See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/271561294

Intelligent automation systems for predictive maintenance: A case study

Article  in  Robotics and Computer-Integrated Manufacturing · October 2006

DOI: 10.1016/j.rcim.2005.12.010

CITATIONS READS

50 1,112

2 authors:

Eduardo Gilabert Aitor Arnaiz

TEKNIKER TEKNIKER
27 PUBLICATIONS   185 CITATIONS    82 PUBLICATIONS   615 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

EEPSA (Energy Efficiency Prediction Semantic Assistant) View project

T-REX: Lifecycle Extension Through Product Redesign And Repair, Renovation, Reuse, Recycle Strategies For Usage&Reusage-Oriented Business Models View project

All content following this page was uploaded by Aitor Arnaiz on 30 January 2018.

The user has requested enhancement of the downloaded file.

 ARTICLE IN PRESS

Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549

www.elsevier.com/locate/rcim

Intelligent automation systems for predictive maintenance: A case study

Eduardo Gilabert, Aitor Arnaiz
Manufacturing Processes Department, Fundación Tekniker, Eibar, Gipuzkoa 20600, Spain
Received 7 October 2005; accepted 15 December 2005

Abstract

A case study is presented, where a predictive maintenance solution for non-critical machinery (such as elevators and machine tools)
was sought. Both cases are different. There is no experience in elevator monitoring and diagnosis, and modeling has been performed
using Neural Networks. On the other hand, machine tools were monitored through vibration systems where some experience exists. In
this case, Bayesian Networks are the paradigm of choice as it was also recommended to include some ‘adaptation’ mechanism for the
knowledge modeled in the network. The final system also includes a sensor processing unit and a remote maintenance module system that
provides an automated remote condition monitoring system, for both applications. Results indicate the feasibility of partial solutions in
monitoring and diagnosis, though future enhancements are needed to compose a complete solution. This paper explains the
characteristics of the Bayesian Network solution finally developed for high-speed machine tools, evaluate their strengths and weaknesses,
and indicate the future enhancements.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Bayesian networks; Knowledge modeling; Monitoring; Diagnosis; Uncertainty; Adaptation

1. Introduction from data as there is no experience available. Finally, there

are cases where continuous adaptation of diagnostic and
Predictive maintenance strategies are very efficient in monitoring solutions is required, as knowledge in main-
mechanical-failure modes, when failure probability in- tenance increases.
creases with time, and one or more condition-monitoring In order to cope with these problems, several algorithm
techniques (vibrations, oil, thermography, sound,y) can paradigms coming from machine-learning field can be
predict the failure before breakage. However, use of these applied to a variety of situations, and with diverse
techniques have been always linked to a very limited kind requirements and problem characteristics (supervised, un-
of assets, where unavailability can lead to safety loss supervised, reinforcement learning, expert knowledge, data
(Aerospace), or big economical losses (Energy). mining). Among these, there are some that have developed
One of the greatest problems in predictive maintenance into robust tools that can be used for the modeling of
is the need to handle great amounts of data and the cost in several kind of classification problems, such as monitoring
personnel training. To overcome this problem, several and diagnosis. Examples of successful paradigms are CBR
decision support systems are in development and research. systems, neural networks, induction trees & rules, and
However, the ill-structure of predictive data makes it Bayesian networks.
difficult to automate the tasks, as there is normally A case study is presented, corresponding to EC partially
uncertainty in the diagnosis process. Moreover, some supported MINICON project (minimum cost, minimum
diagnosis and monitoring models must be built straight size, maximum benefit condition monitoring system),
where a predictive maintenance solution for non-critical
Corresponding author. Tel.: +34 943 206 744; fax: +34 943 202 757. machinery (such as elevators and machine tools) was
E-mail address: [email protected] (E. Gilabert). sought.

