Production Engineering (2022) 16:65–88

https://doi.org/10.1007/s11740-021-01064-0

MECHANICAL ENGINEERING

Predictive maintenance integrated production scheduling by applying

deep generative prognostics models: approach, formulation
and solution
Simon Zhai1 · Meltem Göksu Kandemir1 · Gunther Reinhart1

Received: 11 April 2021 / Accepted: 8 June 2021 / Published online: 29 July 2021
© The Author(s) 2021

Abstract
To harness the full potential of predictive maintenance (PdM), PdM information has to be used to optimally plan production
and maintenance actions. Hence, operation-specific modelling of degradation, i.e. predictions of the health condition under
time-varying operational conditions, has to be realized. By utilizing operation-specific degradation information, maintenance
and production can be planned with regard to each other and thus, predictive maintenance integrated production schedul-
ing (PdM-IPS) is enabled. This publication proposes a novel PdM-IPS approach consisting of two interacting modules:
an operation-specific Prognostics and Health Management (PHM) module and an integrated production scheduling and
maintenance planning (IPSMP) module. Specifically, the mathematical problem of the IPSMP module based on an extended
version of the maintenance integrated flexible job shop problem is formulated. A two-stage genetic algorithm to efficiently
solve this problem is designed and subsequently applied to simulated condition monitoring, as well as real industrial data.
Results indicate that the approach is able to find feasible high quality PdM integrated production schedules.

Keywords Predictive maintenance · Integrated scheduling · Genetic algorithm · Decision support

1 Introduction regarding its value-adding implementation in industry [1].

RUL estimation alone does not lead to value-added, deci-
Predictive maintenance (PdM) is one of the core Industry 4.0 sion support does [2]. One crucial missing link to generate
innovations that substantially increases machine availabil- value-added is the inability of most existing PdM models to
ity by preventing machine failure and enabling maintenance account for varying operational conditions, i.e. to estimate
actions just in time. Using advanced analytics on condition operation-specific degradation [3]. On the one hand, opera-
monitoring (CM) data collected by powerful sensors, the tion-specific predictions are needed for subsequent schedul-
current health condition and remaining useful life (RUL) ing of production and maintenance since different products
of machines can be predicted. This enables the introduc- and their required operations induce different stresses on the
tion of new maintenance strategies. Recent studies indicate machine. On the other hand, existing scheduling algorithms
that while industrial companies do acknowledge the upside are not designed to jointly optimize production scheduling
potential of PdM, this technology still faces challenges and maintenance planning using PdM information. Thus, the
focus of this publication is the mathematical formulation of
the integrated production scheduling and maintenance plan-
* Simon Zhai ning (IPSMP) problem, its interactions with the Prognostics
[email protected] and Health Management (PHM) module developed by Zhai
Meltem Göksu Kandemir et al. [4] and the subsequent solution of the problem using
[email protected] a novel two-stage genetic algorithm (TSGA) that is specifi-
Gunther Reinhart cally designed for interacting with PHM modules.
[email protected] This paper is structured as follows: Sect. 2 gives an over-
1 view of related work and derives the research objective of
Department of Mechanical Engineering, Institute
for Machine Tools and Industrial Management, Technical this publication. Section 3 introduces the concept of the PdM
University of Munich, Garching, Germany integrated production scheduling (PdM-IPS) approach. Its

13
Vol.:(0123456789)
66 Production Engineering (2022) 16:65–88

mathematical model is presented in Sect. 4, while Sect. 5 considered a single machine with several operational pro-
proposes a two-stage Genetic Algorithm for its optimization files. The degradation processes were modeled using expo-
and solution. Computational results applying the approach to nential functions. However, it was not specified how the
benchmark data, as well as real industrial data, are presented parameters of the exponential functions were determined
in Sect. 7. Finally, Sect. 8 concludes the main findings of and in what relation they stand to the individual operating
this paper. profiles of the machine.
Ghaleb et al. [5] followed a similar approach to those
of Sloan and Shanthikumar [10] and Bajestani et al. [11]
2 State of the art and research objective by modelling the degradation process as a continuous-time
Markov chain with discrete states. A single machine with
A detailed literature study was carried out to shed light on different repair policies was studied: full repair, partial
the current advances in IPSMP, as well as to highlight exist- repair (one-step or multi-step) and no repair. The duration
ing shortcomings to derive research objectives. The entity and cost of a repair depend on the policy and the condition
of IPSMP approaches can be divided into the categories of the machine. Furthermore, deteriorating machine condi-
time-based and condition-based maintenance integrated tion was assumed to lengthen the job processing times and
approaches [5]. Time-based approaches plan production increase the energy consumption. The model was solved
and maintenance according to predetermined maintenance using a genetic algorithm (GA) with a total cost minimiza-
actions with fixed starting times or time windows, i.e. inte- tion objective.
grating the preventive maintenance strategy into produc- Pan et al. [12] utilized sensor and prognostic technologies
tion scheduling without considering the real health condi- to continuously monitor a single machine, i.e. to predict its
tion of a machine, see for example [6–8]. Condition-based degradation process during the processing of a job sequence.
approaches, such as PdM, carry out maintenance actions Based on this predictive information, maintenance opera-
based on the assessed or predicted health condition and will tions are scheduled simultaneously with production jobs.
be the focus for the following section. The literature will be Fitouri et al. [13] proposed a decision-making heuris-
reviewed and assessed according to the industrial viability tic with decision rules. They considered a job shop with
of condition-based IPSMP approaches. In detail, these cri- machines subject to operation-dependent degradation. A
teria are: prognostic module continuously monitors the job shop sys-
tem and provides RUL predictions for each operation on
1. consideration of a job shop scheduling problem (JSSP) each machine. Based on the predictive RUL, the cumulative
or flexible job shop scheduling problem (FJSSP) and machine degradation ∆ imposed by an operation is calcu-
their respective production metrics (e.g. makespan) in lated as follows:
order to be applicable for different shop floor layouts and p
not being limited to single machine setups, Δ= (1)
RUL
2. consideration of the machine condition for PdM plan-
ning, where p denotes the operation’s processing duration. The
3. operation-specific modeling of machine degradation, total degradation of a machine imposed by a production
4. consideration of the non-linearity, i.e. the operation schedule equals to the linear accumulation of operation-spe-
sequence-dependency, of the machine degradation pro- cific degradations of the assigned operations. Each machine
cess, and is associated with a minimum and maximum threshold of
5. use of real industrial data for the assessment and predic- degradation. The proposed heuristic aims at finding optimal
tion of machine condition. starting times for PdM actions as well as the production jobs.
Ladj et al. [14] pursued a similar approach as Fitouri
2.1 Literature review et al. [13] and studied the operation-specific degradation
of a single machine. The cumulative degradation of each
Bougacha et al. [9] proposed a PdM integrated IPSMP operation is calculated using Eq. (1) based on RUL predic-
framework with two modules; an algorithm used for long- tions provided by a PHM module. The formulated IPSMP
term RUL predictions of the system and an algorithm used problem was solved using a GA with a total cost minimiza-
for short-term predictions to estimate the system’s future tion criterion. The proposed approach was later extended
state while executing a job or maintenance. by Ladj et al. [15] to take uncertainty in production shops
The proposed framework adopts an iterative approach by into account. They considered a flow shop with uncertain
considering “the effect of a selected decision on the sys- RUL values and degradation of machines, which were both
tem health and the new possible decision that can be intro- represented using fuzzy sets. The fuzzy sets are constructed
duced by a change in the estimated RUL” [9]. The authors

13
Production Engineering (2022) 16:65–88 67

using the probability density function of machine failures total idle time, while also taking into account the stability
based on historical data. robustness and quality robustness of the schedules. In case of
Another study investigating the RUL uncertainty was car- machine breakdown, jobs are rescheduled based on heuristic
ried out by Benaggoune et al. [16]. The authors considered shifting rules.
a multifunctional single machine and studied the impact of Denkena et al. [20] proposed a statistical method to
RUL uncertainty on maintenance planning. A PHM mod- estimate failure durations of machine tools in order to
ule was deployed to continuously monitor the machine improve production scheduling. By using data from prac-
and deliver its estimated RUL after the processing of each tical experiments, their method showcases high accuracy
operation. The operations can be scheduled with respect to of these prognosis that in turn can be used for production
a predefined maximum degradation threshold, which cannot scheduling purposes. While operation-specific prediction of
be exceeded in any situation. An operation-specific degrada- RUL and subsequent scheduling was not within the scope of
tion δ was considered, whereas the evolution of the machine their work, the results have potential for holistically improve
degradation f(δ) was considered linear. The aim of the for- IPSMP.
mulated model is the minimization of total maintenance cost
formulated as follows: 2.2 Research Gap and Research Objective
( )
CPdM = CPdM,fix + CPdM,adv 𝛿i (t) (2) The assessment of the relevant literature regarding the crite-
ria given in the beginning of this chapter is given in Table 1.
where CPdM,fix is the fixed cost of maintenance and CPdM,adv
It is notable that those publications with focus on solving
is the total cost of advancement, which is a linear function
IPSMP problems apply genetic algorithms as solvers due to
of the cost of advancement per unit time.
its ability to solve multi-objective optimization problems in
Zhai et al. [17] studied a job shop scheduling problem
reasonable time. In addition to the presented literature, fur-
(JSSP) under time-varying operational conditions. The
ther publications were studied and listed in the table to fully
machines are subject to degradation and stochastic failures
capture the scientific landscape, but not described in the sec-
that follow the Weibull distribution. An operation-specific
tion above for the sake of brevity. As the analysis indicates,
stress equivalent was introduced to represent the machine
the state of the art does not provide an efficient approach for
degradation imposed by a job operation. A not specified
the PdM integrated production scheduling of flexible job
PHM module supplies RUL values, which are transformed
shops under varying operational conditions. Especially the
into the operation-specific stress equivalents using Eq. (1).
linear accumulation of degradation as found in publications
The authors assumed a linear accumulation of operation-
[13–17] does not hold true in industry and simplifies the
specific degradation values that in turn influence the fail-
problem at hand. Although some publications have dealt
ure probability using the Weibull distribution. A GA was
with the specific aspects of the problem formulation of this
employed to solve the formulated problem.
work, no holistic approach was found.
Morariu et al. [18] explored the potentials of Big Data
Thus, the present work aims at developing a PdM-IPS
techniques and machine learning algorithms in PdM inte-
approach consisting of an optimization model for the main-
grated production planning concerning energy use on the
tenance integrated flexible job shop scheduling problem
shop floor. They considered a large-scale manufacturing
(MIFJSSP) using non-linear health predictions from a PHM
system, in which production scheduling decisions are based
module.
on real time learning from Big Data and data-driven deci-
sion making. Deep learning was employed to identify pos-
sible anomalies of energy consumption in the production
3 Concept
processes. The current energy use is compared to learned
energy consumption patterns, i.e. the predicted energy data
Figure 1 presents the concept of the overall PdM-IPS
based on previous production data. The proposed approach
approach with its modules. Information from the scheduler,
has two stages: long-term batch planning for optimal sched-
production orders and from the shop floor serve as inputs to
uling and resource allocation, and short-term resource health
the approach. The Planning Module receives performance
monitoring to detect anomalies in energy use and maintain
measures and the weights for both production and mainte-
the affected machine.
nance plans. In addition, it receives the job due dates from
Paprocka et al. [19] aimed at generating robust schedules
production orders. The PHM Module receives CM data from
for job shops based on historical data of failure frequency
sensors installed in machines on the shop floor. Correspond-
using the Maxwell distribution. Time-varying machine
ing operating data derived from production orders enable the
failure rates were considered. The optimal schedules are
PHM Module to correlate CM and operating data. An Inter-
selected assessing makespan, total tardiness, flow time and
face Module enables the information exchange. Schedule

