Supporting the Selection of Prognostic-based Decision Support
Methods in Manufacturing
Alexandros Bousdekis
, Babis Magoutas
, Dimitris Apostolou
and Gregoris Mentzas
School of Electrical and Computer Engineering, National Technical University of Athens, Athens, Greece
Department of Informatics, University of Piraeus, Piraeus, Greece
Keywords: Condition based Maintenance, Decision Tree Learning, Method Filtering, Decision Support.
Abstract: In manufacturing enterprises, maintenance is a significant contributor to the total company’s cost. Condition
Based Maintenance (CBM) relies on prognostic models and uses them to support maintenance decisions
based on the current and predicted health state of equipment. Although decision support for CBM is not an
extensively explored area, there exist methods which have been developed in order to deal with specific
challenges such as the need to cope with real-time information, to prognose the health state of equipment
and to continually update decision recommendations. We propose an approach for supporting analysts
selecting the most suitable combination(s) of methods for prognostic-based maintenance decision support
according to the requirements of a given maintenance application. Our approach is based on the ID3
decision tree learning algorithm and is applied in a maintenance scenario in the oil and gas industry.
In manufacturing enterprises, high reliability, low
environmental impact and safety of operations are
important issues for every industry (Peng, Dong, and
Zuo, 2010). Maintenance is a significant contributor
to the total company’s cost, so optimal maintenance
policy in terms of cost, equipment downtime and
quality should be identified (Garg, and Deshmukh,
2006). Condition Based Maintenance (CBM) is a
type of maintenance strategy, which relies on
diagnostic and prognostic models and uses them to
support decisions about the appropriate maintenance
actions based on the current health state of a system
through condition monitoring (Jardine, Lin, and
Banjevic, 2006). Condition monitoring in
manufacturing enterprises is increasingly realised
with equipment-installed sensors, which have the
capability of measuring with high frequencies a
multitude of parameters. This capability leads to
storage of a huge amount of data. Generating and
storing Big Data has become possible due to recent
developments in both hardware and data
management software (Zikopoulos, and Eaton,
Big Data-driven CBM poses challenges to
Decision Support Systems. These challenges are not
easily addressed within the complex manufacturing
environment, especially when dealing with
maintenance where several factors should be
considered simultaneously such as costs of
maintenance actions as a function of time, safety
issues and equipment degradation.
Although decision support for CBM is not an
extensively explored area, there exist several works
focusing on combinations of methods that can be
utilised for CBM decision support. Such methods
deal with real-time data which are gathered in high
frequency, develop prognostic models for the
estimation of Remaining Useful Life (RUL) or
Remaining Life Distribution (RLD) and provide
recommendations for maintenance. In the current
paper, we propose a practical approach for
supporting analysts to select the most suitable
combination(s) of methods for prognostic-based
maintenance on the basis of Big Data according to
the requirements of the application which they are
involved with.
Aiming to support prognostic-based maintenance
in various application domains and for a wide range
of functional and non-functional application
requirements, we follow a practical multistage
decision making approach. The basic idea of our
hierarchical approach is to break up the problem of
selecting the most suitable combination(s) of
Bousdekis A., Magoutas B., Apostolou D. and Mentzas G..
Supporting the Selection of Prognostic-based Decision Support Methods in Manufacturing.
DOI: 10.5220/0005372104870494
In Proceedings of the 17th International Conference on Enterprise Information Systems (ICEIS-2015), pages 487-494
ISBN: 978-989-758-096-3
2015 SCITEPRESS (Science and Technology Publications, Lda.)
methods for prognosis and prognostic-based
maintenance, into a union of several simpler
decisions about the suitability of the method
combination depending on the functional and non-
functional application requirements by developing a
Decision Tree (DT).
To formulate the information found in the papers
examined in our literature review in the form and
structure needed for feeding the DT learning
algorithm and building the DT, the following steps
were performed. First, generic categories of methods
for prognosis and prognostic-based maintenance
were defined in our effort to avoid specific method
extensions or variations used in the various papers
and keep the resulting DT as generic as possible.
