
The proposed dynamic PdM and production planning in predictive maintenance (PdM) in the context of multi-
each rolling horizon consists of a dynamic LSTM deci- ple period planning problems. This characteristic was
sion in the ordinary stage, and model-based planning advantageously exploited to develop the integrated PdM
for the forecast stages, except for the first rolling horizon and tactical production planning model proposed in this
where the sensory data are not available. In fact, a model- paper.
based approach is first used to plan for the entire rolling Two maintenance options are considered in the pro-
horizon; then, in the other rolling horizons, decisions are posed model as a PM action: replacing and doing noth-
updated at the beginning of each ordinary stage, as the ing. To find the best planning strategy, the total cost of
new sensory data feeds the LSTM model and the health each one should be evaluated. Between different possible
condition of the machine can be predicted. That is, by alternatives, we choose the one which has the minimum
starting a new rolling horizon, information on degrada- total cost in both ordinary and forecast stages.
tion, inventory and backorder levels will be updated for First, the integrated model considered in an ordinary
both stages. According to the proposed rolling horizon stage can be formulated as follows:
planning, there are two planning stages:
p
Expected

(1) An ordinary stage that includes only the current Xt r = (Qi,tr .πi + Ii,tr .hi + Bi,tr .bi + Seti,tr .si )
period. At this stage, the LSTM network predicts the i=1
health condition of the machine based on available + (CR,tr .mR,tr ) + (PF,tr .CCM,tr ) (11)
sensory data.
Qi,tr − Ii,tr + Ii,tr −1 − Bi,tr + Bi,tr −1 = di,tr ; ∀i (12)
(2) The second stage is a forecast stage that includes the
remaining periods. Contrary to the ordinary stage, Qi,tr ≤ Seti,tr .Wt , ∀i (13)
machine health information is expected at this stage  
L − (TR .mR,tr ) − (PF,tr .TCM,tr )
using a model-based approach (Weibull distribution Atr = Max 0, (14)
L
in our numerical example).
Gtr = g.Atr (15)

That is, the results of LSTM and the statistical Weibull iP Qi,tr
distribution are separately employed at different stages. ≤ Gtr (16)
L
As the rolling horizon moves, the RUL class and the Seti,tr ∈ {0, 1}, ∀i (17)
failure probability are predicted by the LSTM for the
next ordinary stage. To solve the whole integrated multi- mR,tr ∈ {0, 1} (18)
period optimisation problem, the statistical distribution
is used only for the forecasting stage periods (but not at The objective function (Equation 11) in the ordinary
Expected
the ordinary stage period, for which the machine health stage is the total cost Xtr , which includes the total
prediction is rather based on LSTM). production, holding, backorder, set up, maintenance, and
For the ordinary stage (next period), the LSTM net- expected failure costs. In this equation, the PF,tr repre-
work predicts the health condition of the machine based sents the failure probability during the current period
on available sensory data from the previous period. and the last term illustrates the expected failure cost.
Therefore, for each ordinary stage, the maintenance is of The decision variables are quantity of production, back-
a predictive nature. Only the forecast stage uses statis- order, inventory (Qi,tr , Bi,tr , and Ii,tr ), and the two binary
tical information to feed the multi-period optimisation variables Seti,tr and mR,tr which express the existence or
model. Once new sensory data are available at the end absence of setup and replacement in the current period,
of the next period, the information is updated and new respectively.
PdM decisions are made. In practice, only the decisions Constraint (12) presents the standard inventory and
made at the beginning of each ordinary stage are imple- backorder balance equation in the ordinary stage t r .
mented. The decisions made for the forecast stage are Equation (13) is a production setup constraint where a
not implemented; they will be updated once new sen- binary variable, Seti,tr is used to force Qi,tr = 0 if Seti,tr =
sory data are available at the end of the next period. 0 and free Qi,tr ≥ 0 if Seti,tr = 1. In Constraint (13), the
7982 H. DEHGHAN SHOORKAND ET AL.

quantity Wt is an upper bound of Qi,tr which can be pro- Gtr +k = g.Atr +k (23)
duced during each period. Equation (14) presents the 
iP Qi,tr +k
steady-state availability, which depends on the mainte- ≤ Gtr +k (24)
nance decision in period t r . If the machine’s nominal L
production rate is denoted by g, Constraint (15) rep- Bi,tr = Bi,tr +1 , Ii,tr = Ii,tr +1 ; ∀i (25)
resents the constraint of the average production rate of Seti,tr +k ∈ {0, 1}, ∀i (26)
the machine during a period t r . Equation (16) represents
mR,tr +k ∈ {0, 1} (27)
the constraint of the available production capacity. Con-
straints (17) and (18) represent the binary conditions,
where in (18) mR,tr is used to determine if replacing the The objective function in the forecast stage, Ytr +k ,
Expected
machine is performed. aims to minimise the total cost as presented in Equation
Note that the proposed model covers the situation (19), where mR,tr +k is a binary variable to express the exis-
where the maintenance time is higher than the produc- tence or absence of replacing in the period t r + k. In this
tion period duration. In fact, according to Equation (14), equation, the CM cost is an expected cost which is cal-
when the maintenance time, (TR .mR,tr + PF,tr .TCM,tr ), culated by the failure probability (PF,tr +k ) in period t r +
is higher than the production period duration L, the k. Equation (20) ∼ Equation (24) express the standard
availability is 0 (because the difference between L and inventory balance, production, steady-state availability,
the maintenance time is negative). On the other hand, average production rate, and available production capac-
when the maintenance time is lower than the produc- ity constraints in forecast stages, respectively. Equation
tion period duration, the availability is estimated by (25) presents the quantity of backorder and inventory,
L−(TR .mR,tr )−(PF,tr .TCM,tr )
L . Note however that due to the which are passed from the ordinary stage to the fore-
Max operator, this extension introduces a non-linearity. cast stage. Equation (26) is a binary constraint that is
The resulting optimisation model is therefore a mixed- used to determine if the product i will be produced, so
integer non-linear programming model that is more dif- Seti,tr = 1 or not, Seti,tr = 0. Constraint (27) represents
ficult to solve, for large scale instances, than the mixed- the binary conditions, where mR,tr +k is used to determine
integer linear programming one. if replacing the machine is performed, so mR,tr +k = 1 or
Next, in the forecast stage, the planning for mainte- not, mR,tr +k = 0.
nance actions is performed based on failure probability Eventually, a dynamic integrated production and
distribution. As only the decisions of the ordinary stage maintenance planning for both ordinary and forecast
are implemented, the manager should investigate opti- stages is considered by Equation (28)
mal solutions for the remaining periods, which leads
to an optimisation model for the forecast stage that is
Expected Expected
formulated as follows: Minimise: Zr,Expected = Xtr + Ytr +k ,