13
68 Production Engineering (2022) 16:65–88

Table 1  Evaluation of the reviewed condition-based maintenance integrated production scheduling approaches (✔: considered, ✗: not consid-
ered)
JSSP/ FJSSP PdM planning based on Operation-Spe- Non-Linear Accumulation of Use of Real
Machine Condition cific Degradation Operation-Specific Degradation Industrial
Data

Benaggoune et al. [16] ✗ ✔ ✗ ✗ ✔

Denkena et al. [20] ✗ ✗ ✗ ✔ ✔
Morariu et al. [18] ✗ ✔ ✔ ✗ ✔
Ghaleb et al. [5] ✗ ✗ ✗ ✗ ✗
Bougacha et al. [9] ✗ ✔ ✗ ✗ ✔
Khatab et al. [21] ✗ ✗ ✗ ✗ ✔
Ladj et al. [15] ✗ ✔ ✔ ✗ ✔
Zhai et al. [17] ✔ ✔ ✔ ✗ ✗
Zandieh et al. [22] ✔ ✗ ✗ ✗ ✗
Fitouri et al. [13] ✔ ✔ ✔ ✗ ✔
Aramon Bajestani et al. [11] ✗ ✗ ✗ ✗ ✔
Xiang et al. [23] ✗ ✗ ✗ ✗ ✗
Pan et al. [12] ✗ ✔ ✗ ✗ ✔

Fig. 1  Overall PdM-IPS frame-

work (MP maintenance plan, PS
production schedule)

candidates are sent to the PHM Module and the predicted the IPSMP part is based on a maintenance integrated ver-
degradation imposed by this very schedule candidate will sion of the FJSSP to account for different shop floor layouts
be transmitted back to the Planning Module. Based on both (i.e. MIFJSSP). As mentioned, this publication focuses the
production and degradation metrics, the Planning Module planning optimization using a two-stage Genetic Algorithm.
derives the output of the framework: an integrated produc- Thus, the Interface, Planning Module and the interactions
tion schedule with maintenance actions. Mathematically, with the PHM Module will be explained in detail. Based on

13
Production Engineering (2022) 16:65–88 69

literature analysis [24] and the author’s own studies [17], 4.2 Subsystem 1: machine
genetic algorithms are well-suited to solve IPSMP problems.
The PHM Module that will be applied is based on the The flexible job shop comprises a set of machines, denoted
work of Zhai et al. [4] and is able to assess and predict the as 𝐌, on which the jobs are produced. A single machine is
health condition of a machine after manufacturing a specific denoted as Mk ∈ 𝐌, whereas the total number of machines
production schedule. Specifically, two Conditional Varia- in the shop floor equals to m . The following assumptions
tional Autoencoder Neural Networks are applied to derive (A) hold:
operation-specific health indicator prediction of machines. A.1 A machine can only execute one activity (job opera-
The algorithms presented in this publication are able to work tion or a PdM action) at a time.
with any other PHM module that is capable of returning an A.2 The degree of degradation on each machine is oper-
operation or product-specific health indicator prediction. We ation-specific. The health state, i.e. the degradation level, of
recommend our approach for short-term planning, i.e. for the machine remains constant during setups and idle state.
planning multiple shifts or weekly schedules. Depending A.3 PdM actions reset the machine to an “as good as
on the data quality and availability of the PHM module, and new” state.
thus the quality of predictions, longer planning horizons can A.4 No unexpected machine breakdown or failure occurs
be also realized. during the schedule horizon.
The HIk of machine takes a value between 0 (degraded)
and 1 (new) representing the health of the machine Mk . RULk
4 Modelling refers to the time period in which the machine is expected to
be functional without breakdowns. Appendix Table 5 gives
Based on the conceptual approach presented in the previous an overview of the introduced notations and variables to
chapter, this chapter establishes the mathematical foundation describe the machines in a flexible job shop.
of the PdM-IPS. Hereby, the interaction between the Plan-
ning and the PHM Module is presented in depth. 4.3 Subsystem 2: Product and Operation

4.1 Problem description The flexible job shop can produce numerous product vari-
ants, denoted as Pv. The set of all product variants is referred
The present work studies a flexible job shop scheduling prob- to as product portfolio denoted by 𝐏 . Each product variant
lem with integrated predictive maintenance planning based comprises several successive operations Ovj with index j of
on operation-specific machine degradation. Mathemati- variant v . For each product Pv ∈ 𝐏 , the following holds:
cally, the integrated problem consists of two simultaneous A.5 The operations Ovj of a product variant Pv are subject
optimization problems: namely the production scheduling to precedence constraints.
and maintenance planning in a flexible job shop setting, i.e. The total number of operations comprised by a product
MIFJSSP. The production scheduling further comprises the variant Pv is denoted as ov , with 𝐎v being its ordered set
machine assignment and operation sequencing subproblems. of operations. The set of machines an operation Ovj can be
The MIFJSSP is extended to the PdM-IPS approach by inte- assigned to is referred to as the eligible machine set of that
grating information from a PHM Module and resulting PdM operation and is denoted by 𝐌vj . Each operation Ovj has a
actions. The product portfolio contains several product vari- deterministic processing duration pvjk on machine Mk:
ants, which are produced on the multifunctional machines of A.6 The processing times of operations on machines are
the flexible job shop. The machines are subject to varying fixed, i.e. deterministic.
operating conditions, and thus varying degradation, due to Appendix Table 6 gives an overview of the introduced
the execution of multiple types of operation using the same notations and variables. Operations are associated with dif-
machine. A production order initiates the manufacturing ferent operating conditions which are defined by the com-
process of a specific number of units of a product variant, bination of different operational parameters [25] that lead
each unit corresponds to a job. In the following section, a to distinct CM values. These can be clustered into different
system model for the PdM-IPS is established with the fol- load profiles, so called operating regimes. This concept will
lowing subsystems: (1) machine, (2) product and operation, be introduced in detail in Sect. 4.5.2.
and (3) production order and job. All introduced variables
and descriptions can be found in Appendix Tables 5, 6 and 7.

13
70 Production Engineering (2022) 16:65–88

4.4 Subsystem 3: Production Order and Job The operations of a job are subject to precedence con-
straints, i.e., must be executed in the given order. Each oper-
Each workpiece requested by a production order represents ation is associated with a set of eligible machines 𝐌ℴij on
a job Jℴi with the following data: which it can be processed:

• product variant v(ℴ) to be produced, and

𝐌ℴij = 𝐌v(ℴ)j (4)
• due date dℴi equal to the due date dℴ of the production Figure 2 illustrates the notation and the principle of inher-
order 𝐏𝐎ℴ it belongs to, itance between product variants and jobs.
Appendix Table 7 gives an overview of introduced vari-
where i is the running index of jobs. The total number of ables to describe production orders and jobs.
production orders is denoted as Q. Within the scope of this
work, following holds: 4.5 Mathematical formulation
A.7 Production orders correspond to only one type of
product variant Pv each. This section deals with the mathematical formulation of
Accordingly, each job inherits the specifications of the the PdM-IPS approach based on the assumptions made in
product variant v(ℴ) it corresponds to. Based on the subsys- Sect. 4.1.
tem “products and operations”, following notations are intro-
duced: each job Jℴi comprises ov(ℴ) operations Oℴij , whereas 4.5.1 The production scheduling problem
the j-th operation Oℴij of job Jℴi is an instance of the j-th
operation Ov(ℴ)j of product Pv(ℴ) with the same machine pro- The production scheduling problem of the PdM-IPS com-
gram (cf. Fig. 2) and processing duration pℴijk as follows: prises the machine assignment and operation sequencing
poijk = pv(o)jk (3) subproblems of the FJSSP.
A.10 All jobs are available at the beginning of the plan-
ning horizon.
The starting and completion times of an operation Oℴij Following decision variable Yℴij,ℴ′ i′ j′ ,k is introduced to the
are denoted by Sℴij and Cℴij . The completion time of job model:
Jℴi is denoted by Cℴi . Once a job operation Oℴij has started {
on a machine Mk , it cannot be interrupted or preempted by 1, if Oℴij immediately precedes Oℴ� i� j� on Mk
Yℴij,ℴ� i� j� ,k =
another operation: 0, Otherwise
A.8 Job preemptions are not allowed. (5)
The set of all jobs of the PdM-IPS instance is denoted as
In accordance with A.1, a machine can execute only one
𝐉 , where n denotes the total number of jobs. The job opera-
operation at a time, meaning that the starting time of an
tions of these jobs form an unordered set of all job opera-
operation on machine Mk must be greater than the comple-
tions, denoted as 𝐎, the number of all job operations in the
tion time of its predecessor:
problem instance is denoted as N .
( )
A.9 The jobs in the flexible job shop are independent and Sℴ� i� j� k ≥ Cℴijk − 1 − Yℴij,ℴ� i� j� k ⋅ B (6)
can be sequenced in any order that benefits the objective
function values. where B is a large positive number. For the completion
time Cℴijk of operation Oℴij on machine Mk must hold the
following:
( )
Cℴijk ≥ Sℴijk + pℴijk − 1 − Xℴijk ⋅ B (7)

with the binary decision variable Xℴijk as follows:

