Content uploaded by Sulo Lahdelma
Author content
All content in this area was uploaded by Sulo Lahdelma on Nov 19, 2015
Content may be subject to copyright.
Adaptive Features in Condition Monitoring Systems
Prof. Dr. Jens Strackeljan (*) , Prof. Dr. Sulo Lahdelma (**),
(*) Otto-von-Guericke-Universität Magdeburg, Fakultät für Maschinenbau
Institut für Mechanik
Universitätsplatz 2, 30106 Magdeburg, Germany
phone: +49 0391 67 {18438}
e-mail: {jens.strackeljan}@ovgu.de
(**) Mechatronics and Machine Diagnostics Laboratory, Department of Mechanical
Engineering, P.O.Box 4200, FI-90014 University of Oulu, Finland
phone: +358-8-5532083
e-mail: sulo.lahdelma@oulu.fi
Abstract
This paper presents some aspects concerning the problems of adaptive monitoring sys-
tems. Each automatic monitoring system has to be adapted if it is installed in a new en-
vironment. Characteristic of solving the monitoring task, the number of fault classes and
free parameters in the internal classier are potential switchers to adjust the system. We
discuss general problems in the field, such as fault simulation, provide the necessary
definitions of different levels of adaptivity, describe the state of the art and give some
hints about how the implementation of intelligent data pre-processing can improve the
transfer of data from an existing system to a new one. As an application we use the de-
tection of fault in a roller bearing using the derivative x(4) to obtain a higher sensitivity
in the monitoring system.
Keywords: Condition monitoring, roller bearing diagnostics, adaptivity, fractional de-
rivative
1. Introduction
The use of advanced pattern recognition systems in assuming an objective perspective
in statements concerning the state of technical systems has gained increasing impor-
tance. Thus the economy of highly automated and cost-intensive machines can only be
guaranteed upon the high availability of these machines. The use of advanced, high-
performance monitoring and diagnosis systems can make a significant contribution to
this. Certain processes can be carried out safely for man and the environment only by
means of reliably operating machines, particularly in fields where safety and environ-
mental aspects play an important role. In the automatic control of technical systems,
supervisory functions serve to indicate undesired or non-permitted machine or process
states and to take appropriate actions in order to maintain operation and to avoid dam-
age or accident. The following basic functions can be distinguished(23,24) :
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1360
2
1. Monitoring
Measurable variables are checked with regard to tolerances, and alarms are gener-
ated for the operator.
2. Automatic protection
In the case of a dangerous process state, the monitoring system automatically initi-
ates an appropriate counteraction.
3. Monitoring with fault diagnosis
Based on measured variables, features are determined and a fault diagnosis is per-
formed; in advanced systems, decisions are made for counteractions.
The advantage with the classical level-based monitoring (1. and 2.) is simplicity, but it
is only capable of reacting to a relatively large-scale change in a given feature. If the
early detection of small faults and a fault diagnosis are desired, advanced methods based
on Fuzzy Technology, Neural Networks or their combinations could be used(23). A gen-
eral problem which is very often cited as an argument against the application of adap-
tive methods in monitoring is the risk that the system may gradually and unnoticeably
adapt to undetected changes. As a result, the system may fail to detect real faults, if
these faults likewise develop slowly. Problems could occur if the transition from state A
to C proceed over a long time (Fig. 1), because the classifier adaptation could follow the
transition in small steps. Consequently, adaptive monitoring systems have hardly be-
come established in fields which are especially critical, such as safety-relevant applica-
tions(24).
Figure 1. Transition of machinery condition from a healthy state A to a fault
state C (24).
The second limitation arises from problems to generate a learning data set for the classi-
fier design. In most cases only few or sometimes even no real measurement data are
available to represent information about all possible machine or process states. Conse-
quently the simulation of fault classes becomes more important. This paper will empha-
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1361
3
sise the main aspects regarding the integration of simulation results to improve the
adaptive behaviour of monitoring systems.
2. Smart Adaptive Monitoring Systems
The concept of “smart, adaptive monitoring system” must first be defined. Furthermore,
if a system is to be considered “smart” and “adaptive”, the requirements imposed on
such a system must be specified.
2.1 General Properties of Smart Monitoring Systems
In the present section, the following four properties are employed as necessary criteria
for a smart system(1, 24) :
1. Adapting
Ability to modify the system behaviour to fit the environment, new locations and
process changes. The aspect is identified as the most important feature for smart
adaptive monitoring systems.