Second, unique combinations of the previously
defined generic methods, which are actually the leaf
nodes of the DT, were identified in the papers
reviewed. Third, rules for classifying the method
combinations using criteria that depend on
functional and non-functional application
requirements were defined; such rules and criteria
are used in the decision nodes of the DT. Fourth, for
all method combinations identified, the fulfilment of
the criteria used in the DT’s decision nodes was
assessed. Finally, the appropriate DT learning
algorithm was selected and fed with the training
The rest of the paper is organized as follows:
Section 2 presents the literature review; Section 3
outlines the method filtering approach, while Section
4 illustrates its application in a maintenance scenario
in the oil and gas industry. Section 6 discusses the
method filtering approach and the results and
concludes the paper.
2.1 Literature Search and Pre-filtering
of Results
Several research works have examined and
developed maintenance decision support methods,
based on historical and real-time data as well as
expert knowledge, in order to address different
maintenance challenges. Most of the existing
research works address maintenance issues for
components subjected to condition monitoring in the
context of Condition Based Maintenance (CBM).
Maintenance decision support is related to
reliability, safety and environmental issues as well
as associated with equipment downtime costs in
cases of breakdowns or malfunctions of machines
(Peng, Dong, and Zuo, 2010). First, prognostic
methods are applied and then decision methods are
developed in order to provide prognostic-based
recommendations. Table 1 summarises the
prognostic-based decision support methods
reviewed, as well as their inputs and outputs. The
methods have been separated in two groups: one
group supports the Prognostic (P) and the second
one the Decision (D) step.
The papers examined were identified by
searching Google scholar with the keywordsCBM,
‘recommendations’, ‘decision support’, ‘decision
making’,’ manufacturing, ‘maintenance’ in various
combinations among them. We focused on papers
dealing with the decision step of CBM. However,
we realized that most of them proposed a
combination of methods so that they develop a
prognostic model based on real-time data and then,
based on this, they provide recommendations for
maintenance. The focus was on most recent papers,
after 2008, with exceptions in cases where an older
paper satisfied the keywords and proposed a novel
and useful method which has not been extended
until now.
2.2 Categorising Methods
The methods found in the literature for prognostic-
based decision support can be divided into the
following generic categories:
Bayesian Network (BN), which also
include Dynamic Bayesian Network
Neural Network (NN)
Statistical Analysis (SA), which include
statistical techniques such as Statistical
Quality Control (SQC), Support Vector
Machine (SVM) and moving average.
Degradation Modelling (DM), which
includes all the mathematical techniques
dealing with representing the degradation
Reinforcement Learning (RL), such as
(SARSA) algorithm.
Markov Chain (MC)
Mathematical Programming (MP)-
Optimisation, which includes operational
research methods such as linear, non-linear
and stochastic dynamic programming
Markov Decision Process (MDP), which
also includes Semi-Markov Decision
Process (SMDP) and Partially Observable
Markov Decision Process (POMDP).
Rules (R), which include rule-based
systems such as IF-THEN rules and Event-
Condition-Action (ECA) rules.
2.3 Identifying Combinations of
As shown in Τable 2, there exist in the reviewed
literature ten unique combinations of methods used
for providing prognostic-based recommendations.
3.1 Identifying Method Filtering
Following our analysis of existing methods and
combinations of methods, we propose criteria for
selecting the appropriate ones based on the
functional and/or non-functional requirements of
specific applications. Selection should be based on
desired output that the business analyst expects to
get after the implementation of the method
combination. Depending on the available input,
different combinations of methods can be applicable.
Another criterion is whether Domain Knowledge
(DK) can be expressed in terms of utility functions.
Finally, the existence of degradation knowledge
affects the selection of the appropriate prognostic
methods and thus the selection of the suitable
combination of methods.
3.2 Evaluating Methods
Based on the information given in Table 1, the
identified combinations of methods are evaluated
according to the four specified criteria. Evaluation of
method combinations on the criteria of available
input and desired output was done based on the
information summarized in Table 1. For evaluating
method combinations on the other two criteria, we
examined in more detail the information provided in
the respective papers. Desired output can be either
the optimal time of applying a predefined action
(e.g., optimal time of replacement of some part of
equipment) or the optimal action and the optimal
time of applying it (e.g., lubrication of metal parts
accompanied with the optimal time). Available input
can include historical data about cause (e.g.,
vibration, temperature, etc.) and effect (e.g., failure,
malfunction, etc.) or prognostic information.