 p ∀r, k = 1, . . . , u (28)
Expected

T 
u  Subject to: Constraints (12)−(18) and (20)−(27) (29)
Ytr +k = (Qi,tr +k .πi + Ii,tr +k .hi
t=t r k=1 i=1 Qi,tr , Ii,tr , Bi,tr ≥ 0 and are integers (30)
+ Bi,tr +k .bi + Seti,tr +k .si )

The total cost in each rolling horizon, Zr,Expected , con-
+(C R,tr +k .m R,tr +k ) + (P F,tr +k .C CM,t r +k ) sists of production cost, inventory cost, backorder cost,
set-up cost, expected failure cost, and replacement cost in
(19) both ordinary and forecast stages. As soon as the main-
Qi,tr +k − Ii,tr +k + Ii,tr +k−1 − Bi,tr +k + Bi,tr +k−1 tenance option based on the total cost is selected for
each individual stage, the production capacity for all peri-
= di,,tr ; ∀i, k = 1, . . . , u (20) ods is updated by subtracting the maintenance time. The
Qi,tr +k ≤ Seti,tr +k .Wt ; ∀i, k = 1, . . . , u; (21) expected failure cost is considered to penalise the possi-
⎛ ⎞ ble occurrence of failures. In real conditions, if the failure
L − (TR .mR,tr +k )
⎜ −(PF,tr +k .TCM,tr +k ) ⎟ does not occur, there is no need to assign CM time and
Atr +k = Max ⎜ ⎝0,
⎟;
⎠ cost for each period.
L Considering this part can reduce the total available
time in each period. So, the planning is updated again
∀t, k = 1, . . . , u (22) based on the following equations. The integrated model
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7983

in the ordinary stage is: Bi,tr = Bi,tr +1 , Ii,tr = Ii,tr +1 ; ∀i (45)


Seti,tr +k ∈ {0, 1}, ∀i (46)
p
Planning
 mR,tr +k ∈ {0, 1} (47)
Xt r = (Qi,tr. πi + Ii,tr .hi + Bi,tr .bi + Seti,tr .si )
i=1
Finally, all planning processes in both ordinary and
+ (CR,tr .mR,tr ) (31) forecast stages are performed by minimising the total cost
Qi,tr − Ii,tr + Ii,tr −1 − Bi,tr + Bi,tr −1 as follows:
Planning Expected
= di,tr ; ∀i (32) Minimise : Zr,Planning = Xtr + Ytr +k ,
Qi,tr ≤ Seti,tr .Wt , ∀i (33) ∀r, k = 1, . . . , u (48)
 
L − (TR .mR,tr ) Subject to : Constraints(32)−(38) and (40)−(47) (49)
Atr = Max 0, (34)
L
Qi,tr , Ii,tr , Bi,tr ≥ 0 and are integers (50)
Gtr = g.Atr (35)
 Qi,tr +k , Ii,tr +k , Bi,tr +k ≥ 0 and are integers (51)
iP Qi,tr
≤ Gtr (36) The actual cost in the RHP is the sum of all actual costs
L
Seti,tr ∈ {0, 1}, ∀i (37) of each ordinary stage. As the decisions that are made
in the planning steps and are actually implemented may
mR,tr ∈ {0, 1} (38) cause different results, we use the following equations to
calculate the total actual cost in RHP:
The only difference between expected and planning 
T
steps is considering the expected failure cost in the ZrReal = XtReal
r (52)
first step and assigning CM anytime that is neces- t r =1
sary. So, in the objective function, instead of this term, p

PF,tr +k .CCM,tr +k , we consider CCM,tr . This process will be XtReal
r = (Qi,tr πi + Ii,tr .hi + Bi,tr .bi + Seti,tr .si )
repeated until the last period is reached. i=1
On the other hand, the idea in the forecast stage is + (CR,tr .nR ) + (CCM,tr .nCM ) (53)
calculated the same as the last step.
nR ∈ {0, 1} (54)
 p nCM ∈ {0, 1} (55)
Expected

T 
u 
Ytr +k = (Qi,tr +k .πi + Ii,tr +k .hi
t=t r k=1 i=1 4. Numerical experiments
+B i,tr +k .bi + Set i,tr +k .si ) + (C
R,tr +k .
4.1. Data set

mR,tr +k ) + (PF,tr +k .CCM,tr +k ) (39) In this section, the suggested integrated model is applied
using a benchmarking data set developed by NASA
(Saxena and Geodel 2008). This public data set con-
Qi,tr +k − Ii,tr +k + Ii,tr +k−1 − Bi,tr +k
sists of multiple run-to-failure data simulated using a
+ Bi,tr +k−1 = di,,tr ; ∀i, k = 1, . . . , u thermo-dynamical simulation model, and includes four
(40) subsets generated under different operational conditions
Qi,tr +k ≤ Seti,tr +k .Wt , ∀i, k = 1, . . . , u (41) and health states. All experiments are performed using
⎛ ⎞ the first subset, FD001. This subset contains a training
L − (TR .mR,tr +k ) and testing set which consists of 26 columns, including
⎜ −(PF,tr +k .TCM,tr +k ) ⎟ the unit numbers (Id), cycle numbers, three operational
Atr +k = Max ⎜ ⎝0,
⎟;