{
1, if Oℴij can be executed on Mk
Xℴijk = (8)
0, otherwise

A machine can only be assigned to an operation, if it

belongs to the eligible machine subset 𝐌ℴij = 𝐌v(ℴ)j (see
Eq. 4) expressed by the parameter 𝛾ℴijk = 𝛾v(ℴ)jk:

Fig. 2  Subsystem of production orders and jobs

13
Production Engineering (2022) 16:65–88 71

{
1, if Ov(o)j can be executed on Mk of setup if needed. The transportation times can be consid-
𝛾v(o)jk = (9) ered in Eq. (15), resulting in the following constraint:
0, otherwise
( )
Accordingly, the following must hold: Sℴij� ≥ Cℴij + ttransport ⋅ 𝜃tr,ℴij� − 1 − 𝒫ℴ(v)jj� ⋅ B (18)

Xoijk = 𝛾v(o)jk (10) For production scheduling, the following objective func-
tions are selected: minimization of makespan Cmax , total
indicating that the decision variable Xℴijk can only be 1, tardiness Ttotal and the total penalty cost of the production
when 𝛾v(ℴ)jk equals to 1. Furthermore, the starting and end- schedule (PS), denoted as 𝜁PS.
ing times of operation Oℴij can be formulated as follows:
Cmax = max Cℴi for ∀ℴ, i (19)
Sℴij ≥ Sℴijk for ∀ k ∈ 𝐌 (11)
whereas Cℴi denotes the completion time of job Jℴi:
Cℴij ≥ Cℴijk for ∀k ∈ 𝐌 (12) Cℴi ≥ Cℴij for ∀j ∈ Ov(ℴ) (20)
In accordance with A.5, the operations Oℴij of job Jℴi are Makespan is the most commonly used objective function
subject to precedence constraints, an operation cannot start within production scheduling and is directly related with
prior to the completion of its predecessor: the cost of a PS [26]. The minimization of Ttotal is selected
( ) with respect to the industrial applicability of the model. Any
Sℴij� ≥ Cℴij − 1 − 𝒫ℴ(v)jj� ⋅ B (13) order that exceeds its due date is considered as lost oppor-
tunity associated with costs [9]. Thus, the model aims at
whereas 𝒫ℴ(v)jj� is a binary precedence parameter as follows:
minimizing the total tardiness
{
1, if Oℴ(v)j of product variant Pℴ(v) precedes Oℴ(v)j�
Pℴ(v)jj� = . ∑ ∑
Q qℴ
0, otherwise Ttotal = Tℴi (21)
(14) ℴ=1 i=1

The jobs operations are subject to transportation times where Tℴi denotes the tardiness of job Jℴi:
ttransport and setup times tsetup: ( )
A.11 The setup times between two product variants are Tℴi = max 0, Cℴi − dℴ (22)
sequence and product independent.
A.12 The time to transport a job between machines is Lastly, the model minimizes 𝜁PS . 𝜁PS is introduced to the
taken into account and de-terministic. It has the same dura- model to minimize the transportation of jobs in the job shop
tion regardless of the origin and destination machine. as well as frequent setup actions. Each job transportation and
Transportation time applies when two consequent opera- setup action is associated with costs, denoted as costtransport
tions of a job are not processed on the same machine, and costsetup , respectively:
expressed by the following variable: Q qℴ ov(ℴ)
∑ ∑∑ ( )
{ 𝜁PS = costtransport ⋅ 𝜃transport Oℴij
1, if Xℴijk = 1 and Xℴi(j−1)k ≠ 1 (23)
𝜃tr,ℴij = (15) ℴ=1 i=1 j=1
( )
0, otherwise
+ costsetup ⋅ 𝜃transport Oℴij
The variable of setup time needed for operation Oℴij is
where 𝜃transport (Jℴi ) and 𝜃transport (Jℴi ) are model variables:
expressed as:
{
{ � ( ) 1, if ttr,ℴij ≠ 0
1, if Yoij,ℴ� i� j� ,k = 1 and v(o) ≠ v(o ) 𝜃transport Oℴij = (24)
𝜃stp,ℴ� i� j� = (16) 0, otherwise
0, otherwise
{
meaning that setup is required if Oℴij immediately precedes ( ) 1, if tstp,ℴi� j� ≠ 0
Oℴ′ i′ j′ on machine Mk and the product orders ℴ and ℴ do not
′ 𝜃setup Oℴij = (25)
0, otherwise
correspond to the same product variant. Thus, considering
setup times, Eq. (6) can be extended as follows:
( ) 4.5.2 The maintenance planning problem
Sℴi� j� k ≥ Cℴijk + tsetup Δ𝜃stp,ℴi� j� − 1 − Yℴij,ℴi� j� ,k ⋅ B (17)
Maintenance planning problem is concerned with the plan-
meaning that operation Oℴ′ i′ j′ cannot start on machine Mk
ning of PdM actions based on health state of the machines in
prior to the completion of its predecessor and the completion
the flexible job shop. It is assumed that the flexible job shop

13
72 Production Engineering (2022) 16:65–88

in consideration is equipped with a PHM Module, which ∑ ∑∑

Q qℴ v(ℴ)

continuously collects sensory data and can deliver RULk Zℴijk ≤ 1, ∀Mk ∈ 𝐌 (28)
predictions. In industrial practice, machines rarely deplete ℴ=1 i=1 j=1

all RULk. Thus, total cost of maintenance denoted as 𝜁MP is formu-

The present work adopts the concept of Remaining Main- lated as 𝜁MP:
tenance Life, denoted as RMLk , proposed by Pan et al. [12].
Each machine Mk is associated with two HI values HIk,safe ( ) ∑Q ℴ ∑
∑ ∑
v(ℴ) n q

and HIk,fail , reached at tk,safe and tk,fail , respectively, see Fig. 3. 𝜁MP = costPdM tPdM ⋅ Zℴijk (29)
Regarding maintenance planning, tk,fail represents the last ℴ=1 i=1 j=1 k=1

timestep before a PdM action has to be planned on machine The PdM-IPS approach uses the PHM Module of Zhai
Mk . The time span between tk,safe and tk,fail represents the et al. [4] to consider the machine degradation during IPSMP
ideal time to plan maintenance. An earlier starting time tk,PdM to supply RMLk and RULk predictions. Here, the model
of PdM than tk,safe means unnecessary maintenance costs. In assumes operation-specific machine degradation, referred
order to integrate RMLk and RULk , we model maintenance to as 𝛿vjk ∈ [0,1]. It corresponds to the amount of degra-
costs CPdM (see Fig. 3): dation imposed on Mk during the processing of operation

⎧ � �
� � ⎪ costPdM,fix + 𝛼PdM,adv ⋅ tk,safe − tk,PdM , if tk,PdM < tk,safe
costPdM,k tk, PdM = ⎨ costPdM,fix , if tk,safe ≤ tPdM ≤ tk,fail (26)
⎪ B, if tk,PdM > tk,fail
⎩

whereas costPdM,fix is the fixed cost of maintenance and Ovjk for one timestep, if the machine would only produce
𝛼PdM,adv is the advancement cost per timestep. this operation. The total amount of degradation imposed
The next assumptions integrate maintenance and produc- by the Ovjk is denoted as Δvjk . However, as opposed to the
tion scheduling: publications [13, 15, 17], the present work does not assume
A.13 A planned PdM action will be performed imme- linear accumulation of the operation-specific machine deg-
diately after the processing of a job operation is complete. radation. In order to model operation-specific degradation,
Thus, the binary decision variable Zℴijk is introduced to we consider each timestep of an operation individually
the model: according to its degradation. This allows us to accurately
{ predict the machine health in consideration of the active
1, if a PdM action is scheduled after Oℴijk operation sequence. The model quantifies the effects of dif-
Zℴijk =
0, otherwise ferent processing steps during an operation Ovjk , referred to
(27) as operating regime OpReg. Each OpReg represents a differ-
ent load profile of the machine. OpRegt corresponds to the
A. 14 At most one PdM action is scheduled per machine
active OpReg in timestep t and is dependent on the active job
Mk
operation Ovjk . OpRegs are identified by clustering available
This can be mathematically formulated as follows:
CM data, during the time an operation Ovjk is active, into K
clusters. Thus, an operation Ovjk corresponds to an ordered
set, i.e. sequence, of active operating regimes:
( ) [ ]
OpRegSeq Ovjk = OpRegvjk,1 , OpRegvjk,2 , … , OpRegvjk,pvjk
(30)
where it holds OpRegvjk,t ∈ [OpReg1 , … , OpRegK ] and
t ∈ pvjk . The total degradation Δvjk imposed by an operation
Ovj on machine Mk is dependent on the sequence of operating
regimes the operation comprises:
( ( ))
𝛿vjk = f OpRegSeq Ovj (31)

where OpRegSeqvj denotes the OpReg sequence of operation

Ovj . Accordingly, the degradation of machine Mk imposed
Fig. 3  RML concept [12] and adapted PdM advancement cost model by a schedule, denoted as 𝒮, also depends on the sequence
[27] of OpRegs of that schedule:

13
Production Engineering (2022) 16:65–88 73

and a total of six objective functions presented in the previ-

Δk (𝒮) = f (OpRegSeq(𝒮)) (32)
ous sections. The objective functions regarding the PS are
In accordance with assumption A.2, setup and idle times as follows:
do not change the degradation of the machine, as no OpReg
is active.
min Cmax (38)
The total degradation of machine Mk due to schedule 𝒮 is
formulated as follows: min Ttotal (39)