0736-5845/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.rcim.2005.12.010
 ARTICLE IN PRESS
544 E. Gilabert, A. Arnaiz / Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549

2. State of the art in monitoring and diagnosis Another challenge relates to the need to keep automated
systems adapted. In [2], it is indicated that once we have a
Maintenance actions are launched depending on the knowledge-based system, no matter how has it been
status of a parameter that is usually related to known created, we can distinguish three basic motivations for
equipment malfunctions. Typical indicators here are learning and knowledge change:
vibrations, temperatures, performance losses, etc. This
maintenance is more commonly referred to as condition-
correction of system errors (knowledge acquired
based maintenance (CBM) or predictive maintenance through experience and feedback)
(PdM) and the use of the above indicators is normally
enhancement of knowledge base (new products, new
referred to as a condition-monitoring activity. It is analysis means)
important to point out that other techniques, such as
improvement of knowledge base in terms of efficiency
statistical process control, can also be of help. (dynamic strategies for maintenance cost reduction)
Predictive maintenance seeks for a much more cost-
effective analysis than preventive maintenance, improving
equipment reliability (avoiding unexpected machinery fail- 3. MINICON project
ures).
However, predictive maintenance methods require great This paper introduces the work carried out under a
quantities of information in order to provide adequate larger project (MINICON) that intends to develop a cost-
response to different fault detection sub-problems: identi- effective integrated sensor processing units (SPUs) for
fication of abnormalities, diagnosis, assessment of fault inclusion of on-line condition monitoring systems on all
severity, prognosis, etc. All these tasks are ‘knowledge range of industrial and civil machinery. This includes two
intensive’, as there is an evident need of expertise to be able objectives: the consecution of a cost-effective hardware
to handle them. Also, the answer of previous problems monitoring unit (or sensor processing unit-SPU) and the
requires the inclusion of information from many different availability of automated systems to perform decision
sources (vibrations, temperature, oil, wear debris,y), support on fault diagnosis and troubleshooting.
sources that not very long ago were isolated.
Uncertainty appears in diagnosis problems at different 3.1. MINICON overview and objectives
levels, that can be classified as information and/or model
flaws [1]. In a problem space such as failure diagnosis by MINICON [3] aims to apply cost-effective operation
analysis of vibrations, we can see that, when talking about monitoring and maintenance to small elements of indus-
failure modes, the borderline between dissimilar problems trial facilities (i.e. machine tools) and civil infrastructure
is often fuzzy. For instance, it is difficult to determine if (i.e. elevators and escalators). By introducing flexibility,
there is just misalignment or there is also unbalance. In self-configuration capability, improved performance and
many cases, though there is a crystal-clear difference, knowledge-based modules to the monitoring sensors, it
several issues complicate this: helps to optimize the overall production and life cycle of
the relevant machinery.
Finally, by developing a new remote maintenance

machinery symptoms are likely to be a mix of different
management Scheme for these installations, which makes
problems (unbalance and misalignment appear very
use of actual service data and is based on continuously
often, though one of them is at a minimal level),
monitoring structural integrity and prognosis of relevant

information (e.g. spectra) could be unclear. For
damage, MINICON intends to drastically reduce the
instance, frequency resolution is limited and, depending
relevant maintenance costs (Fig. 1).
on the number of the components in analysis, and the
spectral range, some measurements can be bonded to
3.2. Monitoring and diagnosis algorithms in MINICON
more than one problem,

in many cases, the range of the devices installed can limit
The design of diagnostics models able to adapt to new
information, such as the ability to measure the HFD
conditions is of primary importance within the MINICON
and Spike energy measurements,
final framework system. The consecution of a cost-effective

models also have inexactitudes, due to high human
SPU depends in part on the incorporation of robust fault
dependency when assessing severity of spectral peaks.
detection methods. For these methods to work, we used a
Concerning machine tools, for instance, there is hardly
knowledge-based modeling strategy. The main advantage
any information about peak limits surpassed.
over previous knowledge based approaches is the use of
algorithms that facilitate the re-use of knowledge and the
As a result, a given spectral pattern can be related to inclusion of examples in an adaptive, on-line, learning
several problems, with uncertainty linked to interpretation approach.
(in almost all cases) and, therefore, uncertainty should be Given this, we may see that the different nature of
treated in order to solve diagnostic problems adequately. machine tools and elevators not only relates to the different
 ARTICLE IN PRESS
E. Gilabert, A. Arnaiz / Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549 545