Knowledge of the degradation process is a
prerequisite for some prognostic methods, while this
is not the case for others. Table 3 shows the
evaluation of the methods’ combinations using the
four specified criteria.
3.3 Decision Tree Learning
The data presented in Table 3, were fed into a DT
learning algorithm, i.e. a DT classifier, in order to
produce the DT for supporting analysts perform
prognostic-based maintenance in various application
domains and for a wide range of functional and non-
functional application requirements. DT classifiers
have the ability to handle data which are measured
in different scales, they do not require any
assumptions about the frequency of data in each
class, while they are able to handle non-linear
relationships between features and classes.
Furthermore, the analyst can comprehend and
interpret a decision tree as it is not a ‘black box’
(Pal, M., and Mather, 2003). The ID3 (Iterative
Dichotomiser 3) algorithm, that was used in our
case, classifies all training data provided that there
are enough attributes to do so.
There are several extensions of the ID3
algorithm, such as J48, C4.5 and C5.0, which,
among others, are able to handle continuous
attributes, training data with missing attribute values
and attributes with differing costs. However, these
capabilities are not useful in our case because there
are not any related issues to address. Moreover,
these additional capabilities provide improvements
in terms of speed and memory usage, which are
nevertheless not needed in our case because it
consists of a small number of combinations of
methods and the DT learning is done once. Finally,
the aforementioned extensions of ID3 create smaller
DTs because the probability of over-classifying the
data is much smaller compared to ID3. However, in
our case, we want to separate our method dataset as
much as possible. The four criteria used for the
separation of the combinations of methods and their
alternative values were defined in an abstract level.
This means that if, for example, two combinations of
methods are classified in the same class, they are not
necessarily the same and cannot be used
interchangeably because e.g., their input may require
additional, more specific data or knowledge than
those specified in Table 3.
The ID3 decision trees algorithm is based on
information theory and tries to minimize the number
of comparisons among the data of the training set.
The core idea behind the algorithm Is asking
Table 1: Reviewed Research Works on Prognostic-based Decision Support.
Methods Input Output
(Kaiser, and
P BN; DM Real-time and historical data;
Threshold; Replacements
Estimation of RLD
D R RLD; Costs of maintenance; Process
Compute/update maintenance
P DM; MC (Continuous
Degradation condition monitoring;
Degradation process
RUL; Failure rate
D MC (Continuous time);
RUL; Failure rate; Maintenance and
production knowledge and costs
Optimal maintenance strategy
(Besnard, et
al., 2011)
P - - -
D MP (Stochastic); R Wind forecasting ; Failure rate; List of
actions; Production and maintenance
knowledge and costs
Minimised cost of production
losses and transportation
(Castro, et
al., 2012)
P DM Real-time and historical data; Threshold Mean Residual Life; Times of
D MP (cost minimisation) Mean Residual Life; Times and costs of
Minimised maintenance cost;
Optimum policy
(Wu, et al.,
P NN; MP (Non-linear
programming); SA
(Moving average)
Real-time and historical data; Threshold Residual Life Percentile
D MP (Non-linear
Predicted Residual Life Percentile;
Times of operation; Costs related to
Minimised cost; Optimal
replacement time
(Ivy, and
P SA (SQC) Real-time and historical data; States;
Transition Matrix; Estimation
of observation distribution
D MDP (POMDP) Transition Matrix; Estimation of the
observation distribution parameters;
Maintenance costs
Minimised expected cost ;
Optimal maintenance and
monitoring actions
P RL (SARSA algorithm) Real-time and historical data;
Degradation and maintenance
Solution of SARSA
algorithm; Probabilities of
D MDP Solution of SARSA algorithm;
Probabilities of events
Predictive and corrective
maintenance tasks
P DM Real-time and historical data; Failure
D MP (replacement
RLD; Costs of maintenance; Lead times
of spare parts
Optimal replacement and
inventory ordering times
et al., 2011)
P DM; SA Real-time monitoring; Degradation
process; Threshold
Failure probability; Estimated
degradation path; Time-to-
D MP (maintenance
Failure probability; Time-to-failure;
Maintenance costs
Optimal maintenance cost and
Barros, and,
P DM Condition monitoring; Degradation
Reliability; Probability
density function
DMP (dynamic
replacement model); R
Reliability; Probability density function;
Cost function
Replacement Time
Estimation; Optimised cost
Suhner, and
Iung, 2007)
P BN (DBN) Real-time and historical data; Process Degradation process;
D MC (Discrete time); R Degradation process; Prognosis; List of
actions; Costs
Optimal maintenance policy
Etzion, and
P BN Real-time monitoring; Historical data of
Probability distribution of an
event; Time-to-failure
Probability distribution of an event;
Time-to-failure; States; Actions; Cost
Optimal action; Optimal time
of action
Table 2: Methods and Techniques for Decision Making.