Δk (𝒮) = HIk,pred (𝒮) − HIk (t = 0) (33) min 𝜁PS (40)

where HIk,pred (𝒮) represents the predicted level of The maintenance plan (MP) has the following objective
HIk,pred ∈ [0,1] after the execution of schedule 𝒮. functions:
The total degradation Δtotal of the machines should be
minimal, so the frequency of maintenance actions and asso- min Δtotal (41)
ciated cost can be minimized.
∑
n min Δcritical (42)
Δtotal = Δk , (34)
(43)
k=1
min 𝜁MP
Furthermore, the critical machine degradation Δcritical , i.e.
the maximum degradation of a single machine, is chosen as
minimization criteria:
5 Two‑stage genetic algorithm (TSGA)
min Δcritical = max Δk ∀k ∈ M (35)
We propose a TSGA for the PdM integrated production
This facilitates the balanced use of resources. By using scheduling of flexible job shops. The general procedure of
the machines to a similar extent, the PdM actions can be the designed algorithm is shown in Fig. 4. The optimiza-
synchronized, which would result in time and cost savings tion process refers to the simultaneous scheduling of job
in industrial practice [17]. operations and PdM actions. The integrated scheduling is
carried out by the TSGA based on the feedback provided
4.5.3 The integrated problem by the PHM Module. As mentioned, the genetic algorithm
is chosen due to its viability to solve IPSMP problems as
The IPSMP part of the PdM-IPS approach is formulated the scientific literature points out. Specifically, its ability to
by combining the constraints of the partial problems out- solve the present multi-objective optimization problem in
lined in previous sections. In consideration of the binary reasonable time is crucial for industrial applicability.
decision variable Zℴijk for maintenance planning, Eq. (17) We adopt a two-stage procedure comprising a single-
is reformulated: objective stage S1 and a multi-objective stage S2. The single-
( ) objective S1 represents a preliminary stage for generating a
Sℴi� j� k� ≥ Cℴijk + tsetup ⋅ 𝜃strp,ℴ� i� j� + tPdM ⋅ Zℴijk − 1 − Yℴij,ℴi� j� ,k ⋅ B
(36) high-quality initial population for the S2, aiming at enhanc-
ing the convergence speed of the multi-objective stage [28],
where tPdM denotes the duration of one PdM action. The which represents the main phase where both production
calculation of Cmax of the integrated schedule requires a scheduling and maintenance planning are jointly optimized.
reformulation from Eq. (19):
( ) 5.1 Single‑objective Stage S1
Cmax = max Cℴi + tPdM ⋅ Zℴioi k ∀ℴ, i, j, k, (37)

to consider the case where a PdM action is scheduled after S1 starts by randomly initializing a population and cloning
the last operation Oℴioi of a job Jℴi . All other constraints it into two populations, referred to as PopulationS1(1) and
presented in previous sections remain unchanged for the PopulationS1(2). Each chromosome corresponds to a can-
integrated problem. Thus, the MIFJSSP comprises follow- didate solution and represents a PS. These populations are
ing decision variables: evolved by two single-objective genetic algorithms with fit-
ness functions fS1(1) and fS1(2), GA-S1(1) using the makespan
• Yℴij,ℴ′ i′ j′ ,k for operation sequencing (Eq. 9), criterion min Cmax and GA-S1(2) minimizing the total tardi-
• Xℴijk for machine assignment (Eq. 8), and ness, i.e. min Ttotal . The latest generations of these popula-
• Zℴijk for PdM scheduling (Eq. 28). tions are then merged into a high-quality initial population

13
74 Production Engineering (2022) 16:65–88

Fig. 4  Flowchart of the

proposed TSGA with a
single-objective stage S1, b
multi-objective stage S2 and
c interface between the PHM
module and GA-S2

for S2 , referred to as PopulationS2 , comprising PopSizeS2 fS2 = wPS ⋅ fPS + wMP ⋅ fMP (44)
chromosomes. Thus, it is ensured that the search process of
S2 will start in promising regions of the genetic search space. where fPS and fMP refer to the fitness functions of the PS
In order to enhance population diversity and prevent prema- and MP, their corresponding weights wPS and wMP (with
ture convergence, 10% of PopSizeS2 are arbitrarily selected wPS + wMP = 1) and represent the input data of the TSGA.
and replaced with randomly generated chromosomes. The fitness function of the PS comprises the minimization
of makespan Cmax , total tardiness Ttotal and total cost of PS
5.2 Multi‑objective stage S2 𝜁PS , which were formulated in Sect. 4.5.1. As these objec-
tives have different units, their simultaneous optimization
The second stage aims at the multi-objective optimization using the weighted sum method is not possible. Thus, the
of both production and maintenance using following fitness scaling method proposed by [29] is adopted in order to con-
function: stitute a linear fitness function for the PS and MP:

13
Production Engineering (2022) 16:65–88 75

� �
fPS = Cmax + Ttotal + 𝜁PS
�
(45) those of the training data under the same OpReg and
health state.
� � �
fMP = Δtotal + Δcritical + 𝜁MP (46) The operation-specific machine degradation evaluation of
alternative schedules requires the joint deployment of these
where Cmax , Ttotal and 𝜁PS are scaled fitness values. Here,
′ ′ ′

models. In order to consider the operation-specificity of the

each objective function value fitness of a chromosome,
machine degradation process, the adopted PHM Module clus-
denoted as C , will be converted to a scaled value fitness as
′

ters CM data into operation specific clusters, i.e. OpRegs. A

follows:
� �
⎧ fitness(C) − MIN Genz � � � �
⎪ � � � � , if MAX Genz ≠ MIN Genz
�
fitness (C) = ⎨ MAX Genz − MIN Genz (47)
⎪
⎩ 0 , otherwise

where MIN(G) and MAX(G) correspond to the fitness values sequence of OpReg is denoted as OpRegSeq and mirror the
of the fittest and least fit chromosomes within the chromo- operations necessary to manufacture a product—in other
some’s generation Genz , respectively. words, candidate PSs can be translated to OpRegSeq which are
Hereby, each schedule comprises m job operation then evaluated by the PHM Module to obtain changes in HIk .
sequences, i.e. one for each machine Mk of the flexible job The main steps of the predictive simulation are summarized
shop. The adopted PHM Module simulates each sequence in the following, which occur recursively for each timestep t :
regarding the following:
• The DS-CVAE uses the information of the active OpReg
• current health state of the machine, expressed by HIk , and of the candidate production schedule (PS) by the TSGA
• the production history, i.e. operation sequence previously and the health condition HIk (t) to generate synthetic sen-
executed on the machine. sor observations.
• The sensor observations generated by the DS-CVAE are
Here, the latter is expressed in form of HIk history of the passed to the HA-CVAE for health assessment. Its output
machine, denoted as 𝐇𝐈k,history . Hence, assuming a TSGA is the predicted HIk,pred (t + 1).
population with PopSize chromosomes, the algorithm
requires m × PopSize predictive simulations per generation. The described procedure is illustrated in Appendix Fig. 6.
Interface Between the TSGA and PHM Module The
5.3 Interface to the PHM module feedback regarding the health states of machines is consid-
ered by the TSGA within the fitness function fMP of the
The application of the proposed approach enables the deci- multi-objective stage S2 . As expressed in Eq. (46), this
sion maker (scheduler) to consider the machines’ health function comprises the minimization of total degradation
states during production planning, and thus to simultane- of machines Δtotal , critical machine degradation Δcritical and
ously optimize the PP and MP. As visualized in Fig. 4c, the total cost of MP 𝜁MP . Considering the Eqs. (29), (34) and
the proposed TSGA and the PHM Module have an itera- (35) of these objectives, the following can be identified as
tive exchange of information during the entire optimization the output data required by the PHM Module:
process.
Functionality and Architecture of the PHM Module • predicted health indicator HI k,pred,
The PHM Module simulates candidate production schedules, • predicted timestep tk,safe in which the health state will
i.e. chromosomes, generated
{ by the TSGA and} forecasts the reach HI k,safe, and
future HIk trajectory HIk (t), HIk (t + 1), … based on can- • predicted timestep tk,fail in which the health state will
didate schedules. It comprises two generative deep learn- reach HI k,fail.
ing models, namely conditional variational autoencoders
(CVAE) as modelled by Zhai et al. [4]: In correspondence, the PHM Module requires the follow-
ing input data:
• a health assessor model (HA-CVAE), used for deriving
a health indicator (HI) of a machine, and • current health state expressed by HIk,
• a data simulator model (DS-CVAE), which generates • job operation sequence of the machine in form of
realistic synthetic multivariate CM data that resemble OpRegSeqk , and