L parameters, and 21 data of various sensors, respectively.
We divided the training data into two parts. The first
∀t, k = 1, . . . , u (42) part, which includes the first five records, is considered
for numerical examples and the others for training the
Gtr +k = g.Atr +k (43) model.

iP Qi,tr +k
The training set is composed of instances with com-
≤ Gtr +k (44) plete run-to-failure data, which can be used to develop
L
7984 H. DEHGHAN SHOORKAND ET AL.

life prediction models. The test set, on the other hand, confusion matrix. The results of the confusion matrix on
is made of instances with data recorded before a spe- the first data set are represented in Figure 3.
cific time prior to system failure, which can be employed In this figure, the number of correctly classified
for RUL estimation and algorithm performance evalua- records is shown in the original diameters. The accuracy
tion. The test instances are modelled as run-to-failure; of our model is 94.63% . This result for the neural net-
although only an earlier part of the history is given in work is 92% based on our data set. The above results show
the data set, the actual lives of the test instances are still that we have a model that better fits the data. Therefore,
known – the ground true RUL of the test instances is also we can safely use the model to make a prediction on new
provided for prediction validation purposes. data in the context of the proposed integrated production
and maintenance optimisation model.

4.2. Running and validating the LSTM model for the 4.3. Running the model-based approach for the
ordinary stage forecast stage
In this study, the proposed models are programmed in Since our framework requires also a model-based
Python and the Keras (Chollet 2015) Deep Learning approach for the forecast stage, we define the best statis-
library with the open-source software library TensorFlow tical method that fits the run-to-failure data set FD001.
(Abadi et al. 2015) as the backend. All our experiments We found that a Weibull distribution fits with this data
are performed on a personal computer with an Intel Core set. Weibull distribution is a two-parameter distribution
i7 (3.3 GHz) CPU, 16 GB memory and mac OS 11 operat- function represented by a dimensionless shape parame-
ing system. The first step of designing the LSTM model is ter β and scale parameter α in units of hours (Chaurasiya,
defining the network configuration. Table 2 presents the Ahmed, and Warudkar 2018). The Weibull density func-
configuration parameters of the LSTM model. tion, f (x) is as follows:
   
The number of epochs is a hyperparameter that con- β  x β−1 x β
trols how many times the learning algorithm runs over f (x) = exp − (56)
α α α
the whole training dataset. The batch size is a hyperpa-
rameter that specifies how many samples must be pro- While several different approaches can be taken to esti-
cessed before the internal model parameters are updated. mate the parameters β and α, the maximum likeli-
Dropout is a basic method of preventing overfitting in the hood estimator (MLE) is commonly used, because of its
model. During the training process, some layer outputs numerical stability (Cohen 1965). The calculated shape
are ignored or ‘dropped out’ randomly. This causes the and scale parameters, by maximising the likelihood func-
layer to look like and behave as a layer with a different tion for our data set, are β = 4 and α = 225. The shape
number of nodes and connections to the previous layer. parameter of the Weibull distribution represents the fail-
Units define the number of neurons in a particular layer. ure rate behaviour; when it is greater than 1, it means the
When this number is higher, the model has more parame- failure rate increases with time, which is known as wear-
ters to update during the learning process. The same may out life. In mechanical components, this is mainly due
be said about layers. to thermal-mechanical fatigue, stress, corrosion, or the
The second phase that needs to be taken into consid- degradation process (Abernethy 2004).
eration is compiling and fitting the network. An optimi- In this example, we consider a machine with a plan-
sation algorithm is used to train the model and a loss ning horizon consisting of 12 months, where the deci-
function is applied to evaluate the network are applied sions are determined monthly at the beginning of each
during the compiling process. In our binary classification, ordinary stage. At this point, new sensory data, which
an Adam optimiser is used to update the network weights are collected during the previous period, feeds the LSTM
and a binary cross-entropy loss function is implemented classifier, and RUL classification and associated proba-
to calculate the model’s error. The weights on training bilities are calculated. Based on the health condition of
data are extracted in the fitting step, and then the net- the machine, maintenance and production planning are
work uses the training data to be determined. The model performed for the ordinary stage at the basis of the deep
is prepared to make a prediction after optimisation. learning LSTM model.
At this point, the test set is used to evaluate the per- As the operating time increases, some functions of
formance of the conducted model during the learning the machine are degraded. So, to make an accurate pre-
process. diction, at each decision point, new sensory data are
As already mentioned, to evaluate the performance of required to evaluate the machine’s lifetime. Since the sen-
our prediction model, we calculated the accuracy, and sory information is available only at the beginning of
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7985

Table 2. Configuration parameters of the LSTM model.


Number of
epochs Batch size Dropout rate 1st layer units 2nd layer units
20 32 0.2 100 50

Figure 3. Confusion matrix for the LSTM model on FD001 data set.

Table 3. Some characteristics of the machine. Table 4. Demands of products.


g (items/month) CCM ($) CR ($) T CM (month) T R (month) Period t 1 2 3 4 5 6 7 8 9 10 11 12
50 2800 800 0.02 0.09 d1,t (items) 22 21 24 25 22 24 20 23 23 22 22 20
d2,t (items) 23 22 20 23 23 22 23 21 20 26 22 23

the ordinary stage, we use the Weibull distribution for


the rest of the periods, which are known as the forecast Table 5. Cost data of products.
stages. In this case, the Weibull distribution is respon-
Product i hi ($) bi ($) si ($) π i ($)
sible for determining the maintenance and production
1 40 160 1000 90
planning for the forecast stages. However, from a practi- 2 40 160 1000 90
cal viewpoint, these decisions are not implemented. This
decision procedure is repeated by collecting new infor-
mation at the beginning of each ordinary stage. Let us
consider a machine with the additional characteristics each product i are constant for all periods and are given
given in Table 3. in Table 5.
In this example, each period of 1 month consists of
50 cycles and the machine should produce two types of
products (i = 1, 2) to satisfy the demands. The informa-
4.4. Results
tion of periodic demands for each product i is given in
Table 4. This Subsection presents the total cost of the introduced
We also assume that the inventory holding cost hi , model. Moreover, to evaluate the performance of this
backorder cost bi , set-up cost si , and production cost πi of model, three other approaches are considered as follows.
7986 H. DEHGHAN SHOORKAND ET AL.