2. Sensing
Ability to acquire information from the surrounding world and respond to it with
consistent behavior. Chemical and nuclear power plant monitoring are large scale
sensing systems. They accept input from hundreds of sensors to regulate tempera-
ture, pressure and power throughout the plant. In spite of the fact that different sen-
sors exist, more and better sensors are continuosly required. Along with a prolifera-
tion of sensors comes a greatly increased need to combine, or fuse data from multi-
ple sensors for more effective monitoring. The required range extends from the
processing of multiple data streams from a simple array of identical sensors to data
from sensors based on entirely different physical phenomena operating asynchro-
nously at vastly different rates.
3. Inferring
Ability to solve problems using embedded knowledge and draw conclusions. This
expert knowledge is in general a combination of theoretical understanding and a
collection of heuristic problem solving rules that experience has shown to be effec-
tive. A smart monitoring system should be able to detect sensor faults to prevent the
use of nonsensical input values for a classification.
4. Learning
Ability to learn from experience to improve the system performance. Learning and
adaptive behaviour are closely combined and without the implementation of learning
strategies adapting will not work.
2.2 Adapting Monitoring Systems
The property of "adapting" is of special interest here and is therefore considered in
greater detail. At this point, it may be helpful to provide an example of monitoring in
which no adaptive capability exists. The roller bearings of a rotating industrial centri-
fuge must be monitored for possible damage. For this purpose, the vibration signals are
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1362
4
frequently recorded by means of an acceleration sensor. If the value exceeds a limit that
has been manually preset on the basis of experience or as specified in standards and
guidelines, an alarm is triggered. The alarm threshold can be adjusted, of course, but its
value then remains constant for the monitoring phase. A system of this kind, which is
still employed for most monitoring applications in industry today, is not adaptive and
certainly not smart. Which prerequisite for this designation is lacking in such a system?
A change in vibratory behaviour can be due to a wide variety of factors that do not re-
sult from a fault in the machine itself. Minor conversion work may have been performed
on the machine, for instance, in the course of maintenance and servicing; such measures
can cause a change in behaviour without resulting in a malfunction. The following lev-
els of adaptability are conceivable in conjunction with monitoring systems(24).
1. Level 1
An adaptive monitoring system is capable of recognising variations in the surround-
ings and process conditions. Modifications, such as the adaptation of limiting values,
can be performed automatically by the system itself. The previously mentioned
monitoring system for roller bearings remains unaltered on a machine.
2. Level 2
An adaptive monitoring system can be transferred from one machine to another
without the need of readjustment by an expert. Any necessary adjustment work
should be reducible to an absolute minimum. However, the monitoring task itself
should not be altered in this connection. That is, monitoring of a roller bearing is
still the specified task, although the type of roller bearing and the parameters are dif-
ferent.
3. Level 3
An adaptive monitoring system can be employed for other monitoring tasks without
the need of altering the basic structure. The necessary limiting values or control pa-
rameters of the classification algorithm are, to a large extent, specified independ-
ently. At this third level, the monitored object itself can also be varied. The system
that had previously been employed for detecting damage to roller bearings should
now be employed for recognising imbalance in the rotor or for monitoring the proc-
ess. For allowing the system to function at this level, learning strategies are imple-
mented, rather than pre-programmed algorithms for calculating problem-specific
features, such as the effective value of the acceleration.
In its present status, the technology usually does not even attain the first of the levels
just defined. This situation may at first be surprising, and perhaps also somewhat disap-
pointing, but is easy to understand from an engineering standpoint. The decisive exter-
nal parameters causing variations in the monitoring parameters are highly diversified,
and the mutual interactions among them are often unknown, consequently, a considera-
tion of these parameters in a diagnostic model is difficult or impossible(24). Once
trained, a system is capable of performing a monitoring task as long as the prerequisites
for the training status are satisfied. If these conditions change, however, problems will
occur in the monitoring system, and the system must be retrained. For example, a defec-
tive roller bearing generates a typical fault signal. The results of experimental investiga-
tions indicate that the level of vibration generated by a bearing with identical damage
can vary by a factor as high as 100 with different machines. Thus, they cannot be ap-
plied for attaining level 2, that is, transferability to other machines. However, methods
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1363
5
of this kind are typical of industrial monitoring to the present day and are applied in a
wide variety of fields. Slight variations in vibration results from one survey to the next
are assumed to be due to "process conditions", and regarded as not significant. If we
were able to collect relevant quantitative data regarding the "process conditions" exist-
ing at the time that the vibration data were collected, and correct the vibration data for
these conditions, then our diagnostic capability would become far more accurate and
adaptive(2). The aforementioned level 3 currently exists only as a vision in the minds of
researchers.