Application software
and database

Data store

Maintenance
management system
providing decision
Elevator or support
Machine tool to
be monitored
CMMS
On-board
intelligent
alarm system

SPU

Service
Sensors engineer with
portable device

Fig. 1. Conceptual design of MINICON project.

strategies required for fault detection. What is more, the software system may encounter at the end of the project.
differences also include the way modules should be Some new applications could be integrated through
constructed [4]: adaptation of new knowledge on final diagnostic structure
An adaptive system for condition monitoring in eleva- (through supervised learning, such Bayesian networks),
tors can be tackled in an efficient manner through and some other would be less clear to model due to lack of
math techniques, since here two conditions appear (less expertise knowledge, and could be best tackled through
knowledge on the machine faults to be modeled—less mathematical techniques based on automated feature
understood failure mechanisms and related observable selection and extraction and multi-parameter fusion
symptoms—and availability on demand of a big sample systems, using non-supervised clustering systems such as
base with different elevator status). Moreover, many neural networks.
incipient faults can derive into critical in a short time.
In machine tools, however, the situation is opposite: 4. Bayesian networks
there are a few data samples since monitoring schedule is
normally based on quarterly/monthly/bimonthly data If we stress on the two problems described (uncertainty
acquisition, with few failure occurrences and slow mechan- management and adaptation) we could find a knowledge
ical degradation. Also, there is already a high knowledge modeling methodology that suits our aims: Bayesian
on failure components (gears, bearings, y) and on networks.
observable symptoms. The most important point here can
be to model the diagnosis system in a flexible way to allow 4.1. Overview of Bayesian networks
adaptation to new components:
elevators: math/statistical fusion techniques are foreseen A Bayesian network [5] is a compact model representa-
tion for reasoning under uncertainty. It reflects the states

first: multiple feature sets generation of some part of a world that is being modeled and it

second: statistical clustering and ANN to extract models describes how those states are related though conditional

develop model without knowledge probabilities.
A problem domain—diagnosis of mechanical failures,
machine tools: knowledge-based techniques for instance—consists of a number of entities or events.
These entities or events are, in a Bayesian network,

no feature selection needed as features are already represented as random variables. One random variable
known can, for instance, represent the event that a piece of

first: generic models for known fault diagnostics mechanical hardware in a production facility has failed.

second: validate knowledge-based models The random variables representing different events are
connected by directed edges to describe relations between
Finally, the most important conclusion is that this events. An edge between two random variables X and Y
division also reflects very well what a final SPU-adaptive represents a possible dependence relation between the
 ARTICLE IN PRESS
546 E. Gilabert, A. Arnaiz / Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549

events or entities represented by X and Y. For instance, an 4.2. A Bayesian network for diagnosing machine tools
edge could describe a dependence relation between disease
and a symptom—diseases causes symptoms. The following diagram illustrates an example of a
Thus, edges can be used to represent cause–effect Bayesian network (Fig. 2).
relations. The dependence relations between entities of In the network prototype shown above, the qualitative
the problem domain are organized as a graphical structure. relationship indicated by the direction of the link arrows
This graphical structure describes the possible dependence corresponds to dependence and independence between
relations between the entities of the problem domain. The events. That is, the top nodes in the diagram tend to
uncertainty of the problem domain is represented through influence those below rather than the other way around.
conditional probabilities. Conditional probability distribu- On the other hand, the quantitative relationships between
tions specify our belief about the strengths of the nodes are defined by conditional probability tables
cause–effect relations. Thus, a Bayesian network consists (Table 1), where it can be extracted that a priori
of a qualitative part, which describes the dependence probabilities of having unbalance are of 40%, and
relations of the problem domain, and a quantitative part, conditional probabilities of unbalance having evidence of
which describes our belief about the strengths of the high amplitude peaks at fundamental frequency (1X) are
relations. of 99%,
This representation is known as a directed acyclic graph
(DAG) consisting of nodes, which correspond to random that is; PðUÞ ¼ 0:4 and PðUj1X Þ ¼ 0:99.
variables and arcs that represent the probabilistic depen-
dencies between the variables [6,7]. Here, we can see a directed acyclic graph since each arc is
Many practical tasks can be reduced to the problem of directed and there are no loops. It contains 5 nodes
classification. Fault diagnostics is one of these examples. (discrete random variables) that corresponds to different
A Bayesian network helps tackle the problem of classifica- status that can occur in our system: unbalance (U),
tion in a way that helps to overcome problems that other misalignment (M), amplitude increase in vibrations as speed
methods partially address: increases (IS), high amplitude peaks at fundamental
frequency (1X) and high amplitude peaks at second harmonic
(2X). This graphical representation serves to represent the