13
76 Production Engineering (2022) 16:65–88

• the health indicator history HIk,history up to the time of neighborhood and flip mutation operators with mutation rate
prediction, i.e. the current timestep. pm [30], as well as binary flip mutation operator designed
by the present work for the MP segment. The OS segment
The TSGA and PHM Module interact through a defined is executed by the swapping and neighborhood mutation
interface realized by the function simulate_schedules , of operators. The general procedure adopted for the mutation of
which the pseudo-code is shown in Appendix Algorithm 3. generations is analog to the crossover process. Thus, either
Each chromosome corresponds to m simulations, i.e. the the OS and MA segments are mutated with a probability
number of machines in the flexible job shop. When all pre- of pm.
dictive simulations for every machine are completed, TSGA Fitness evaluation and selection operators Selection
calculates the total degradation of machines Δtotal , critical operators are responsible for choosing the parent chromo-
machine degradation Δcritical and the total cost of MP 𝜁MP of somes from a generation, denoted as Genz , that will produce
the chromosome. the offspring chromosomes for the next generation Genz+1.
Thus, the result of the selection process is a set of chromo-
5.4 Genetic Operators somes, referred to as the parental population, for the crosso-
ver process. The proposed TSGA adopts the selection pro-
Chromosome Encoding Chromosomes correspond to can- cedure of Li and Gao [30] using the elitist and tournament
didate solutions to the MIFJSSP, i.e. production schedules selection operator.
with PdM actions. The proposed TSGA adopts a chromo- The described selection procedure is summarized as
some representation with three segments as visualized in pseudo-code in Appendix Algorithm 2. The combined
Appendix Figs. 7, 8 and 9. Each segment is a vector con- application of the elitist and tournament selection operators
sisting of integers and corresponds to a subproblem of the allows the TSGA to make a trade-off between exploration
MIFJSSP (see Sect. 4.1) as follows: and exploitation of the genetic search space. In order to pre-
vent the loss of the best-found solutions between genera-
• operation sequencing segment OS, tions, an external archive is employed based on publications
• machine assignment segment MA, [33, 34]. For a population with PopSize chromosomes, the
• maintenance planning segment MP. archive contains archiveRate * PopSize of the best chromo-
somes found by the entire TSGA generations. In each TSGA
OS and MA are adopted from [30, 31] and [32]. To include iteration, the chromosomes stored in the archive are added to
the maintenance planning subproblem, this representation the population of the most recent generation Genz.
is extended by MP . Each segment uses an operation-based Chromosome decoding By decoding, the chromosome
encoding and comprises N genes, where N is the total num- is translated to the PdM integrated schedule. We adopt the
ber of job operations. By choosing and applying suitable priority-based decoding and reordering by [31]. The decod-
genetic operators for each segment, the creation of infeasible ing process is carried out by scanning the OS segment of the
chromosomes during iterations are avoided and no repair chromosome from left to right, with the machine-assignment
mechanism is needed. information for each operation provided by MA. Finally, the
Crossover operators The precedence operation crosso- MP segment specifies whether the operation is followed by
ver (POX), job-based crossover (JBX), and 2-point crossover a PdM action. The main steps of the decoding process are
(2PX) adopted from Li & Gao [30] are applied in the TSGA. summarized in Appendix Algorithm 4.
For the OS segment, the TSGA employs either the POX or Termination The termination of both stages S1 and S2
JBX operator, where one is selected randomly. The MA and use the following criteria:
MP segments are executed by the 2PX operator. In S2, the
TSGA chooses randomly whether to crossover the OS and • the convergence of the optimization process, i.e. the
MA segments or the MP . In other words, at each iteration of allowed maximum of generations with stagnation of the
the TSGA, either the production plan (represented by the OS fitness value, and
and MA segments) or the maintenance plan (represented by • the allowed maximum number of generations, denoted
the MP segment) is altered in order the generate offsprings as maxGenS1 and maxGenS2.
for the next generation. The described crossover procedure
in the multi-objective stage S2 is summarized as pseudo- Note that due to the stochastic nature of the health
code in Appendix Algorithm 1. predictions in fS2 , the fitness of S2 will not converge in a
Mutation operators Mutation operators are applied to monotonic decreasing manner. Thus, the termination cri-
the new generation Genz generated by the crossover pro- teria will be based on the average total cost of production
cedure. In each new TSGA generation, a chromosome is schedule and maintenance plan, makespan and tardiness
selected randomly. The proposed TSGA uses swapping,

13
Production Engineering (2022) 16:65–88 77

𝜁PS + 𝜁MP + Ttotal + Cmax for the 10 fittest chromosomes in Table 2  TSGA hyperparameter setting for case studies
each generation. Maxi- Maximum Popula- Crosso- Mutation Reproduc-
mum gen- stagnation tion size ver rate rate pm tion rate
eration PopSize pc (elitism) pr
6 Computational results MaxGen

S1: 300, 0.1 ∗ MaxGen300 0.90 0.20 0.05

In this section, the results of computational experiments of S2: 100
the TSGA are presented. The objectives of the computa-
tional experiments are as follows:
respective optimization metrics and the associated total
• check the plausibility of the implemented code, i.e. deter- cost of the schedule and not visualize every schedule, since
mine whether the implemented algorithms are capable of the visualizations do not offer any more information than
generating feasible solutions for FJSSP instances, already displayed in Fig. 5.
• investigate the performance of the different components
of the proposed algorithm, 6.1 Case 1: Partial MIFJSSP with simulated dataset
• calibrate the hyperparameters of the implemented algo-
rithm, and In order to assess the applicability of the developed algo-
• investigate the convergence of the search process. rithm, an exemplary PdM-IPS instance is studied in this
section. Here, a partial flexible job shop is considered using
One of the common challenges in the field of PdM inte- the NASA’s simulated C-MAPSS dataset [36]. Note that
grated scheduling is that there are no benchmark data avail- this dataset comprises run to failure CM data of jet engines
able for the testing of proposed algorithms [15]. Thus, for and not manufacturing data. Therefore, virtual products and
the following experiments in this section, a commonly used operations, as well as corresponding machines that mirror
FJSSP benchmark dataset by Brandimarte [35] was adopted. the jet engine’s degradation behavior are generated.
The used test instance (“Mk01”) belongs to a 10 × 15 FJSSP Dataset description Case 1 adopts the C-MAPSS
with 10 machines and 15 jobs. The test instance comprises FD002 sub-dataset.
{ The dataset was } clustered into a total
a total of 55 operations and is aimed at makespan mini- of 6 OpRegs: OpReg1 , … , OpReg6 . Accordingly, the
mization. Thus, for the following experiments in this sec- OpRegSeq of an operation may comprise several operating
tion, a simple single-stage structure of the proposed GA regimes, i.e. consists of several sub-operations with differ-
is considered, which corresponds to the GA-S1(1) with ent degrading effects on the machines. At each time step,
fS1(1) = min Cmax of the TSGA. the sensor observation was mapped to the currently active
For the hyperparameter optimization, a full-factorial OpReg. The HA-CVAE and DS-CVAE models were trained
experimental design with four levels was used, see Appendix and supplied by [4]. The results presented in this section are
Table 8. Based on these experiments, the setting in Table 2 obtained by integrating the trained models into the devel-
is chosen for the application of the proposed TSGA in the oped TSGA.
following sections: Case description The OpRegs identified for the
Note that the reported computational study of hyperpa- C-MAPSS are used by the present work to generate virtual
rameters was conducted using the GA-S1(1) in the single- products and operations. An exemplary middle-sized PdM-
objective stage S1 of the proposed TSGA. However, all GAs IPS instance was created considering a partial flexible job
of the TSGA are implemented using the same chromosome shop:
representation and genetic operators. Furthermore, the cho- Machine Data. The partial flexible job shop comprises 5
sen test instance of Brandimarte [35] exhibits a problem size multifunctional machines. It is assumed that the machines
which is representative of the case studies in the following are identical, and thus have the same HIk,safe and HIk,fail val-
chapters. Hence, the results of these computational experi- ues. However, they show different levels of machine health
ments can be transferred to the TSGA. In order to improve HIk (t = 0) at the time of planning, i.e. are degraded to
the usability of the approach, a graphical user interface varying degrees, see Appendix Table 9. The PHM Module
(GUI) was developed. Job shop data, as well as parameters requires the historical records of operational conditions up
for cost and time and TSGA hyperparameters can be easily to the time of prediction. Thus, five time series with varying
set. In Fig. 5, the GUI shows an exemplary PdM integrated machine degradation rates were selected from the training
production schedule with corresponding costs. dataset used for the CVAE models. The selected time series
Next, a simulated dataset and a real industrial dataset are correspond to the HIk,history of machines.
applied to investigate the general applicability of the algo- Product Data The product portfolio 𝐏 of the problem
rithm. For the sake of brevity, we will focus on reporting instance comprises 5 product variants Pv , corresponding

13
78 Production Engineering (2022) 16:65–88

Fig. 5  Exemplary visualization of a PdM integrated production schedule

to a total of 16 operations Ovj . A detailed overview of the fittest chromosome represents the PdM integrated schedule
processing times pvj and eligible machine sets 𝐌vj of these selected by the TSGA with Cmax=230, Ttotal=22, 𝜉PP=1480
operations is given in Appendix Table 10. and 𝜉MP=1600.
Product Order Data The problem comprises 8 produc- The total cost of the selected schedule corresponds to
tion orders POℴ , corresponding to 20 jobs Jℴi with a total 3080€, composed of setup costs (1150€), transportation
of 61 job operations Jℴi . The product variants Pv , due dates costs (330€) and cost of maintenance (1600€). (see Appen-
dℴ and quantities qℴ of these production orders are given in dix Table 12 for cost positions). As shown in Table 3, each
Appendix Table 11. of these PdM actions were scheduled with a starting time
Cost and Time Parameters The cost and time parameters tk,PdM between tk,safe and tk,fail , thus constituting no advance-
of PdM actions, setup actions between product variants and ment costs. Table 3 also shows the predicted health condi-
transportation of jobs are given in Appendix Table 12. tions of machines at time thorizon , i.e. after the execution of
Application of the TSGA The hyperparameters of S1 the schedule. The critical machine degradation Δmax imposed
are given in Table 2. For S2 , the number of iterations was by the integrated schedule is printed in bold and belongs to
set to MaxGenS2 = 100 . The production plan and main- machine M2.
tenance plans are weighted with same importance, i.e. The predictions indicate that the machines M1 and M4
wPP = wMP = 0.5. The algorithm was run on a macOS sys- do not reach their failure states. M1 reaches its HI1,safe value
tem with 2.4 GHz Intel Core i5 processor and 8 GB RAM. and was accordingly scheduled by the TSGA to be main-
As described in Sect. 5.1, S1 evolves two populations tained after the completion of its job operations. Although
in parallel, namely PopulationS1(1) and PopulationS1(2) . M2 exhibits the lowest HIk (i.e. worst health condition), the
For the MIFJSSP instance studied in this case, the evolv- PHM Module predicts that M5 will reach its HI5,safe before
ing process of both populations were terminated due to the M2 reaches HI2,safe. However, the time period between t5,safe
convergence criterion after 110.48 seconds. In particular, the and t5,fail is longer than it is for M2 . Furthermore, M4 is
GA-S1(1) terminated after 191 generations with a makespan expected to degrade the least. It can be concluded that the
Cmax = 191, whereas GA-S1(2) terminated after 94 generation PHM Module is able to assess and predict operation-spe-
and was able to minimize the total tardiness to Ttotal = 0. cific machine degradation based on the assigned operations.
The most recent generations of PopulationS1(1) Using this predictive information, the TSGA schedules the
and PopulationS1(2) were merged by the TSGA to a necessary PdM actions so that (1) the machines do not reach
high-quality initial population PopulationS2 , which their failure states at time tk,fail and (2) the multi-objective
evolved for MaxGenS2 = 100 iterations in 567.0 s. The fS2 of the schedule is minimized. The analysis shows that the