2.3 Requirements and Acceptance of Adaptive Monitoring Systems
If monitoring problems are already being solved sufficiently well today without adap-
tive behaviour, a logical question is why there is a demand for smart systems at all and
whether this demand can be justified? For this purpose, the type of monitoring currently
applied in industry must first be considered. From a methodical standpoint, this kind of
monitoring no longer satisfies all of the requirements that must be imposed on a modern
monitoring system. The concept of preventive maintenance implies the application of
techniques for the early detection of faults and thus the implementation of appropriate
maintenance measures in due time. As far as the monitoring of machines in industry is
concerned, however, a change of this kind has hardly taken place at all in practice.
Changes in the process or in the machine, of the kind not resulting from a fault, must be
distinguished from ones that do result from a fault. Precisely this adaptivity at level 1
still presents serious problems for many systems, however. If a given system functions
correctly, weeks or even months are often necessary for transferring this property to
similar machines. Support for the expert in charge by the system itself is severely lim-
ited or completely absent. The requirement for human experts and the associated labour
costs still severely restrict the acceptance of monitoring systems. On the other hand,
precisely this situation offers a special opportunity for those who develop and supply
monitoring systems, since an adaptive system becomes independent of particular appli-
cations and can thus provide specialised solutions at acceptable prices(24).
3. Learning Procedures for Smart Adaptive Monitoring Systems
Besides adaptivity, the learning ability is a decisive factor for a smart technical system.
To a certain extent, these two properties are inseparably related, since the previously
mentioned concepts of learning anew or relearning for achieving adaptive behaviour are
not possible without the implementation of learning algorithms. The two fundamentally
distinct approaches of "supervised" and "unsupervised" learning also apply to monitor-
ing systems. The following three learning objectives are essential for a monitoring sys-
tem(24):
1. Number of fault classes and quality states
Especially in the case of machine monitoring, faults are often the result of a gradual
transition from a good to a poor condition, rather than instantenous occurrence. Con-
sequently, the number of status classes to be selected is not clearly evident. For clas-
sical monitoring with limiting values, only two classes are employed. The transition
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1364
6
then occurs very quickly, and the operator's indicating lamp suddenly changes from
green to red. However, one would expect an entirely different kind of information
behaviour from a smart system. A possible solution is the introduction of additional,
intermediate states. Thus, an alternation class can improve the transition between
"good" and "poor". Another possibility is the use of fuzzy classifiers, which allow
the appraisal of a condition not only for a status class, but also for gradual differen-
tiations. In this case too, however, the number of fuzzy classes to be defined must
first be specified. If one prefers to apply an algorithm for controlling the perform-
ance of this task, rather than doing it oneself, the use of "Kohonen feature maps" or
Artificial Immune Systems are advisable. The appropriate number of class parame-
ters can be determined automatically and a change in the number of classes is also
conceivable, if new error classes occur.
2. Limiting values and classification parameters
For dealing with these parameters, training is certainly the most important task for
monitoring systems. Manual entry and management of several hundred limiting val-
ues is not practicable; however, the need for such a large number of values is quite
normal with the use of many features. Of course, a system capable of learning and
which has recognised that new classes are necessary for the monitoring task can also
be trained to employ the associated classifiers in a second step.
3. Features
An adaptive system should recognise the condition that the previously applied fea-
tures no longer suffice for solving a monitoring problem with sufficient reliability.
Hence, a further learning objective is the calculation - or at least a new selection - of
new features which are better suited for the purpose. For achieving this objective,
the application of automatic selection methods is necessary. The selection has to
consider constraints of the classification algorithm. In this sense only so called
wrapper approaches for the feature selection process will find suitable feature com-
binations.