able to mix a priori knowledge together with data/ dependence and independence assumptions, that leads to
experimental knowledge, develop a joint probability distribution. For instance we

explanatory abilities, can say that 2X is conditionally dependent of M, but is

uncertainty management—causality management, conditionally independent of U given M.

learning both parametric and structural issues.
that is Pð2X jU; MÞ ¼ Pð2X jMÞ.

We can also say that 1X and 2X are conditionally

independent given shaft failure
Unbalance (U) Misalignment(M)
that is Pð1X jM; 2X Þ ¼ Pð1X jMÞ.

By probability theory, we know that we can infer many

probabilities. Pð2X jMÞ ¼ PðMÞ  PðMj2X Þ=Pð2X Þ by the
use of Bayes theorem and conditional probabilities. This is
Incrspeed(IS) 1X 2X used to infer the probability of misalignment (M) having
evidence on appearance of a peak of amplitude in second
Fig. 2. Excerpt from original Bayesian network. harmonic (2X), but it also can work in the opposite

Table 1
Excerpt from original chart of failure probabilities

Failure cause A priori Freq. probability Meas. probability Amplitude incr.

1X 2X 42X Radial Axial Speed Load

Unbalance 40 99 90 10 90
Bent shaft 10 90 10 20 80 90
Angular mis. 30 70 20 10 20 80 90
Parallel mis. 30 20 70 10 80 20 90
Mechanical loss 5 80 20 90 10
Other causes 20 30 50 50 50 25 25
 ARTICLE IN PRESS
E. Gilabert, A. Arnaiz / Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549 547

direction, making it useful—for instance- to—reinforce been used in the Bayesian network for diagnosing machines
other causes (such as unbalance) if we already know that tools.
there is no misalignment in the system. By the appropriate The simplest example is that referring to fractional
inference mechanisms, the management of the partial updating tables, were an statistical task is meant to modify
conditional distributions allows the calculation of prob- the estimates of the parameters gradually with the cases
abilities for any node in the network given any evidence. used. We can consider the conditional probability table
The modeling of Bayesian networks is completely suited (CPT) of unbalance, that is, the prior probability of having
to the new trends existing in knowledge engineering: the unbalance in a given machinery type every time an
change from knowledge acquisition (represented by expert abnormality is detecting:
systems) to knowledge modeling. Both Bayesian networks
and knowledge modeling methodologies share the need of Unbalance
a proper knowledge representation through problem
modeling within available templates and have arisen in False 0.60
last 10 years. In both cases, knowledge-based systems can True 0.40
be seen as antecedents. Concerning this, Bayesian networks
can be seen [8] as special knowledge representation systems
fitted for the representation of causal semantics. It is clear that this relationship may not reflect particular
cases in different environments. For instance, a given
company may present a particular highly rate of abnorm-
4.3. Learning in Bayesian networks alities related to this problem. We can add a new feature on
this CPT, called ‘experience’, represented by a number that
Learning a graphical model has become a very active indicates the value that we assign to experience in the ‘a
research topic and many algorithms have been developed priori’ design. Now, we can also include feedback from
for it. Introductory and advanced information on prob- application of the system. For example, we can assume that
abilistic network learning can be found in [6]. Here, we can our belief in the correctness of the current conditional
distinguish between three approaches. distribution for unbalance is low, thus we can set the initial
experience count to a small number, say 10.