Table 6. Evolution process of production in each rolling horizon.


Period
1 2 3 4 5 6 7 8 9 10 11 12
Rolling 1 Q(22, 23) Q (21, 22) Q (24, 20) Q (25, 23) Q (22, 23) Q (24, 22) Q (20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
2 – Q(21, 22) Q (24, 20) Q (25, 23) Q (22, 23) Q (24, 22) Q (20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
3 – – Q(27, 20) Q (22, 23) Q (22, 23) Q (24, 22) Q (20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
4 – – – Q(22, 23) Q (22, 23) Q (24, 22) Q (20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
5 – – – – Q(22, 23) Q (24, 22) Q (20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
6 – – – – – Q(24, 22) Q (20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
7 – – – – – – Q(20, 23) Q (23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
8 – – – – – – – Q(23, 21) Q (23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
9 – – – – – – – – Q(23, 20) Q (22, 26) Q (22, 22) Q (20, 23)
10 – – – – – – – – – Q(22, 26) Q (22, 22) Q (20, 23)
11 – – – – – – – – – – Q(22, 22) Q (20, 23)
12 – – – – – – – – – – – Q(20, 23)

Table 7. Optimal decisions at each ordinary stage of the proposed model.


Production decisions for Product 1 Production decisions for Product 2
Period Maintenance decisions Production Inventory Backorder Set-up Production Inventory Backorder Set-up
1 0 22 0 0 1 23 0 0 1
2 0 21 0 0 1 22 0 0 1
3 0 27 3 0 1 20 0 0 1
4 1 22 0 0 1 23 0 0 1
5 0 22 0 0 1 23 0 0 1
6 0 24 0 0 1 22 0 0 1
7 0 20 0 0 1 23 0 0 1
8 0 23 0 0 1 21 0 0 1
9 1 23 0 0 1 20 0 0 1
10 0 22 0 0 1 26 0 0 1
11 0 22 0 0 1 22 0 0 1
12 1 20 0 0 1 23 0 0 1

4.4.1. Approach 1 also these decisions are updated after finishing each ordi-
In the simulated run-to-failure data set, the LSTM clas- nary stage. This corresponds to a pure statistical method
sifier predicts the RUL and associated probabilities. The (using only Weibull distribution).
actual RUL is still known to determine the accuracy of Table 6 represents the obtained results of the inte-
prediction. In this approach, we assume the decision- grated production and maintenance planning in the con-
maker accesses these actual RULs before anything, and text of RHP by minimising the total cost while satisfying
he can accurately plan the maintenance and production the demands for all products in the proposed model. Each
before any failures. So, this is the ideal scenario access row of this Table shows one of the 12 rolling horizons that
that can provide the minimum total cost. In reality, it is consist of 12 periods. Each column presents the quantity
impossible to access the ‘perfect information’ on the exact of product 1 and product 2 for each ordinary and fore-
degradation behaviour of each specific machine to per- cast stages. For example, in the fourth row of Table 6, the
form accurate planning. This scenario is used only as a fourth period is the ordinary stage and periods 5 ∼ 12
reference to evaluate the other scenarios. are forecast stages. Q(22, 23) in the fourth rolling hori-
zon and fourth period presents the optimal quantity of
products 1 and 2 that are 22 and 23 items in the pro-
4.4.2. Approach 2 posed model, respectively. We should consider that only
We assume that the RUL probabilities of forecast stages decisions of the ordinary stage are implemented in prac-
are the same as the ordinary stage. In this case, the whole tice. That is, in practice only the decisions of the ordinary
decision procedure is performed based on the results of stages that are bold in Table 6 are implemented.
the LSTM classifier at the ordinary stage. In the proposed model, the LSTM is used to accurately
predict the RUL of the machine and the probabilities
that it will fail in the time windows w0 that equals the
4.4.3. Approach 3 length of the ordinary stage L. Maintenance actions are
In this case, we only take advantage of the Weibull func- executed only when the total cost when replacing is less
tion for both stages to determine the best PM tasks that than the total cost when doing nothing. Next, the pro-
must be performed at the beginning of each period, and duction capacity is updated in both ordinary and forecast
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7987