Figure 2. Trendline showing manually set alarm levels and adaptive levels (bold
lines) and the overall vibration measurement data of rotating machine(24)
One of the major problems is the manual setting of alarm levels. As two identical ma-
chines will not run or wear out in the same time the levels need to be adjusted during the
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1365
7
machine lifetime. There are several approaches to calculate the levels automatically
based on historical data concerning outlier elimination 3. Even the use of a moving av-
erage with adjustable window length and different methods of the exponential weight-
ing of data in the window will provide very sufficient results. Comparing the two bold
printed solid lines in Figure 2, it can be seen that the levels set by the algorithm are ap-
proximately the same as those set manually. Events where the vibration level was higher
than the current alarm level are identified in both cases. The markers indicate measure-
ment points where the vibration is below or above the level(24).
3.1 Signal Processing Techniques for Adaptive Monitoring Systems
The measurement parameters traditionally used in condition monitoring are displace-
ment, velocity and acceleration, i.e. )0(
xx ≡,)1(
xx ≡
&and )2(
xx ≡
&& . The first time deriva-
tive of acceleration, i.e. the jerk )3(
xx ≡
&&& , is commonly used to examine the comfort of
travelling in vehicles. The parameters )3(
x and )4(
x are very suitable for the condition
monitoring of slowly rotating bearings(3,4,5, 6, 7). This is due to the fact that although the
acceleration pulses are weak and occur at long intervals, the changes in acceleration are
rapid and become emphasised upon differentiation of the signal )2(
x. The use of frac-
tional derivative )(
α
xallows stepless differentiation(6). The general suitability of this
technique even in practical real world applications is without any controversy. An ad-
vantage of this method is the ability to obtain integer, real or complex order derivatives,
which all can be used for machine diagnosis.
High frequency vibrations are excited in roller bearings as the rolling element passes a
damaged area. Figure 3a shows the raw acceleration time signal )2(
x gained by an accel-
erometer with a length of 0.5 s, which is in accordance to a number of 65536 digital
data and a sampling rate of 131072 Hz. From the original and non-filtered time domain
signal, it is not possible to detect a high frequency vibration. But after zooming in
neighbourhood of 0.04 s, one can see how the signal structure changes at about 0.041 s
(Figure 3b). Another way to make this change noticeable is calculation of the amplitude
spectra of the signal with data sets before and after the excitement. Two sets (each 1024
data points) were taken out from )2(
x. Both have the same length. The first data set ends
at 0.04 s and the second one starts at 0.04 s. The comparison of amplitude spectra in
Figure 3e shows that the high frequency vibration generated by the rolling contact be-
tween ball and the damaged surface is not present before the excitement. The spectrum
on the right shows the high frequency vibration in 40 – 60 kHz range shortly after the
excitement. High frequency vibration can be shown also via filtering. In Figure 3c and
3d filtered signal is shown by two different band pass settings (1-40 kHz and 10-
40kHz). The first band pass filtered signal in Figure 3c (1-40 kHz) does not signifi-
cantly differ from the raw signal in Figure 3a. A significant change can be seen in Fig-
ure 3d, in which the second band pass filtered signal (10-40kHz) is shown, at about
0.041 s. The similar effect can be obtained via second derivative of the acceleration sig-
nal(8,9). The advantage is that it is not necessary to filter the raw signal (Figure 3f). This
method can be very helpful in searching small faults in roller bearings, hence high fre-
quency vibrations are excited via small damages.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1366
8
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-100
0
100 x(2)(t) raw signal
x(2) (m/s2)
0.03 0.032 0.034 0.036 0.038 0.04 0.042 0.044 0.046 0.048 0.05
-40
-20
0
20
40 zoom in x(2)(t), from 0.03 to 0.05 s
x(2) (m/s2)
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-100
0
100 filtered signal, bandpass 1 - 40 kHz
x(2) (m/s2)
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-10
0
10 filtered signal, bandpass 10 - 40 kHz
Time (s)
x(2) (m/s2)
020 40 60
0
2
4
61024 points, left side of t= 0.04 s
Frequency (kHz)
x(2) (m/s2)
020 40 60
0
2
4
6
1024 points, right side of t= 0.04 s
Frequency (kHz)
x(2) (m/s2)
00.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
-2000
0
2000
2. derivative of raw signal
Time (s)
x(4) (Gm/s4)
Figure 3. Example for demonstrating the usefulness of higher derivatives(25).