Learning the structure: this kind of learning tries to
make the whole structure of the Bayesian networks
through a fusion of data and expert knowledge. Unbalance
Methods for structural learning include naı̈ve Bayes
approaches, search and scoring based methods (K2) and False 0.60
dependency analysis (PC, NPC). These methods can True 0.40
only be considered as promising. Experience 10

Learning the probabilities in batch: the learning of the
information regarding the conditional distributions.
Parameter estimation uses algorithms such as estima- Now, if we get feedback of 5 new cases (suppose that in
tion–maximization (EM) to look for the best parameter all of them evidence on unbalance is found) and go again to
distribution given a a priori graph configuration. This the CPT of unbalance, we will see an experience value of
and previous approaches need a much greater sample/ 15. These additional 5 counts all pertain to state ‘‘true’’.
cases database. Therefore, the adapted probability distribution of unba-

Learning the probabilities sequentially: this approach is lance becomes:
used when we have certain structure, but we want NðtrueÞ
P ðunbalanceÞ ¼ ,
probabilities to adapt to a particular context. It is also experience
called ‘Adaptation’. where N(true) indicates the number of true events recorded
so far, which accounts for 5 in the last 5 observations, plus
Adaptation is the process of refining the (conditional) 4 in the first 10 observations (0.40  10). That is:
probabilities specified for a Bayesian network by taking
ð0:40  10Þ þ 5
into consideration the real experiment outcomes. This is P ðunbalanceÞ ¼ ¼ 0:6.
15
probably the most interesting type of learning mechanism
that can be used in machinery diagnosis, as the most This gives the following CPT as results:
important input (in learning terms) should be expected
from local usage of the automated tools, as long as they Unbalance
start to be applied in maintenance and diagnosis systems.
For example, every time a machine is diagnosed, the False 0.40
information about their symptoms and problems can be True 0.60
used to adapt to the network’s probabilities. This focus has Experience 15
 ARTICLE IN PRESS
548 E. Gilabert, A. Arnaiz / Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549

Finland

Nokia Communication
GSM
Gateway module
Elevator
SPU

DB MMS

Spain
Data INTERNET
Odyssey transfer
DB
tool

Intranet Any location Estonia

Milling Machine SPU GUI, Diagnostic MMS

Fig. 3. Overview of system installations and their links.

To summarize, an adaptation step consists of entering proof-of-concept systems designed for two pilot applica-
evidence, propagating, and updating (adapting) the condi- tions—an elevator and a milling machine (Fig. 3).
tional probability tables and the experience tables. The final system demonstrated consists of two different
The previous adaptation procedure gives equal weight to SPUs (one at Finland consisting of the sensors and GSM
recent experiences and older ones. Sometimes, however, comm. for elevator condition monitoring, and the other
old observations do not count as much as more recent one at Spain, to monitor the condition of a high-speed
ones. Thus we have to unlearn or forget some of them. This machine tool). A remote maintenance management
is the same as saying that new observations (evidence) are (RMM) system can interact with both SPUs, whether
more important than older, and hence should have more directly through GSM communication, or indirectly
weight. Also, fractional updating has a problem of over- through Internet. Data stored in the main database can
estimation of counts’ importance, so that counts can make be accessed from any location, and the final software
parameters in CPT too resistant to change. configuration reflects the option of having very different
Such situations can be overcame by fading tables, which condition monitoring systems integrated. Schematic draw-
accounts for the rate at which previous observations are ing of the complete experimental system is shown at figure
forgotten. Thus, with an intermediate fading value of 0.5 above, indicating locations of individual components and
and a new true case of unbalance: links among them.
ðfading_factor  NðtrueÞÞ þ 1
P ðunbalanceÞ ¼ 6. Conclusions
ðfading_factor  experienceÞ þ 1
0:5  ð0:40  10Þ þ 1
¼ ¼ 0:5. In the first place, this work exposes the necessities that
6
can arise in the near future, within the application of
A very significant change with just one observation. intelligent systems for the maintenance of industrial assets.
Thus, we have worked with maintenance that are currently
covered by working inferential systems (i.e., expert
5. System integration systems) or are demanded within the industry. These
problems have allowed to identify learning processes that
The research conducted in the frame of the MINICON can be necessary to complement and to support the
project has addressed a number of elementary problems of inferential processes.
condition monitoring in various applications. The results This paper also presents a description of available
of the research have been validated by building of two learning algorithms within Bayesian networks. The
 ARTICLE IN PRESS
E. Gilabert, A. Arnaiz / Robotics and Computer-Integrated Manufacturing 22 (2006) 543–549 549