Table 8. Total cost comparison results. Table 9. Sensitivity analysis results under demand variation.
Proposed model Approach 1 Approach 2 Approach 3 Variation of Proposed
demand model Approach 1 Approach 2 Approach 3
Total cost $74, 760 $74, 640 $75, 120 $77, 600
−15% $66,040 $66,000 $68,870 $68,040
−10% $68,720 $68,680 $71,430 $70,720
−5% $71,840 $71,680 $73,450 $74,840
stages separately. This process is repeated from one to the Base case $74,760 $74,640 $75,120 $77,600
next planning horizon until the last period is reached (see +5% $77,360 $76,960 $77,640 $80,560
+10% $80,120 $79,960 $84,200 $85,760
Table 6). +15% $98,650 $98,050 $100,450 $106,160
In Approach 1, by considering the access to the per-
fect information, the quantity of production, inventory,
backorder, and set-up for each ordinary stage is already
is, under Approach 2, the whole decision procedure is
planned. So, there is no need to collect new informa-
performed based on the results of the LSTM classifier
tion and analyse it. In Approach 2, we assume the failure
at the ordinary stage. Such superiority of the proposed
probabilities of the machine for forecast stages are the
approach is explained here by the fact that the LSTM pre-
same as the ordinary stage. First, the failure probability
diction period is too large under Approach 2, which affect
of the ordinary stage is predicted by the LSTM classi-
LSTM prediction accuracy.
fier and next, it is generalised for the rest of the periods.
In Approach 3, after each period, if no maintenance is
performed, the age of the machine is increased by one 4.6. Sensitivity analysis
month.
We have 12 periods and 12 rolling horizons. For each Sensitivity analysis is carried out to indicate how small
rolling horizon, there are different PM alternatives. The variations of input parameters can affect the optimal solu-
aim is to select the best PM strategy and production plan tion. In this section, a set of experiments will be con-
minimising the total cost in the integrated optimisation ducted to clarify the sensitivity of the proposed model
model. to the variations of demand. That is, to investigate the
Table 7 shows the obtained results related to mainte- effect of uncertainty, this sensitivity analysis is carried out
nance and production decisions. It is suggested to per- by changing the demand by ±15% of its basic level. The
form replacements at the beginning of periods 4, 9 and obtained results are reported in Table 9.
12. Table 7 also presents the production quantity, inven- Table 9 shows that the proposed model outperforms
tory, backorder, and set-up decision for each ordinary the others in all cases, with the exception of Approach
stage and each product. 1, which is ideal because of its access to perfect informa-
tion. Furthermore, Approach 3 provided the worst results
for high demand, whereas Approach 2 is the worst for
4.5. Comparison results low demand. Figure 4 shows that under all variations,
The total cost of each approach is presented in Table 8. As the results of the proposed approach remain close to
it can be expected, the minimum total cost is associated the results of the reference ideal situation presented by
with Approach 1 as it assumes an access to the perfect Approach 1. In summary, this sensitivity analysis illus-
information. The proposed approach is close to the ideal trates the effectiveness of the proposed approach under
situation presented by Approach 1 (a small difference of different demand variations.
$120). The difference between Approach 2 and Approach
1 is $480; and the total cost of Approach 2 is higher than
5. Conclusion
the proposed model. The difference between Approach
3 and Approach 1, is higher and equals $2960, which is This paper has presented a dynamic integrated model
$2840 more than the proposed model. for tactical production and predictive maintenance plan-
In conclusion, we remark that the proposed approach ning. Based on a rolling horizon approach, the model
provides a solution with a total cost closer to the reference dynamically integrated data-driven predictive mainte-
approach than the other approaches. In fact, as expected, nance and production planning. In the context of a multi-
the proposed approach is better than the pure statistical period multi-product capacitated lot-sizing problem, the
method (Approach 3). It is also better than Approach 2. objective was to minimise the sum of maintenance, setup,
While the proposed approach uses LSTM for the ordi- holding, backorder, and production costs, while satisfy-
nary stage and Weibull distribution for the forecast stage, ing the demand for all products over the horizon under
Approach 2 rather considers that the RUL probabilities consideration. Each rolling horizon consists of ordinary
of forecast stages are the same as the ordinary stage. That and forecast stages, and by collecting new sensor data,
7988 H. DEHGHAN SHOORKAND ET AL.

Figure 4. The impact of demand variation on the total cost.

the decisions were simultaneously updated. A long short- possible using the available benchmarking data set devel-
term memory model was used to accurately predict the oped by NASA, and is used as a reference to evaluate the
health condition of the machine. In deep learning lit- other approaches. This public data set is always used in
erature, the LSTM network is known as a sequence to the literature as a reference to test new deep learning algo-
sequence (Seq2Seq) model, which is well-suited to classi- rithms. Our numerical results show that the newly pro-
fying, processing and making predictions based on time posed integrated model provides a solution with a total
series data, since there can be lags of unknown duration cost closer to the reference model than other approaches.
between important events in a time series. LSTM can We are currently developing an integrated model for a
process not only single data points, but also the entire system composed of several machines. Besides, defining
sequence of data. This characteristic makes it very suit- different levels of PM is another possible extension of the
able for studying multiple-period planning problems. In proposed model.
the proposed framework, each period can be seen as a
sequence. PdM actions are performed at the beginning of Acknowledgements
the period. The obtained results confirm the LSTM effi-
The authors would like to thank the editors, and the anonymous
ciency to predict the RUL class for the next period based referees for their constructive comments and recommenda-
on the data available from the previous period (‘period to tions, which have significantly improved the presentation of
period’). The proposed framework was validated using a this paper. They would like to thank also NASA Ames Research
benchmarking data set. The numerical results not only Center for developing and providing the run-to-failure bench-
validate the LSTM prediction model by evaluating the marking data set used in this paper.
accuracy and the confusion matrix, but also compare the
performance of the proposed integrated model to a ref- Disclosure statement
erence model (Approach 1) corresponding to the actual No potential conflict of interest was reported by the authors.
run-to-failure data. In this (reference) case, we assume
that the decision-maker has a ‘perfect information’ on Funding
the actual RUL to accurately plan the maintenance and
production. So, this is the ‘ideal’ situation that can pro- This work was financially supported by grants awarded by
the Natural Sciences and Engineering Research Council of
vide the minimum total cost. In reality, it is impossible Canada through the Discovery Grant Program [numbers
to access the exact degradation behaviour of each spe- RGPIN/04484-2016 and RGPIN/084141-2020]. This support is
cific machine to perform accurate planning. This is made gratefully acknowledged.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7989