13
Production Engineering (2022) 16:65–88 79

scheduling results of the implemented TSGA are effective The hyperparameters of the TSGA are configured as in
and reasonable. the previous Case 1. The algorithms were run on a Windows
system with 3.3 GHz Intel Xeon processor and 64 GB RAM.
6.2 Case 2: Total MIFJSSP with industrial CM data The single-objective stage S1 was terminated after
416.11 seconds. The GA-S1(1) terminated after 145 gen-
Dataset Description For the following case, a real indus- erations due to the convergence with a makespan value of
trial dataset from an automotive manufacturing company is Cmax = 526 . The GA-S1(2) terminated after 99 generations
applied. Note that due to confidentiality, raw data and met- with a total tardiness of Ttotal = 0 . To showcase the effec-
rics associated with efficiency and costs cannot be shown. tiveness of the archive, a simulation run without archive
The dataset belongs to a machine park with five multifunc- was conducted. Without archive, GA-S1(1) converged after
tional machines and comprises CM data collected from 140 236 with Cmax = 575 , while GA-S1(2) terminated after
sensors covering a time period of 9 months. The signifi- 142 generations with Ttotal = 0. This clearly shows that the
cant differences in the real industrial dataset compared to archive approach ensures better results while requiring less
C-MAPSS are as follows: computation.
Table 4 shows the results of the PdM integrated schedule
• During the preprocessing step, a total of K = 32 OpRegs selected by the TSGA after S2 for the PdM-IPS instance
were defined. 15 of these belong to the manufacturing under study. The algorithm terminated after 1998.1 s. The
of a product, i.e. operating conditions during which the table gives an overview of the machines’ health conditions
machine experiences degradation. as predicted by the PHM Module and the starting times tk,PdM
• Compared to C-MAPSS, the OpRegs are defined high- of the PdM actions as scheduled by the TSGA. The critical
level (workpiece ID). Thus, OpRegSeq of each product machine degradation Δmax imposed by the integrated sched-
comprises a single operating regime. ule is printed in bold.
With a predicted machine degradation Δ1 = 0.1577 ,
Case Description machine M1 is expected to experience the highest degra-
Machine Data The flexible job shop comprises three dation imposed by the selected schedule. Another finding
identical multifunctional machines with different health is that the predicted machine degradations are significantly
conditions as listed in Appendix Table 13. lower than in Case 1, although the problem size is bigger
Product Data The product portfolio 𝐏 of the studied (i.e. the prediction horizon is longer). This may be expli-
problem instance comprises 10 product variants Pv. As men- cated by the fact that the OpRegs in the industrial dataset
tioned above, each OpReg corresponds to a product variant. have been defined on product level, leading to a generaliza-
Thus, each product has one operation Ov1 with a OpRegSeq tion of degradation processes and reduced accuracy of the
consisting of a single OpReg. The processing times pvj are operation-specific machine predictions.
listed in Appendix Table 14. The eligible machine sets 𝐌vj
are omitted as a total flexible job shop is studied, i.e. each
operation can be produced on any machine. 7 Conclusion and outlook
Product Order Data The problem comprises a total of
50 jobs Jℴi that belong to 12 production orders POℴ . The Conclusion
product variants Pv , due dates dℴ and quantities qℴ of these This publication presented the PdM-IPS approach in
production orders are given in Appendix Table 15. order to realize integrated production and maintenance
Cost and Time Parameters. The problem uses the same planning using operation-specific predictions from a PHM
cost and time parameters of Case 1 given in Appendix Module. After review of related literature, five main require-
Table 12. ments for value-adding IPSMP in industrial application were
Application of the TSGA identified:

Table 3  Predicted health ( )

Mk Δk HIk (t = 0) HIk,pred thorizon tk,safe tk,fail tk,PdM
conditions of machines in the
selected schedule in Case 1 M1 0.3481 0.7912 0.4431 163 – 170
M2 0.5036 0.6579 0.1543 58 86 66
M3 0.3551 0.7144 0.3593 80 118 90
M4 0.1095 0.8777 0.7682 – – –
M5 0.3926 0.7174 0.3248 37 118 85

13
80 Production Engineering (2022) 16:65–88

Table 4  Predicted health ( )

Mk Δk HIk (t = 0) HIk,pred thorizon tk,safe tk,fail tk,PdM
conditions of machines in the
selected schedule in Case 2 M1 0.1577 0.8963 0.7386 242 321 255
M2 0.0704 0.8503 0.7799 218 − 260
M3 0.1495 0.8406 0.6911 101 205 122

R1: Consideration of a JSSP or FJSSP in order to be learns the representation of data and accumulates degrada-
applicable for different shop floor layouts and not being tion non-linearly.
limited to single machine setups, R5: The modelling of the IPSMP based on a MIFJSSP
R2: consideration of the machine condition for mainte- ensures high applicability to different production layouts
nance planning, and is thus industrial viable. The applied PHM Module is
R3: operation-specific modeling of machine degradation, able to handle industrial-scale CM data. The application of
R4: consideration of the non-linearity, i.e. the operation the TSGA enables the timely generation of schedules while
sequence-dependency, of the machine degradation process, at the same time requiring reasonable computation time.
and Finally, the two use cases presented in Sect. 6 showcase
R5: use of real industrial data for the assessment and pre- that the approach indeed can be applied to an industrial set-
diction of machine condition. ting. Results indicate that the TSGA converges and critical
To fulfill these requirements, the presented PdM-IPS machine conditions and potential failures can be avoided
approach models the production scheduling part after the through scheduling PdM actions.
FJSSP, while the maintenance planning is based on non- To conclude, by considering the future machine condition
linear accumulation of degradation. The mathematical before the execution of a production schedule, the imple-
model was formulated and solved using a two-stage genetic mented approach enables the decision maker to avoid fail-
algorithm. Two case studies using both artificial and real ures due to machine degradation. Job operations and PdM
industrial data showcased that the approach is applicable to actions are jointly optimized regarding overall cost of the
different scenarios. Furthermore, results indicate that fea- manufacturing system, resulting in improved planning and
sible production and maintenance integrated schedules can cost efficiency. It is noteworthy, however, that the benefits
be retrieved. Given its nature, the genetic algorithm cannot of PdM-IPS depend on the shop floor layout of the manu-
guarantee a global optimum. However, results indicate that facturing company. In a flexible job shop layout, i.e. where
the fitness function significantly improves over generations jobs can be freely routed, the benefits are the biggest due
and does not converge prematurely and thus, (local) optima to the inherent flexibility of the system to schedule jobs on
concerning production and maintenance objectives are less degraded machines. In traditional flow shops where the
found. Following statements concerning the requirements routing of jobs is limited, PdM-IPS can still optimally sched-
can be made: ule maintenance and production, but has less room to con-
R1: The exemplary applications of the TSGA in Sect. 6 trol future degradation on machines. Together with recent
confirmed that the proposed algorithm is able to generate developments of “Industrie 4.0” that enables the so called
production plans for both partial and total flexible job shops. matrix production with flexible production cells [37], which
R2: By adopting a three-segment chromosome, each represent a flexible job shop, the relevance of PdM-IPS will
corresponding to a subproblem of the IPSMP, the TSGA further increase.
is able to simultaneously optimize both production and Outlook
maintenance. Thus, the starting times of job operations and Further research can and should be conducted on the basis
PdM actions are jointly scheduled with regard to the overall of this publication. Specifically, three major fields should be
performance. investigated upon.
R3 + R4: These requirements have been met by the pro- Handling of uncertainties. In this work, we consider
posed PdM-IPS approach through the integration of the deterministic times (e.g., deterministic processing and main-
unsupervised deep learning PHM Module developed by [4]. tenance times, new job arrivals, etc.). By considering uncer-
As described, this model was adopted by the present work tainties, i.e. non-deterministic times, robustness measures of
in order to obtain operation-specific predictions regarding the generated predictive schedules can be improved. Robust
machine deterioration. Each job operation to be scheduled by and stable schedules may help to decrease the costs due to
the developed TSGA corresponds to a sequence of OpRegs, unexpected events. Moreover, this would represent the first
that in turn is evaluated by the PHM Module according to step toward dynamic flexible job shop scheduling.
its operation-specific degradation. The deep learning model Determining HI thresholds for maintenance. This publi-
cation assumed the thresholds for HIk,safe and HIk,fail based

13
Production Engineering (2022) 16:65–88 81

on observation of degradation trends and expert knowledge. 8 Appendix

Future works should focus on an automated approach to
derive these thresholds based on data and study its effect The appendix consists of Tables 5, 6, 7, 8, 9, 10, 11, 12, 13,
and sensitivity on subsequent scheduling approaches. 14, 15, Figures 6, 7, 8, 9 and Algorithms 1, 2, 3, 4.
Development of a methodological approach for the evalu-
ation and validation of PdM frameworks. One of the major 8.1 Notations and variables
challenges in the field of data-driven PdM is the lack of
methodologies that enable the systematic evaluation of the
operational and monetary performance of a proposed PdM
approach. Future research needs to investigate which KPIs
are suitable for this purpose, allowing the benefits prom-
ised by a PdM approach to be quantified. This is especially
important for post-prognostic decision support approaches
like the presented PdM-IPS.