(Kaiser, and Gebraeel, 2009) v v v
(Besnard, and Bertling, 2010) v v v v
(Besnard, et al., 2011) v v
(Castro, et al., 2012) v v
(Wu, et al., 2007) v v v
(Ivy, and Nembhard, 2005) v v
(Aissani, Beldjilali, and Trentesaux, 2009) v v
(Elwany, and Gebraeel, 2008) v v
(Bouvard, et al., 2011) v v
(Huynh, Barros, and, Berenguer, 2012) v v v
(Muller, Suhner, and Iung, 2007) v v v
(Engel, Etzion, and Feldman, 2012) v v
Table 3. Methods’ combinations evaluation.
of methods
Desired Output
Available input
DK expressed in
utility function
Knowledge of the
degradation process
Time of
and time
Prognosis Yes No Yes No
NN-SA-MP v v v v
BN-DM-R v v v v
BN-MC-R v v v v
RL-MDP v v v v
DM-MC-MP-R v v v v
SA-MDP v v v v
DM-MP-R v v v v
DM-MP v v v v
MP-R v v v v
BN-MDP v v v v
questions the answers of which provide the most
information. The splitting criteria are prioritized
according to the information gain; splitting criteria
with more information gain are used first. The
decision tree is constructed by employing a top-
down, greedy search through the given sets to test
each attribute at every tree node. Information is
measured by the entropy which represents the
amount of uncertainty of a data set D (Chen, Zhang,
and Tong, 2014). Based on the entropy, the
information gain can be measured. Information gain
is the difference in entropy from before to after the
data set D is split on an attribute A or equally, how
much uncertainty in the data set was reduced after
splitting it on an attribute A (Gaddam, Phoha, and
Balagani, 2007).
The DT was built by feeding the combinations of
methods identified in the literature as training data to
the decision tree learning algorithm and shows the
sequence of steps needed to be followed by an
analyst in order to decide which combination(s) of
methods are the most appropriate ones for a specific
problem. The pseudo-code of the application of the
ID3 algorithm for the classification of the
combinations of methods according to the four
criteria is shown below (adapted from (Jin, De-lin,
and Fen-xiang 2009)):
ID3 (Set of combinations of methods D,
Set of criteria-attributes S, Criteria-
Attributes_values V)
Return Decision Tree.
Load set of combinations of methods D
first, create decision tree root node
'rootNode', add learning set D into
root node as its subset.
For rootNode, we compute
H(rootNode.subset) first
If H(rootNode.subset)==0, then
rootNode.subset consists of
records all with the same
value for the categorical
return a leaf node with
decision criterion-
attribute: criterion-
attribute value;
If H(rootNode.subset)!=0, then
compute IG for each
criterion-attribute left
(have not been used in
splitting), find attribute S
with Maximum(IG(D, S)).
Create child nodes of this
rootNode and add to rootNode
in the decision tree.
For each child of the
apply ID3(D, S, V)
recursively until reach
node that has H=0 or
reach leaf node.
End ID3
Our problem was formulated in the ID3 notation
using the RapidMiner machine learning software,
while the DT obtained after running ID3 is shown in
Figure 1.
In this section, we outline how our method filtering
approach can help in selecting the most appropriate
method combination for supporting decision making
in a maintenance scenario in the oil and gas industry.