Notes on contributors References


Hassan Dehghan Shoorkand is a Ph.D. Abadi, M., A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro,
candidate in Mechanical Engineering at G. S. Corrado, et al. 2015. TensorFlow: Large-Scale Machine
Laval University, Canada. His thesis con- Learning on Heterogeneous Systems. https://www.tensorflow.
cerns integrating predictive maintenance org/.
and production planning by taking advan- Abernethy, R. B. 2004. The New Weibull Handbook: Reliability
tage of deep learning methods. He also and Statistical Analysis for Predicting Life, Safety, Supporta-
received his M.Sc. in Economics from bility, Risk, Cost and Warranty Claims. 5th ed. North Palm
Allameh Tabataba’i University, Iran, in Beach, FL: Dr. RobertAbernethy.
2016. From 2016 to 2019, he was a member of the Central Aghezzaf, E. H., M. A. Jamali, and D. Ait-Kadi. 2007. “An Inte-
Bank of Iran (CBI) as a data analyst. His research interest grated Production and Preventive Maintenance Planning
includes artificial intelligence (AI), industry 4.0 concepts, Inter- Model.” European Journal of Operational Research 181 (2):
net of Things (IoT), predictive maintenance (PdM), Big Data, 679–685. doi:10.1016/j.ejor.2006.06.032.
optimisation, and their applications in the industry. Alimian, M., V. Ghezavati, and R. Tavakkoli-Moghaddam.
2020. “New Integration of Preventive Maintenance and
Mustapha Nourelfath is a professor of
Production Planning with Cell Formation and Group
Industrial Engineering at the Department
Scheduling for Dynamic Cellular Manufacturing Systems.”
of Mechanical Engineering of Laval Uni-
Journal of Manufacturing Systems, 341–358. doi:10.1016/j.
versity (Canada). He received his Ph.D.
jmsy.2020.06.011.
from the National Institute of Applied Sci-
Arani, M., M. Dastmard, Z. D. Ebrahimi, M. Momenitabar, and
ence (INSA) of Lyon (France) in 1997.
X. Liu. 2020. “Optimizing the Total Production and Main-
He is a regular member of CIRRELT
tenance Cost of an Integrated Multi–Product Process and
(Interuniversity Research Centre on Enter-
Maintenance Planning (IPPMP) Model.” In IEEE Interna-
prise Networks, Logistics and Transportation). His specific
tional Symposium on Systems Engineering (ISSE), Vienna,
topics of interest are operations research and artificial intel-
Austria, 1–8. IEEE.
ligence applications in supply chains, reliability, logistics and
Arena, M., V. Di Pasquale, R. Iannone, S. Miranda, and S.
manufacturing. He is Associate Editor of INFOR: Information
Riemma. 2022. “A Maintenance Driven Scheduling Cock-
Systems and Operational Research (Area: Queueing Theory,
pit for Integrated Production and Maintenance Operation
Stochastic modelling and simulation), member of the Editorial
Schedule.” Advances in Manufacturing, 1–15. doi:10.1007/
Board of Reliability Engineering and Systems Safety, Interna-
s40436-021-00380-z.
tional Journal of Performability Engineering (System Analy-
Aydin, O., and S. Guldamlasioglu. 2017. “Using LSTM Net-
sis and Computational Intelligence), International Journal of
works to Predict Engine Condition on Large Scale Data
Mathematical, Engineering and Management Sciences, Jour-
Processing Framework.” In 2017 4th International Confer-
nal of Applied Mathematics, and Mathematical Problems in
ence on Electrical and Electronic Engineering (ICEEE), April,
Engineering. Prof. Nourelfath has received several research
281–285. Ankara: IEEE.
grants and awards, including the Brockhouse Canada Prize for
Ayvaz, S., and K. Alpay. 2021. “Predictive Maintenance System
Interdisciplinary Research in Science and Engineering.
for Production Lines in Manufacturing: A Machine Learn-
Adnène Hajji is a full professor of oper- ing Approach Using IoT Data in Real-Time.” Expert Systems
ations and decision systems at Université with Applications 173. doi:10.1016/j.eswa.2021.114598.
Laval in Québec City, Canada. He is the Bahria, N., I. H. Dridi, A. Chelbi, and H. Bouchriha.
director of the Center for Research on 2020. “Joint Design of Control Chart, Production and
Intelligent Communities (CeRCI). He is Maintenance Policy for Unreliable Manufacturing Sys-
a member of the Interuniversity Research tems.” Journal of Quality in Maintenance Engineering.
Center on Enterprise Networks, Logis- doi:10.1108/JQME-01-2020-0006.
tics and Transportation (CIRRELT). He Bajestani, M. A. 2014. “Integrating Maintenance Planning
received the Engineering degree in Mechanical Engineering and Production Scheduling: Making Operational Decisions
from Ecole Nationale d’Ingénieurs de Tunis, Tunisia (1999) and with a Strategic Perspective.” Doctoral diss., University of
his M.Eng (2003) and PhD (2007) in Automated Production Toronto.
Engineering both from Ecole de Technologie Supérieure, Mon- Bampoula, X., G. Siaterlis, N. Nikolakis, and K. Alexopou-
tréal. His main research interest includes production system los. 2021. “A Deep Learning Model for Predictive Mainte-
modelling and control, simulation, integrated reactive mod- nance in Cyber-Physical Production Systems Using LSTM
els in ERP systems, performance management systems design Autoencoders.” Sensors, 972. doi:10.3390/s21030972.
and implementation, Industry 4.0 concepts, and reconfigurable Beheshti Fakher, H., M. Nourelfath, and M. Gendreau.
manufacturing systems. 2017. “A Cost Minimisation Model for Joint Produc-
tion and Maintenance Planning Under Quality Con-
straints.” International Journal of Production Research 55.
Data availability statement doi:10.1080/00207543.2016.1201605.
Bougacha, O., C. Varnier, N. Zerhouni, and S. Hajri-Gabouj.
The data that supports the findings of this study is openly avail- 2019. “Integrated Production and Predictive Maintenance
able on NASA at https://ti.arc.nasa.gov/tech/dash/groups/pcoe/ Planning Based on Prognostic Information.” In International
prognostic-data-repository/#turbofan (Saxena and Geodel Conference on Advanced Systems and Emergent Technologies,
2008). 363–368. doi:10.1109/ASET.2019.8871007.
7990 H. DEHGHAN SHOORKAND ET AL.