(b)
(a)
(c)
(d)
(e)
(f)
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1367
9
4. Application - Roller Bearing Diagnostics
The idea suggested here is to use only parts of the time signal which stand in close rela-
tion to possible bearing faults. For that purpose a peak in the time signal is defined as a
local extremum in the measured acceleration signal )2(
x. All peaks in the time signal
which fulfil the requirements of the peak definition are detected and for each all signifi-
cant information like position and amplitude, and adjacent data points are stored. This
data are used as an input for the calculation of features. The calculation of the fourth
derivative x(4) was carried out by using a fourth-order centered difference formula on
uniform grid(9). Therefore a set of adjacent data points on the left and right side of each
peak have to be considered. Now the calculation of different features on the basis of this
set of peak information containing position and amplitude of peak, and adjacent data
points is possible.
A feature vector is defined as a set of parameters extracted from the considered signal,
which gives indications about the current state of the operating system. In condition
monitoring statistical methods have been widely used for investigation, where measured
data are time series. Extensive literature is available on diagnostic techniques using
RMS, Kurtosis, Crest Factor and histograms and other statistical moments (10,11,12,13).
The method suggested here uses peaks as a source of information corresponding to
bearing faults. Possible features may be calculated from the ratio of local maxima of
measured )2(
x and )4(
x. Also the distance variation between local maxima on adjacent or
non-adjacent locations over a predefined offset value can be considered. Further fea-
tures can be obtained by the number of local maxima with absolute values, which are
over some predefined threshold values. The histogram of peak amplitude and peak dis-
tance distribution are additional features which could be considered. Taking the norm of
signal values - for instance with root mean square - may be useful to eliminate the influ-
ence of signal energy of the measured impact sound. A selection of features from a total
of 32 is listed in Table 1 in Section 4.1.
4.1 Detection of small bearing fault in simple demonstrator
For the investigation of the method described in Section 4 a simple bearing test rig was
used. In Figure 4, on the left the very simple test assembly is shown. The components
are the outer race of a roller bearing and a cage, in which the outer ring is mounted.
Only one ball driven by compressed air is rotating and the inner ring is replaced by a
whole shaft with eight nozzles around it. The vibration signal is measured by an accel-
erometer mounted at the outer ring. This assembly allows the measurement of the iso-
lated vibration generated by the rotating ball and all other sources of excitation are omit-
ted. The pathway from the source of vibration and the sensor is well defined and the
number of join patches is minimized. Figure 4, right side, shows the fault in the outer
race, which was induced by creating a small groove using electric spark erosion.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1368
10
Figure 4. (left) Test rig is air driven and consists of outer race of a deep groove ball
bearing type 6310, cage and accelerometer. (right) Outer race of the ball bearing.
A point fault (diameter 510 µm) was introduced using electric spark erosion.
The comparison of the signals in Figures 5 (bearing without a fault) and 6 (faulty bear-
ing) indicates that a separation of the signal from a bearing without a fault and the outer
race groove is possible in the time and frequency domain without any problems. The
overall vibration level differs significantly. To eliminate the influence of the vibration
level signals are normalised. The impulse of the ball in contact with the groove excites
natural frequencies of the bearing and the surrounding elements (assembly parts).
00.1 0.2 0.3 0.4
-20
0
20
x(2) (m/s2)
Time (s) 010 20 30 40
0
0.5
1
x(2) (m/s2)
Frequency (kHz)
good bearing
00.1 0.2 0.3 0.4
-10
-5
0
5
10
x(2) rms-normed
Time (s) 010 20 30 40
0
0.05
0.1
0.15
0.2
x(2) rms-normed
Frequency (kHz)
good bearing
Figure 5. Time signals and the amplitude spectra of intact outer race. Above
measured signal, below signal was normalised by its rms value(9) .
510 µm
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1369
11
The amplitude spectra in Figures 5 and 6 have amplitudes in a frequency range up to 30
kHz, which is typical for a small damage size but also for a random excitation of an
intact bearing. One has to consider that the duration of contact between the damage and
the ball is very short in comparison to the complete measurement time. In consequence
the normalised spectra in Figure 5 (lower right) and 6 (lower right) do not show signifi-
cant differences between the two states.
00.1 0.2 0.3 0.4
-500
0
500
x(2) (m/s2)
Time (s) 010 20 30 40
0
2
4
6
8
x(2) (m/s2)
Frequency (kHz)
fault y bearing
00.1 0.2 0.3 0.4
-10
-5
0
5
10
x(2) rms-normed
Time (s) 010 20 30 40
0
0.05
0.1
0.15
0.2
x(2) rms-normed
Frequency (kHz)
fault y bearing
Figure 6. Time signals and the spectra of faulty outer race. Above measured signal,
below signal was normalised by its rms value(9).