Table 5  Notations and variables of the subsystem "Machine"

Symbol Designation

HIk Health index of machine

Mk , HIk ∈ [0, 1.0]
k Index of machines
m ∈ ℤ+ Number of machines Mk in the
problem instance,|𝐌| = m
{ }
𝐌 = Mk ∶ k ∈ {1, … , m} Set of m ∈ ℤ+ machines Mk
Mk Machine
RULk Remaining useful life of machine
Mk in [timesteps]

Table 6  Notations and variables Symbol Designation

of the subsystem "Products and
Operations" j Running index of operations comprised by a product variant
𝐌vj ⊆ 𝐌 Eligible machine set (unordered) of operation Ovj
𝐎 = ∪v∈{1,…,P} 𝐎v Set of all operations of the problem instance
ov ∈ ℤ+ Number of operations of product v, ||𝐎v || = ov
{ { }}
𝐎v = Ovj ∶ j ∈ 1, , ov Ordered set of operations of product variant Pv
Ovj j-th operation of product v
pvjk Unit processing time of operation Ovj on machine Mk in [timesteps]
𝐏 Product portfolio
Pv Product
v∈𝐏 Index of product variants

13
82 Production Engineering (2022) 16:65–88

Table 7  Notations and variables Symbol Designation

of the subsystem "Production
Orders and Jobs" Cℴi Completion time of job Jℴi in [timestep]
Cℴij Completion time of job operation Oℴij in [timestep]
dℴ Due date of 𝐏𝐎ℴ
dℴi Due date of jobs Jℴi ∈ 𝐏𝐎ℴ
i Running index of the jobs comprised by a production order
𝐉 = ∪ℴ∈{1,…,Q} 𝐏𝐎ℴ Unordered set of all jobs of the problem instance
Jℴi job Belonging to 𝐏𝐎ℴ
n ∈ ℤ+ Total number of jobs Jℴi in the problem instance, |𝐉 | = n
N ∈ ℤ+ Total number of job operations in the problem instance
𝐎 Set of all job operations of the problem instance
ov(ℴ) ∈ ℤ+ Number of job operations comprised by job Jℴi ,
| |
Jℴi , ||Oℴi || = |Ov(ℴ) | = ov(ℴ)
| |
{ { }}
Oℴi = Oℴi ∶ j ∈ 1, … , ov(ℴ) Ordered set of operations of job Jℴi
Oij j-th operation of job Jℴi
pℴijk unit Processing time of operation Oℴij on machine Mk
𝐏𝐎ℴ Production order, i.e. set of qℴ jobs
Q Total number of 𝐏𝐎ℴ in the problem instance
qℴ Production quantity requested by 𝐏𝐎ℴ
Sℴij Starting time of operation Oℴij in [timestep]
v(ℴ) ∈ 𝐏 Product variant requested by POℴ

8.2 Algorithms

13
Production Engineering (2022) 16:65–88 83

Note that S1 adopts the same procedure, except the random

choice on line 7 is always True, i.e. lines 11 and 12 are never
visited. For implementation, [38] was adopted.

13
84 Production Engineering (2022) 16:65–88

8.3 Predicting Future Operation‑specific Health

Indicator

Training Two CVAEs are trained to predict the future health

indicator given a production schedule. The Data Simulator
(DS) is trained on all historical CM data of a run to failure,
i.e. from the maintenance action until the failure. The Health
Assessor (HA) is trained on “healthy” data only, i.e. the first
20% of data of the run to failure.
Application Both DS and HA are jointly applied. The
decoder part of the DS-CVAE is used as a generative model
to create simulated sensor signals (i.e. CM data) conditioned
on a given future production sequence. This simulated data
is in turn assessed by the HA-CVAE to derive an operation-
specific HI prediction given a future production schedule
(see Fig. 6).

Fig. 6  Health assessment and prediction procedure as outlined by [4]

13
Production Engineering (2022) 16:65–88 85

Fig. 7  Encoding of the OS seg-

1 1 3 1 2 3 3 1 2 1 1 2 1 2 2 1 3 1 0 0 0 1 0 1 0 0 0
ment of chromosomes

1 1 3 1 2 3 3 1 2

111 112 311 113 211 312 313 114 212

Fig. 8  Encoding of the MA seg-

1 1 3 1 2 3 3 1 2 1 1 2 1 2 2 1 3 1 0 0 0 1 0 1 0 0 0
ment of chromosomes

= 1,
1 1 2 1 2 2 1 3 1 211 3

211 is assigned to 3

111 112 113 114 211 212 311 312 313

11 21 31

Fig. 9  Encoding of the MP seg-

ment of chromosomes

8.4 GA Encoding binary encoding proposed by [7] for maintenance planning

(see Fig. 9). An allele of 1 means that a PdM action sched-
Encoding of the 𝐎𝐒 segment The OS segment of the chro- uled after the completion of the job operation, directly cor-
mosome adopts a permutation-based representation as visu- responding to the to the decision variable Zℴijk.
alized in Fig. 7.
The operation sequencing decision variables Yℴij,ℴ′ i′ j′ ,k are 8.5 Selected Hyperparameters
implicitly expressed in this segment and are to be decided
by the decoding algorithm. By scanning through the OS seg- The selected values of these hyperparameters are those often
ment from left to right, the j-th appearance of an index refers encountered in the reviewed literature and are listed in the
to the j-th operation Oℴij of the job Jℴi , to which the index following Table 8.
belongs. The main advantage of this representation is that 11 possible combinations of the shown levels of each
each chromosome represents a feasible schedule [31]. Fur- hyperparameter were tested. Each setting was run five times
thermore, the precedence constraints between operations are by the algorithm. The computations have been performed on
not explicitly expressed in the chromosomes. The decoding a macOS High Sierra [Version 10.13.6] operating system
algorithm tracks the number of appearances of an index in with 2.4 GHz Intel Core i5 processor and 8 GB RAM. The
the OS segment and determines the corresponding job opera- behavior of the algorithm in each run was recorded with
tion for each gene. regard to the following performance measures:
Encoding of the 𝐌𝐀 segment The MA segment con-
tains information concerning the machine assignments of Table 8  GA hyperparameter levels
job operations. It adopts a partitioned permutation of job
indices as visualized in Fig. 8. Hyperparameter Level 1 Level 2 Level 3 Level 4
Encoding of the 𝐌𝐏 segment The MP segment con- MaxGen 100 200 300 400
tains information regarding the PdM actions and adopts the PopSize 100 200 300 400
pc 0.75 0.80 0.85 0.90
pm 0.05 0.10 0.15 0.20
pr 0.005 0.01 0.05 0.10

13
86 Production Engineering (2022) 16:65–88

Table 9  Machine data for Case 1 • obtained objective function value, i.e. the makespan in
Mk HIk (t = 0) HIk,fail HIk,safe timesteps and
• the computational run time in seconds.
M1 0.7912 0.45 0.55
M2 0.6579 0.45 0.55 8.6 Machine and Product Data for Case 1 & 2
M3 0.7144 0.45 0.55
M4 0.8777 0.45 0.55 The cost and time parameters used in case 1 are listed in the
M5 0.7174 0.45 0.55 following. Note that the same parameters were also used for
case 2 (see Table 12).

Table 10  Product data in Case 1 Product variant Pv Operation Ovj Processing time pvj Eligible Machine Set 𝐌vj
{ }
P1 O11 p11 = 10 𝐌11 = M1 , M4 , M5
{ }
O12 p12 = 8 𝐌12 = M1 , M2 , M3
{ }
O13 p13 = 10 𝐌13 = M1 , M2 , M3
{ }
O14 p14 = 4 𝐌14 = M1 , M3
{ }
P2 O21 p21 = 8 𝐌21 = M1 , M2 , M3 , M4 , M5
{ }
O22 p22 = 10 𝐌22 = M1 , M2 , M3 , M4 , M5
{ }
O23 p23 = 6 𝐌23 = M1 , M3 , M4
{ }
P3 O31 p31 = 7 𝐌31 = M1 , M2 , M3
{ }
O32 p32 = 9 𝐌32 = M1 , M5
{ }
O33 p33 = 10 𝐌33 = M1 , M2 , M3
{ }
P4 O41 p41 = 9 𝐌41 = M1 , , M4 , M5
{ }
O42 p42 = 10 𝐌42 = M1 , M2 , M3
{ }
O43 p43 = 9 𝐌43 = M1 , M2 , M3
{ }
P5 O51 p51 = 11 𝐌51 = M1 , M2 , M3 , M4 , M5
{ }
O52 p52 = 9 𝐌52 = M1 , M2 , M3
{ }
O53 p53 = 10 𝐌53 = M1 , M5

Table 11  Product order data in Case 2: Total FJSSP with a real industrial dataset
POℴ v(ℴ) qℴ dℴ
Case 1 Detailed data of the PdM-FJSSP instance studied in Case
PO1 P1 2 100 2 (see Sect. 6.2) are given below. Note that the indices opera-
PO2 P2 4 150 tions and processing durations are separated by a comma,
PO3 P1 2 200 as some comprise two digits in this case (see Table 13, 14
PO4 P3 2 240 and 15).
PO5 P4 1 240
PO6 P5 2 280
PO7 P3 3 320 Table 13  Machine data for Mk HIk (t = 0) HIk,safe HIk,fail
PO8 P2 4 400 Case 2
M1 0.8963 0.80 0.75
M2 0.8503 0.80 0.75
M3 0.8406 0.80 0.75
Table 12  Cost and time parameters in Case 1 and 2
Cost component Time parameter Cost parameter [€/action]
[timesteps/action]

PdM action tPdM = 60 costPdM,fix = 400

costPdM,adv = 50
Setup tsetup = 15 costsetup = 50
Transportation ttransport = 10 costtransport = 10

13
Production Engineering (2022) 16:65–88 87

Table 14  Product data of Case 2 Index Pv Ovj pvj

[1] P1 O1,1 t1,1 = 15

[2] P2 O2,1 t2,1 = 12
[3] P3 O3,1 t3,1 = 18
[4] P4 O4,1 t4,1 = 12
[5] P5 O5,1 t5,1 = 22
[6] P6 O6,1 t6,1 = 10
[7] P7 O7,1 t7,1 = 14
[8] P8 O8,1 t8,1 = 10
[9] P9 O9,1 t9,1 = 15
[10] P10 O10,1 t10,1 = 22

Table 15  Product order data of POℴ v(ℴ) qℴ dℴ POℴ v(ℴ) qℴ dℴ

Case 2
PO1 P1 8 140 PO7 P8 3 460
PO2 P2 4 200 PO8 P6 2 460
PO3 P3 2 280 PO9 P2 3 480
PO4 P4 4 350 PO10 P3 5 500
PO5 P7 6 380 PO11 P9 3 500
PO6 P5 2 400 PO12 P10 8 530