Cassady, C. R., and E. Kutanoglu. 2005. “Integrating Preven- Engineering & System Safety, 107560. doi:10.1016/j.ress.2021.107
tive Maintenance Planning and Production Scheduling for 560.
a Single Machine.” IEEE Transactions on Reliability 54 (2): Hesabi, H., M. Nourelfath, and A. Hajji. 2022. “A Deep Learn-
304–309. doi:10.1109/TR.2005.845967. ing Predictive Model for Selective Maintenance Optimiza-
Chand, S., V. N. Hsu, and S. Sethi. 2002. “Forecast, Solution, tion.” Reliability Engineering & System Safety 219: 108191.
and Rolling Horizons in Operations Management Problems: doi:10.1016/j.ress.2021.108191.
A Classified Bibliography.” Manufacturing & Service Oper- Hnaien, F., F. Yalaoui, A. Mhadhbi, and M. Nourelfath. 2016.
ations Management 4 (1): 25–43. doi:10.1287/msom.4.1.25. “A Mixed-Integer Programming Model for Integrated Pro-
287. duction and Maintenance.” IFAC-PapersOnLine 49 (12):
Chaurasiya, P. K., S. Ahmed, and V. Warudkar. 2018. “Study 556–561. doi:10.1016/j.ifacol.2016.07.694.
of Different Parameters Estimation Methods of Weibull Dis- Hochreiter, S., and J. Schmidhuber. 1996. “LSTM Can Solve
tribution to Determine Wind Power Density Using Ground Hard Long Time Lag Problems.” Advances in Neural Infor-
Based Doppler SODAR Instrument.” Alexandria Engineer- mation Processing Systems 9: 473–479.
ing Journal 57 (4): 2299–2311. doi:10.1016/j.aej.2017.08. Kaddachi, R., A. Gharbi, and J. P. Kenné. 2022. “Integrated
008. Production and Maintenance Control Policies for Failure-
Chen, C., C. Wang, N. Lu, B. Jiang, and Y. Xing. 2021. “A Data- Prone Manufacturing Systems Producing Perishable Prod-
Driven Predictive Maintenance Strategy Based on Accurate ucts.” The International Journal of Advanced Manufacturing
Failure Prognostics.” Eksploatacja i Niezawodność 23 (2): Technology, 1–23. doi:10.1007/s00170-021-08273-y.
387–394. doi:10.17531/ein.2021.2.19. Khatab, A., C. Diallo, E. H. Aghezzaf, and U. Venkatadri.
Chollet, F. 2015. Keras. https://keras.io/. 2019. “Integrated Production Quality and Condition-Based
Cohen, A. C. 1965. “Maximum Likelihood Estimation in Maintenance Optimisation for a Stochastically Deteriorating
the Weibull Distribution Based on Complete and on Cen- Manufacturing System.” International Journal of Production
sored Samples.” Technometrics 7 (4): 579–588. doi:10.1080/ Research 57. doi:10.1080/00207543.2018.1521021.
00401706.1965.10490300. Liu, Q., M. Dong, and F. F. Chen. 2018. “Single-Machine-Based
Cui, P. H., J. Q. Wang, and Y. Li. 2021. “Data-Driven Mod- Joint Optimization of Predictive Maintenance Planning and
elling, Analysis and Improvement of Multistage Production Production Scheduling.” Robotics and Computer-Integrated
Systems with Predictive Maintenance and Product Qual- Manufacturing 50: 238–247. doi:10.1016/j.rcim.2018.01.002.
ity.” International Journal of Production Research, 1–18. Medjaher, K., and N. Zerhouni. 2013. “Hybrid Prognostic
doi:10.1080/00207543.2021.1924412. Method Applied to Mechatronic Systems.” The International
Dargan, S., M. Kumar, M. R. Ayyagari, and G. Kumar. Journal of Advanced Manufacturing Technology 69 (1–4):
2019. “A Survey of Deep Learning and Its Applica- 823–834. doi:10.1007/s00170-013-5064-0.
tions: A New Paradigm to Machine Learning.” Archives Morariu, C., O. Morariu, S. Răileanu, and T. Borangiu. 2020.
of Computational Methods in Engineering 27: 1071–1092. “Machine Learning for Predictive Scheduling and Resource
doi:10.1007/s11831-019-09344-w. Allocation in Large Scale Manufacturing Systems.” Comput-
de Sampaio, Raimundo J. B., Rafael R. G. Wollmann, and ers in Industry 120. doi:10.1016/j.compind.2020.103244.
Paula F. G. Vieira. 2017. “A Flexible Production Planning Namuduri, S., B. N. Narayanan, V. S. P. Davuluru, L. Burton,
for Rolling-Horizons.” International Journal of Production and S. Bhansali. 2020. “Deep Learning Methods for Sen-
Economics 190: 31–36. doi:10.1016/j.ijpe.2017.01.003. sor Based Predictive Maintenance and Future Perspectives
El Cadi, A. A., A. Gharbi, K. Dhouib, and A. Artiba. 2021. for Electrochemical Sensors.” Journal of the Electrochemical
“Joint Production and Preventive Maintenance Controls for Society 167 (3): 037552. doi:10.1149/1945-7111/ab67a8.
Unreliable and Imperfect Manufacturing Systems.” Jour- Nguyen, K. T., and K. Medjaher. 2019. “A New Dynamic Pre-
nal of Manufacturing Systems, 263–279. doi:10.1016/j.jmsy. dictive Maintenance Framework Using Deep Learning for
2020.12.003. Failure Prognostics.” Reliability Engineering & System Safety
Fitouri, C., N. Fnaiech, C. Varnier, F. Fnaiech, and N. Zer- 188. doi:10.1016/j.ress.2019.03.018.
houni. 2016. “A Decision-Making Approach for Job Shop Nourelfath, M., and E. Châtelet. 2012. “Integrating Produc-
Scheduling with Job Depending Degradation and Predic- tion, Inventory and Maintenance Planning for a Parallel Sys-
tive Maintenance.” IFAC-PapersOnLine 49 (12): 1490–1495. tem with Dependent Components.” Reliability Engineering
doi:10.1016/j.ifacol.2016.07.782. & System Safety, 59–66. doi:10.1016/j.ress.2012.02.001.
Géhan, M., B. Castanier, and D. Lemoine. 2014. “Integra- Nourelfath, M., M. C. Fitouhi, and M. Machani. 2010. “An
tion of Maintenance in the Tactical Production Planning Integrated Model for Production and Preventive Mainte-
Process under Feasibility Constraint.” In IFIP International nance Planning in Multi-State Systems.” EEE Transactions
Conference on Advances in Production Management Systems, on Reliability, 496–506. doi:10.1109/TR.2010.2056412.
467–474. Berlin: Springer. Okoh, C., R. Roy, and J. Mehnen. 2017. “Predictive Mainte-
Ghaleb, M., S. Taghipour, M. Sharifi, and H. Zolfagharinia. nance Modelling for Through-Life Engineering Services.”
2020. “Integrated Production and Maintenance Schedul- Procedia CIRP 59: 196–201. doi:10.1016/j.procir.2016.09.033.
ing for a Single Degrading Machine with Deterioration- Pan, E., W. Liao, and L. Xi. 2012. “A Joint Model of
Based Failures.” Computers & Industrial Engineering 143. Production Scheduling and Predictive Maintenance for
doi:10.1016/j.cie.2020.106432. Minimizing Job Tardiness.” The International Journal of
Han, X., Z. Wang, M. Xie, Y. He, Y. Li, and W. Wang. Advanced Manufacturing Technology 60 (9–12): 1049–1061.
2021. “Remaining Useful Life Prediction and Predic- doi:10.1007/s00170-011-3652-4.
tive Maintenance Strategies for Multi-State Manufacturing Paprocka, I., W. M. Kempa, and B. Skołud. 2021. “Predic-
Systems Considering Functional Dependence.” Reliability tive Maintenance Scheduling with Reliability Characteristics
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7991