All feature combinations in Figure 7 are calculated after the normalisation of )2(
x, and
are well suited to separate the both classes. The distances between the class centres are
much higher than the variation of the feature within a single class. In Table 1 a descrip-
tion of the selected features used in the figures and the corresponding dimensions are
listed. The numbering indicates that these features are only a subset of the complete
feature pool. All feature combinations could be used to design an automatic classifica-
tion algorithm, but it is not the objective of this investigation to test such a classifier(9) .
The two combinations of features in Figure 7 are selected to demonstrate the potential
of the peak features. Comparing the scatter-plots of various feature combinations from
different surfaces, one can observe the separation of data points which belong to differ-
ent surfaces. This property can be used to monitor bearing health state and to perform
damage detection. When using the combination 5 | 8 and 9 | 10 an 100% classification is
possible. In general the selection of feature combinations could be executed by a soft-
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1370
12
ware algorithm, see (14,15,16) . In addition a couple of algorithms are available for a broad
range of classification tasks.
Table 1. Description of features, which are used for roller bearing tests.
No. Description
5 Mean value of the peak amplitudes from )2(
xat local maxima (peak) after
normalisation. )2(
x is the normalised form of )2(
x. Dim: -
7 RMS of )4(
x. )4(
x was built, using local maxima (peak) in )2(
x. Derivation
method was named in Section 4. )2(
x is the normalised form of )2(
xand )4(
x is
the normalised form of )4(
x.Dim: (1/s2)
8 Mean amplitude value of )4(
x. )4(
x was built as described in Feature 7. Dim:
(1/s2)
9 Standard deviation of )4(
x. )4(
x was built as described in Feature 7. Dim:
(1/s2)
10 Mean value of the ratios, which are obtained using the amplitude of the local
maxima (peak) in )(
x2 and )(
x4 at the corresponding position. Dim: (1/s2)
∑
=
=p
n
i)(
i
)(
i
px
x
n
F12
4
10 1 np = Number of peaks
11.5 22.5
0
1
2
3x 1010
Feature 5
Feature 8
0.5 11.5 2
x 1010
0
5
10 x 1010
Feature 9
Feature 10
Figure 7. Scatter-plots of good and faulty outer races in air driven test rig(9). Each
measured time signal was normalised by its RMS value before feature extraction.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1371
13
4.2 Hybrid Systems for Monitoring
An approach of this kind can be useful especially in applications where hybrid monitor-
ing systems utilise various trial solutions for intelligent data processing in a parallel
structure. The following components have already been implemented for improving the
adaptivity of monitoring systems(24):
1. Neural networks (NN)
Neural networks are certainly suited for applications in machine monitoring. Nu-
merous reports on successful applications have been published. Because of their ca-
pability of operating with noisy signals and of representing the non-linear relation-
ships between sources of error and the resulting signal, neural networks are predes-
tined for this field of application. A further advantage is the high potential for gener-
alisation, that is, the capability of reaching a sensible decision even with input data
that have not been learned.
2. Rule-based fuzzy systems
Despite the general reservations with respect to expert systems, the application of
rule-based systems for machine monitoring still makes sense. Especially the possi-
bility of expanding an initially rudimentary system speaks in favour of this ap-
proach. During the initial stage of installation, such a system is very well suited for
representing expert knowledge by means of simple rules in the form of 'if-then' rela-
tionships. The rule base is frequently very broad and applicable to a wide variety of
machines and plants. The degree of adaptivity is therefore relatively high. A general
disadvantage in such a system is the fact that it cannot operate more efficiently than
the rules that are introduced by the human expert. The loss of information resulting
from the conversion of knowledge for deriving the rule base must be taken into con-
sideration here. Thus, it can be concluded that the diagnostic reliability of such con-
trol systems is no better than that of a human expert with average experience. How-
ever, advantages result from the fact that the expert can define or establish rules
even for classes and states for which no data are available.