complete deployment and testing of learning capabilities is mechanisms, understood as utility or decision nodes in
being made with HUGIN tool [5], which includes the Bayesian networks, where normal diagnosis nodes (non-
necessary algorithms to implement the learning system. troubleshooting) are just referred to as belief nodes.
Learning can be considered as the true characteristic that In conclusion, we can say that, there is a need for
makes a system to look like an intelligent device, developing adaptive systems that can improve software
automatically preventing the repetitive occurrence of silly support for maintenance decisions in many application
errors. We have shown that Bayesian networks are fields besides machine tools and elevators. The work shown
algorithms worthy to be considered within the diagnosis in this paper is a sound alternative to provide the necessary
area and failure detection. We include one prototype that adaptation mechanisms.
shows how to adapt a system already built with partial
information coming from incomplete sample sets.
There are, however, some issues that should devote Acknowledgement
additional work, beyond the scope of current activities.
The gathering of appropriate data samples is a painstaking This work is partially sponsored under the EC contract
need, in order to properly ensure that any model works GRD1-2000-25639, which refers to MINICON project.
properly. Even though the data needs are low, since no http://minicon.iccs.ntua.gr/
training and testing sets are needed when experience is
available, validation is still a necessary step. Validation has
been performed using the data collected from several References
machines during the data analysis stage, and some more
samples are coming from the pilot test. [1] Dı́ez FJ. Introduction to approximate reasoning. Introducción al
Razonamiento Aproximado. PhD Course, UNED, Madrid, 2000.
We could also integrate synthetic data, taking into [2] Arnaiz A, Gilabert E. Diagnosis with Bayesian Networks. Euro-
account controlled deviations from real samples, as well as maintenance, 2004.
theoretically known sample configurations. However, to [3] http://minicon.iccs.ntua.gr/.
become a true validation dataset, real samples with most of [4] Arnaiz A, Arzamendi J, Atxaerandio A, Addison D, Helle A.
MINICON Report D6a. Diagnostic modelling and confidence level
the diagnosis problems, and different ‘scenarios’ (machin-
determination—Algorithms for robust determination of confidence
ery type, working conditions, etc.) should be taken. The levels, Internal report, 2002.
collection of reliable data should be understood as of prime [5] http://www.hugin.com/.
necessity in the machine tool sector, in order to validate [6] Jensen FV. Bayesian networks and decision graphs. Berlin: Springer;
any diagnostic alternative, and it can deserve a whole 2001.
multi-national effort, bringing together past and future [7] Castillo E, Gutierrez JM, Hadi AS. Expert systems and probabilistic
network models. Sistemas Expertos y modelos de redes probabilı́sticas,
monitoring and diagnostic results. 1997.
Concerning the BN development, the work performed [8] Benjamins R. PSM for Diagnosis and their role in Knowledge
has not deployed the potential of the troubleshooting acquisition. Int J Expert Syst: Res Appl 1995;8(2):93–120.

View publication stats