Depending on the Phase of the Machine Life Cycle.” Wang, L., Z. Lu, and Y. Ren. 2020. “Integrated Produc-
Engineering Optimization 53. doi:10.1080/0305215X.2020. tion Planning and Condition-Based Maintenance Consid-
1714041. ering Uncertain Demand and Random Failures.” Proceed-
Pourghader Chobar, A. 2022. “Mathematical Modeling and ings of the Institution of Mechanical Engineers, 310–323.
Problem Solving Integrated Production Planning and Pre- doi:10.1177/0954405419852479.
ventive Maintenance with Limited Human Resources.” Jour- Yildirim, M. B., and F. G. Nezami. 2014. “Integrated Mainte-
nal of New Researches in Mathematics. nance and Production Planning with Energy Consumption
Regler, A. 2020. “Data-Driven Integrated Production and and Minimal Repair.” The International Journal of Advanced
Maintenance Optimization.” Operations Research Proceed- Manufacturing Technology, 1419–1430. doi:10.1007/s00170
ings, 43–49. doi:10.1007/978-3-030-48439-2_6. -014-5834-3.
Saxena, A., and K. Geodel. 2008. Turbofan Engine Degradation Yousefi, N., S. Tsianikas, and D. W. Coit. 2021. “Dynamic Main-
Simulation Data Set, NASA Ames Prognostics Data Reposi- tenance Model for a Repairable Multi-Component System
tory. Moffett Field, CA: NASA Ames Research Center. Using Deep Reinforcement Learning.” Quality Engineering,
Shao, X., Z. Chen, and B. R. Sarker. 2022. “An Joint Deci- 1–20. doi:10.1080/08982112.2021.1977950
sion of Production and Maintenance Plan (Q, N) for Zhai, S., M. G. Kandemir, and G. Reinhart. 2022. “Pre-
a Two-Stage Deteriorating JIT Production System with dictive Maintenance Integrated Production Scheduling by
Random Breakdowns.” Production Engineering, 89–107. Applying Deep Generative Prognostics Models: Approach,
doi:10.1007/s11740-021-01079-7. Formulation and Solution.” Production Engineering, 1–24.
Tobon-Mejia, D. A., K. Medjaher, N. Zerhouni, and G. doi:10.1007/s11740-021-01064-0
Tripot. 2012. “A Data-Driven Failure Prognostics Method Zhao, R., J. Wang, R. Yan, and K. Mao. 2016. “Machine Health
Based on Mixture of Gaussians Hidden Markov Mod- Monitoring with LSTM Networks.” In 2016 10th Interna-
els.” IEEE Transactions on Reliability 61 (2): 491–503. tional Conference on Sensing Technology (ICST), 1–6. Nan-
doi:10.1109/TR.2012.2194177. jing: IEEE.
Triska, Y., ÍRS Agostino, P. M. Penna, L. F. Braghirolli, and Zheng, S., K. Ristovski, A. Farahat, and C. Gupta. 2017. “Long
E. M. Frazzon. 2021. “Integrated Production and Mainte- Short-Term Memory Network for Remaining Useful Life
nance Planning Method with Simulation-Based Optimiza- Estimation.” In 2017 IEEE International Conference on Prog-
tion.” IFAC-PapersOnLine, 349–354. doi:10.1016/j.ifacol. nostics and Health Management (ICPHM), 88–95. Dallas,
2021.08.160. TX: IEEE. doi:10.1109/ICPHM.2017.7998311.
Wang, S., and M. Liu. 2013. “A Branch and Bound Algorithm Zonta, T., C. A. da Costa, F. A. Zeiser, G. de Oliveira Ramos,
for Single-Machine Production Scheduling Integrated with R. Kunst, and R. da Rosa Righi. 2022. “A Predictive Main-
Preventive Maintenance Planning.” International Journal of tenance Model for Optimizing Production Schedule Using
Production Research 51 (3): 874–868. doi:10.1080/00207543. Deep Neural Networks.” Journal of Manufacturing Systems,
2012.676683. 450–462. doi:10.1016/j.jmsy.2021.12.013.

You might also like