3. Neuro-fuzzy diagnostic systems
Neuro-fuzzy systems are one key to the implementation of learning capability in in-
trinsically static rule bases. Many possible methods are available for the automatic
generation of rules. For instance, the NEFCLASS(17) software is capable of perform-
ing this function. The rule bases thus derived are often considerably better than those
resulting from a purely linguistic description of expert knowledge. Furthermore, the
dependence of process parameters can be modelled in this form even if it is not pos-
sible for the experts to express it in such a form. Vibration signals can frequently be
analysed only as functions of process situations. These relationships can be recog-
nised and represented by a neuro-fuzzy system if the process parameters are re-
corded together with the vibration signals as a learning set. With the application of a
neuro-fuzzy system, a high priority also results for the simulation, since the gener-
ally valid fuzzy rules thus derived are of special interest here, rather than the exact
numerical values, for instance, of an individual natural frequency of a machine. For
this purpose, an error of a few percent in the result of a numerical simulation is not
important. The decisive features are the variation of the value in modelling a fault
and the conclusions that can be reached from this result.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1372
14
4. Artificial Immune Systems (AIS)
They are a relatively new area of research in Condition monitoring(18,19). There has
been an increasing interest to development algorithms and engineering tools inspired
by natural and human immune system. The classical self/non-self model for the im-
mune system as the classical approach dominating research activities over more than
40 years. There are many publications that discuss artificial immune systems appli-
cations in robotics, optimization, control, computer science and other areas. Various
mechanisms or processes in the human immune system are investigated in the de-
velopment of AIS. In a system of high complexity (because of its many connections
and diversity of equipment) it is difficult to make a complete catalog of all the pos-
sible, and probable, anomalous situations. With its ability to detect and react to
novel situations, the immune system seems to be an adequate source of inspiration to
develop algorithms for early detection of anomalous behavior in mechanical and
electrical systems.
5. Best Practice
Best practices. These two words represent benchmarking standards- nothing is better or
exceeds a best practice(20,24). The area of monitoring is much too broad to give a general
best practice guideline.
But the questions in the following list should be answered before the selection of a
method or combination of various methods, in order to ensure that the methods consid-
ered are really suited for the problem involved. That will not assure the best practice but
the use of suitable methods that represent the state-of-the-art.
1. Are enough data available for all (as many) conditions or qualities (as possible)?
This question must be answered in order to decide, for instance, whether the auto-
matic learning of the classifiers can be accomplished with neural networks. For a
neural network employed as a monitoring system, the situation is more favourable
than for one employed as a process controller. With a monitoring system, an
unlearned condition results either in no indication at all or, in the worst case, in a
false indication, but the process itself is not affected. If only data for a "good" condi-
tion are available, these data can still be employed as starting material for a monitor-
ing operation. All significant deviations from this condition are then detected. The
adaptive system then evolves in the course of the operation with the continuing oc-
currence of new events.
2. How many features are available?
Features should be closely related to the monitored object. How many features of
this kind can be provided? It is necessary to determine the number of the features
from that on a feature selection, and this can only be decided by considering the cur-
rent monitoring task. As a provisional value you can take a number of 10 features. If
more features are available, a selection should be carried out. Notice the existence of
the close relationship between the number of features and the number of necessary
random samples for a lot of classifiers. If it is necessary to reduce dimensionality,
prefer feature selection if possible. The interpretability of new features, for instance,
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1373
15
calculated by a principal component analysis is difficult for both an expert in the
field and most of all for an operator.
3. Can every set of process or machine data be unambiguously correlated with a class
attribute (quality or damage)?
If this is possible, methods of "supervised learning" should be applied. If this is not
possible, or if class information is only available for portions of the data sets, cluster
methods, such as Kohonen networks, should be employed. The problem is especially
difficult if the task involves a very large number of available features, and the asso-
ciated databases cannot be classified by an expert. In such cases two questions must
be considered in parallel: the question of feature selection and that of clustering.
Nearly all methods for the selection of features utilise the classification efficiency of
the learning or test data set as a selection criterion. However, if no class designation
is available, the classification efficiency is not calculable, and so-called filter meth-
ods must then be employed. As a rule, these methods yield features that are decid-
edly less efficient.
4. Do I have enough theoretical background, system information and suitable software
to carry out a precise simulation, which should include fault simulation?
The development of a model-based prognostic and diagnostic model requires a
proven methodology to create and validate physical models that capture the dynamic
response of the system under normal and faulted conditions(21). There are a couple of
model update algorithms that could improve the accuracy of a FEM-Model but you
should never underestimate the work load and expert knowledge which is necessary
for the implementation of the corresponding experiments and program usage.
5. What is of better quality or more reliable? The system model or the response of your
sensors indicating changes in the machine or process?