Funding Open Access funding enabled and organized by Projekt health prognostics with generative deep learning. J Manuf Syst
DEAL. 104:284. https://​doi.​org/​10.​1016/j.​jmsy.​2021.​02.​006
5. Ghaleb M, Taghipour S, Sharifi M et al (2020) Integrated produc-
Open Access This article is licensed under a Creative Commons Attri- tion and maintenance scheduling for a single degrading machine
bution 4.0 International License, which permits use, sharing, adapta- with deterioration-based failures. Comput Ind Eng. https://d​ oi.o​ rg/​
tion, distribution and reproduction in any medium or format, as long 10.​1016/j.​cie.​2020.​106432
as you give appropriate credit to the original author(s) and the source, 6. Cassady CR, Kutanoglu E (2005) Integrating preventive mainte-
provide a link to the Creative Commons licence, and indicate if changes nance planning and production scheduling for a single machine.
were made. The images or other third party material in this article are IEEE Trans Rel 54:304–309. https://​doi.​org/​10.​1109/​TR.​2005.​
included in the article’s Creative Commons licence, unless indicated 845967
otherwise in a credit line to the material. If material is not included in 7. Ye J, Ma H (2015) Multiobjective joint optimization of produc-
the article’s Creative Commons licence and your intended use is not tion scheduling and maintenance planning in the flexible job-shop
permitted by statutory regulation or exceeds the permitted use, you will problem. Math Probl Eng 2015:1–9. https://d​ oi.o​ rg/1​ 0.1​ 155/2​ 015/​
need to obtain permission directly from the copyright holder. To view a 725460
copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. 8. Chen X, Xiao L, Zhang X (2014) A production scheduling prob-
lem considering random failure and imperfect preventive mainte-
nance. Proc Instit Mech Eng Part O: J Risk Reliab 229(1):26–35.
https://​doi.​org/​10.​1177/​17480​06X14​545834
References 9. Bougacha O, Varnier C, Zerhouni N et al. (2019) Integrated Pro-
duction and Predictive Maintenance Planning based on Prognostic
1. Zhai S, Achatz S, Groher M et al. (2020) An empirical expert Information. In: 2019 International Conference on advanced sys-
study on the status quo and potential of predictive maintenance tems and emergent technologies (IC_ASET). IEEE, pp 363–368
in industry. In: International Conference on Sensing, Diagnostics, 10. Sloan TW, Shanthikumar JG (2010) Combined production and
Prognostics, and Control (SDPC). https://d​ oi.o​ rg/1​ 0.1​ 109/S
​ DPC4​ maintenance scheduling for a multiple-product, single-machine
9476.​2020.​93531​77 production system. Prod Oper Manag 9:379–399. https://​doi.​org/​
2. Iyer N, Goebel K, Bonissone P (2006) Framework for Post-Prog- 10.​1111/j.​1937-​5956.​2000.​tb004​65.x
nostic Decision Support. In: 2006 IEEE Aerospace Conference 11. Aramon Bajestani M, Banjevic D, Beck JC (2014) Integrated
proceedings: March 4–11, 2006, Big Sky, MT. [IEEE], [Piscata- maintenance planning and production scheduling with Markovian
way, N.J.], pp 1–10 deteriorating machine conditions. Int J Prod Res 52:7377–7400.
3. Zhai S, Reinhart G (2018) Predictive maintenance als Wegbere- https://​doi.​org/​10.​1080/​00207​543.​2014.​931609
iter für die instandhaltungsgerechte Produktionssteuerung. ZWF 12. Pan E, Liao W, Xi L (2012) A joint model of production sched-
113:298–301. https://​doi.​org/​10.​3139/​104.​111912 uling and predictive maintenance for minimizing job tardiness.
4. Zhai S, Gehring B, Reinhart G (2021) Enabling predictive main- Int J Adv Manuf Technol 60:1049–1061. https://​doi.​org/​10.​1007/​
tenance integrated production scheduling by operation-specific s00170-​011-​3652-4
13. Fitouri C, Fnaiech N, Varnier C et al (2016) A Decison-making
approach for job shop scheduling with job depending degradation

13
88 Production Engineering (2022) 16:65–88

and predictive maintenance. IFAC-PapersOnLine 49:1490–1495. 27. Benbouzid-Sitayeb F, Guebli SA, Bessadi Y et al (2011) Joint
https://​doi.​org/​10.​1016/j.​ifacol.​2016.​07.​782 scheduling of jobs and Preventive Maintenance operations in the
14. Ladj A, Benbouzid-Si Tayeb F, Varnier C et al (2017) A hybrid flowshop sequencing problem: a resolution with sequential and
of variable neighbor search and fuzzy logic for the permutation integrated strategies. IJMR 6:30–48. https://​doi.​org/​10.​1504/​
flowshop scheduling problem with predictive maintenance. Proc IJMR.​2011.​037912
Comput Sci 112:663–672. https://​doi.​org/​10.​1016/j.​procs.​2017.​ 28. Rooyani D, Defersha FM (2019) An efficient two-stage genetic
08.​120 algorithm for flexible job-shop scheduling. IFAC-PapersOnLine
15. Ladj A, Tayeb FB-S, Varnier C et al. (2019) Improved genetic 52:2519–2524. https://​doi.​org/​10.​1016/j.​ifacol.​2019.​11.​585
algorithm for the fuzzy flowshop scheduling problem with pre- 29. Gao J, Gen M, Sun L (2006) Scheduling jobs and maintenances in
dictive maintenance planning. In: 2019 IEEE 28th International flexible job shop with a hybrid genetic algorithm. J Intell Manuf
Symposium on industrial electronics (ISIE). IEEE, pp 1300–1305 17:493–507. https://​doi.​org/​10.​1007/​s10845-​005-​0021-x
16. Benaggoune K, Meraghni S, Ma J et al. (2020) Post prognostic 30. Li X, Gao L (2016) An effective hybrid genetic algorithm and tabu
decision for predictive maintenance planning with remaining use- search for flexible job shop scheduling problem. Int J Prod Econ
ful life uncertainty. In: 2020 Prognostics and Health Management 174:93–110. https://​doi.​org/​10.​1016/j.​ijpe.​2016.​01.​016
Conference (PHM-Besançon). IEEE, pp 194–199 31. Gao J, Sun L, Gen M (2008) A hybrid genetic and variable neigh-
17. Zhai S, Riess A, Reinhart G (2019) Formulation and solution for borhood descent algorithm for flexible job shop scheduling prob-
the predictive maintenance integrated job shop scheduling prob- lems. Comput Oper Res 35:2892–2907. https://​doi.​org/​10.​1016/j.​
lem, pp 1–8. https://​doi.​org/​10.​1109/​ICPHM.​2019.​88193​97 cor.​2007.​01.​001
18. Morariu C, Morariu O, Răileanu S et al (2020) Machine learning 32. Gen M, Tsujimura Y, Kubota E (1994) Solving job-shop sched-
for predictive scheduling and resource allocation in large scale uling problems by genetic algorithm. In: Proceedings of IEEE
manufacturing systems. Comput Ind 120:103244. https://​doi.​org/​ International Conference on systems, man and cybernetics. IEEE,
10.​1016/j.​compi​nd.​2020.​103244 pp 1577–1582
19. Paprocka I, Kempa WM, Skołud B (2020) Predictive maintenance 33. Zhang G, Zhang L, Song X et al (2019) A variable neighborhood
scheduling with reliability characteristics depending on the phase search based genetic algorithm for flexible job shop scheduling
of the machine life cycle. Eng Optim 31:1–19. https://​doi.​org/​10.​ problem. Cluster Comput 22:11561–11572. https://​doi.​org/​10.​
1080/​03052​15X.​2020.​17140​41 1007/​s10586-​017-​1420-4
20. Denkena B, Dittrich M-A, Keunecke L et al (2020) Continuous 34. Huang X, Guan Z, Yang L (2018) An effective hybrid algorithm
modelling of machine tool failure durations for improved produc- for multi-objective flexible job-shop scheduling problem. Adv
tion scheduling. Prod Eng Res Devel 14:207–215. https://​doi.​org/​ Mech Eng 10:168781401880144. https://​doi.​org/​10.​1177/​16878​
10.​1007/​s11740-​020-​00955-y 14018​801442
21. Khatab A, Diallo C, Aghezzaf E-H et al (2019) Integrated produc- 35. Brandimarte P (1993) Routing and scheduling in a flexible job
tion quality and condition-based maintenance optimisation for a shop by tabu search. Ann Oper Res 41:157–183. https://​doi.​org/​
stochastically deteriorating manufacturing system. Int J Prod Res 10.​1007/​BF020​23073
57:2480–2497. https://​doi.​org/​10.​1080/​00207​543.​2018.​15210​21 36. Saxena A, Goebel K (2008) turbofan engine degradation simula-
22. Zandieh M, Khatami AR, Rahmati SHA (2017) Flexible job tion data set: NASA Ames Prognostics Data Repository. http://​ti.​
shop scheduling under condition-based maintenance: Improved arc.​nasa.​gov/​proje​ct/​progn​ostic-​data-​repos​itory. Accessed 15 Jan
version of imperialist competitive algorithm. Appl Soft Comput 2021
58:449–464. https://​doi.​org/​10.​1016/j.​asoc.​2017.​04.​060 37. KUKA Systems GmbH (2019) Industrie 4.0—Matrix production.
23. Xiang Y, Cassady CR, Jin T et al (2014) Joint production and https://​www.​kuka.​com/-/​media/​kuka-​downl​oads/​impor​ted/​2d0d7​
maintenance planning with machine deterioration and random c7d65​73491​a8b43​eb5f6​6aec5​96/​kuka_​matri​xprod​uktion_​scree​
yield. Int J Prod Res 52:1644–1657. https://d​ oi.o​ rg/1​ 0.1​ 080/0​ 0207​ npdfs_​en_​160419.​pdf. Accessed 10 Mar 2021
543.​2013.​843037 38. Bour G (2018) Python implementation of a genetic algorithm for
24. Ladj A, Varnier C, Tayeb FB-S (2016) IPro-GA: an integrated FJSP [source code]. https://​github.​com/​guill​aumeb​our/​flexi​ble-​
prognostic based GA for scheduling jobs and predictive mainte- job-​shop. Accessed 10 Mar 2021
nance in a single multifunctional machine. IFAC-PapersOnLine
49:1821–1826. https://​doi.​org/​10.​1016/j.​ifacol.​2016.​07.​847 Publisher’s Note Springer Nature remains neutral with regard to
25. Wang T (2010) Trajectory similarity based prediction for remain- jurisdictional claims in published maps and institutional affiliations.
ing useful life estimation. Dissertation, University of Cincinnati
26. Chaudhry IA, Khan AA (2016) A research survey: review of flex-
ible job shop scheduling techniques. Intl Trans in Op Res 23:551–
591. https://​doi.​org/​10.​1111/​itor.​12199

13