In the scenario under consideration, sensors collect
data with a very high frequency, and these data
accompanied with historical data and domain
knowledge are used for detecting the current health
state of the equipment examined, estimating RUL
and calculating the probability distribution of an
undesired event, e.g., breakdown of the gearbox of
an oil drilling company’s equipment. Historical data
show the patterns of the monitored parameters,
which are used as indicators of degradation till
failure. Domain knowledge can include a list of
maintenance actions, failure threshold as well as
utility functions considering criteria such as cost,
time and safety. Then, the optimal action and the
optimal time of applying it are recommended.
The DT flow in the aforementioned scenario is
shown in Figure 2. First, as far as the available input
is concerned, in the aforementioned scenario there
are historical data about causes (sensed parameter)
and effects (failure) but not prognostic information
(e.g. RUL, probability distributions of the
occurrence of failure, etc.), while data are
continuously updated with the ones coming from
The prognostic information is not known in
advance, but it will derive from the processing and
analysis of data by using the appropriate method.
The output of this method will feed into another
method for providing recommendations. Then, the
desired output is the optimal action and the optimal
time for this action, because the objective is to
identify the best maintenance action out of a list of
actions as well as the best time to implement it in
Figure 1: Method filtering process for choosing the appropriate combination of methods.
Figure 2: Method filtering process in a maintenance scenario.
terms of cost.
DK can be expressed in utility functions as there
is extensive industrial experience on the domain
which can be expressed in a systematic way. Issues
about cost function, safety, current maintenance
policy, etc. can be embedded to a utility function
which can be used in optimization techniques in
order to provide reliable recommendations. Finally,
there is knowledge of the degradation process.
Hence, there are four options: BN-MC-R, DM-MC-
For example, for the BN-MDP combination, BN
are used for data-driven estimation of probability
distribution of an undesired event (e.g. gearbox
breakdown) and MDP for generating
recommendations about optimal maintenance
actions and optimal time for applying these actions.
Moreover, this particular combination can
effectively support decision-making in
manufacturing enterprise and especially for CBM
(Engel, Etzion, and Feldman, 2012). This case
shows that MDP and MC are the most suitable
methods for extracting this output provided that
some domain knowledge exists and probability
distribution of an undesired event has been extracted
from machine learning or statistical methods.
CBM relies on prognostic models and uses them to
support decisions about the appropriate maintenance
actions based on the current health state of a system
through condition monitoring (e.g. using sensors).
To do this, combinations of both machine learning
and decision methods are required in a way that they
are able to handle real-time data and provide
recommendations for maintenance decisions based
on predictions about future health state of the
We examined literature that deals with methods
supporting decision making in the context of CBM.
The method filtering approach that we propose
supports the business analyst to select the most
appropriate combination(s) of methods based on the
requirements of the specific maintenance scenario.
Our method may recommend more than one
alternative method or method combinations, which
are applicable in specific maintenance scenarios and
under specific conditions. In such cases however,
methods or combinations of methods that are
classified in the same class are not the same and,
while our filtering method cannot discern between
them, a human expert should be able to do so by
taking into account additional data or knowledge in
order to select the most appropriate. Although we
identified several combinations of methods that are
used for prognostic-based maintenance
recommendations, most are not able to adequately
support proactive decision making by
recommending the optimal action and the optimal
time of applying it based on predictions.
Furthermore, there are limitations regarding the
continuous improvement of the recommendations
using these combinations of methods.
Our future research will focus on the
examination and incorporation of additional machine
learning and decision methods specifically targeting
proactive decision support. Moreover, we will
extend our method filtering approach with a
feedback loop, which will support collection of data
about the effectiveness of the recommended
decisions and will utilize the collected data as a basis
for improving the recommendation generation
process. Finally, we will test and evaluate our
approach in real maintenance scenarios in the oil &
gas and automotive industries.
This work is partly funded by the European
Commission project FP7 STREP ProaSenseThe
Proactive Sensing Enterprise” (612329).
Aissani, N., Beldjilali, B., Trentesaux, D., 2009. Dynamic
scheduling of maintenance tasks in the petroleum
industry: A reinforcement approach. Engineering
Applications of Artificial Intelligence, 22(7), 1089-
Besnard, F., & Bertling, L., 2010. An approach for
condition-based maintenance optimization applied to
wind turbine blades. Sustainable Energy, IEEE
Transactions on, 1(2), 77-83.
Besnard, F., Patriksson, M., Stromberg, A.,
Wojciechowski, A., Fischer, K., Bertling, L., 2011. A
stochastic model for opportunistic maintenance
planning of offshore wind farms.In PowerTech, IEEE
Trondheim, pp. 1-8.IEEE.
Bouvard, K., Artus, S., Berenguer, C., Cocquempot, V.,
2011. Condition-based dynamic maintenance
operations planning & grouping. Application to
commercial heavy vehicles. Reliability Engineering &
System Safety, 96(6), 601-610.
Castro, I. T., Huynh, K. T., Barros, A., Berenguer, C.,
2012. A predictive maintenance strategy based on
mean residual life for systems subject to competing
failures due to degradation and shocks. In Proc. of the
11th International Probabilistic Safety Assessment and
Management Conference & the Annual European
Safety and Reliability Conference-PSAM 11/ESREL
2012, pp. 375-384.
Chen, X. J., Zhang, Z. G., Tong, Y., 2014. An Improved
ID3 Decision Tree Algorithm. In Advanced Materials
Research, Vol. 962, pp. 2842-2847.
Elwany, A. H., Gebraeel, N. Z., 2008. Sensor-driven
prognostic models for equipment replacement and
spare parts inventory. IIE Transactions, 40(7), 629-
Engel, Y., Etzion, O., Feldman, Z., 2012. A basic model
for proactive event-driven computing. In Proceedings
of the 6th ACM International Conference on
Distributed Event-Based Systems, pp. 107-118.ACM.
Gaddam, S. R., Phoha, V. V., Balagani, K. S., 2007. K-
Means+ ID3: A novel method for supervised anomaly
detection by cascading K-Means clustering and ID3
decision tree learning methods. Knowledge and Data
Engineering, IEEE Transactions on, 19(3), 345-354.
Garg, A., Deshmukh, S. G., 2006. Maintenance
management: literature review and directions. Journal
of Quality in Maintenance Engineering, 12(3), 205-
Huynh, K. T., Barros, A., Berenguer, C., 2012.
Maintenance decision-making for systems operating
under indirect condition monitoring: value of online
information and impact of measurement uncertainty.
Reliability, IEEE Transactions on, 61(2), 410-425.
Ivy, J. S., Nembhard, H. B., 2005. A modeling approach to
maintenance decisions using statistical quality control
and optimization. Quality and Reliability Engineering
International, 21(4), 355-366.
Jardine, A. K., Lin, D., Banjevic, D., 2006. A review on
machinery diagnostics and prognostics implementing
condition-based maintenance. Mechanical systems and
signal processing, 20(7), 1483-1510.
Jin, C., De-lin, L., Fen-xiang, M., 2009. An improved ID3
decision tree algorithm. In Computer Science &
Education. ICCSE'09. 4th International Conference
on, pp. 127-130. IEEE.
Kaiser, K. A., Gebraeel, N. Z., 2009. Predictive
maintenance management using sensor-based
degradation models. Systems, Man and Cybernetics,
Part A: Systems and Humans, IEEE Transactions on,
39(4), 840-849.
Muller, A., Suhner, M. C., Iung, B., 2007. Maintenance
alternative integration to prognosis process
engineering. Journal of Quality in Maintenance
Engineering, 13(2), 198-211.
Pal, M., & Mather, P. M., 2003. An assessment of the
effectiveness of decision tree methods for land cover
classification. Remote sensing of environment, 86(4),
Peng, Y., Dong, M., Zuo, M. J., 2010. Current status of
machine prognostics in condition-based maintenance:
a review. The International Journal of Advanced
Manufacturing Technology, 50(1-4), 297-313.
Wu, S. J., Gebraeel, N., Lawley, M. A., Yih, Y., 2007. A
neural network integrated decision support system for
condition-based optimal predictive maintenance
policy. Systems, Man and Cybernetics, Part A:
Systems and Humans, IEEE Transactions on, 37(2),
Zikopoulos, P., Eaton, C., 2011. Understanding big data:
Analytics for enterprise class hadoop and streaming
data. McGraw-Hill Osborne Media.