Having both would be the best situation because you are free to decide on the use of
a forward model, which calculates a prognosis of the system response, and compares
the estimation with the actual sensor signal. This procedure allows an estimation of
the time to failure of the system. Additionally the difference between the sensor and
the model will give you useful information about possible faults. For the inverse
problem, the quality of the model is less important because the describing features
and all condition information are extracted from the sensor signal independently
from the model. For the diagnostics of roller bearings we have a rough model about
the signal structure but in most cases not enough information for the prediction.
6. Do I have a safety relevant application?
Until now there are a couple of problems concerning self-adaptive monitoring sys-
tems in safety relevant applications. These limitations result from problems that
could occur during the evaluation and testing phase of such systems in all conceiv-
able situations. The behaviour today and in future after system adjustment could dif-
fer significantly. This will encounter legal aspects because national and international
testing institutions might have problems to certificate a system. In nuclear power
plants, advanced self-learning and adaptive systems are today additional information
sources but in general not the trigger for an automatic shut-down.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1374
16
7. The cost factor. Do I have sensor technique and communication infrastructure for
on-line monitoring?
Current commonly used sensor technology only permits the most rudimentary form
of signal processing and analysis within the sensor. The use of advanced data analy-
sis techniques will need more signal data. This means that large quantities of data
must be transmitted from the sensor to a separate data collector for subsequent proc-
essing and analysis. If permanent on-line vibration monitoring is required - and for
adaptive techniques it is absolutely necessary - then at present, for anything other
than an overall vibration alarm, the cost of providing the required communications
infrastructure and the data collection and analysis equipment far outweighs the bene-
fits to be obtained(22).
6. Conclusions
In analysing the current status the gap between new research developments in monitor-
ing ideas, methods and algorithms and real-world applications in use in industry cannot
not be neglected. New hybrid diagnostic systems using combinations of expert knowl-
edge, data-driven modelling and advanced soft computing algorithms for classification
and prediction proved their performance in the past. The combination of different tech-
niques is the key to implementing learning ability and in consequence-adaptive behav-
iour. No expert in this field will argue the predominance of these techniques in compari-
son to standard methods that use simple threshold monitoring or a fixed classification
algorithm. But there are still many problems to solve. The automatic determination,
extraction and selection of features that are best suited for a given process is a challeng-
ing field of research in general and also in condition monitoring. Even future research
activities will not lead to an exclusive monitoring strategy because the field of applica-
tion for the various monitoring applications is too broad. But the permanent improve-
ment of the simulation techniques will offer new possibilities by replacing the experi-
mental learning part with computer-based fault simulation. Then we will also find adap-
tive techniques in areas where safety-relevant aspects nowadays prevent a real fault
simulation and therefore possible broader applications of adaptive monitoring systems.
References
1. CSC (2001) Get smart. How Intelligent Technology will enhance our world.
CSC Consulting, www.scs.com.
2. A Adgar, C Emmanouilidis , J MacIntyre, et.al., ‘The application of adaptive
systems in condition monitoring’. International Journal of Condition Monitor-
ing and Engineering Management, Vol. 1, pp. 13-17, 1998.
3. S Lahdelma, ‘On the higher order derivatives in the laws of motion and their
application to an active force generator and to condition monitoring’. University
of Oulu, Research report No. 101, Department of Mechanical Engineering,
1995, (Academic Dissertation), 1995.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1375
17
4. S Lahdelma, J Strackeljan and D Behr, ‘Combination of Higher Order Deriva-
tives and a Fuzzy Classifier as a New Approach for Monitoring Rotating Ma-
chinery’. Proceedings of COMADEM 1999, Sunderland, pp. 231-241, 1999.
5. S Lahdelma, ‘On the Derivative of Real Number Order and its Application to
Condition Monitoring’. Kunnossapito, Vol. 11, No. 4, 1997, pp. 25-28, 1997.
6. S Lahdelma and V Kotila, ‘Real Order Derivatives - New Signal Processing
Method’. Kunnossapito, Vol. 17, No. 8, 2003, pp. 39-42. (in Finnish), 2003.
7. S Lahdelma and E Juuso, ‘Signal Processing in Vibration Analysis’, Proceedings
of Fifth International Conference on Condition Monitoring & Machinery Failure
Prevention Technologies CM 2008 – MFPT 2008, Edinburgh, 15-18 July, pp
867-878, 2008.
8. D Behr and J Strackeljan, ’Method and Device for Categorizing Damage to a
Roller Bearing’. Patent US020080195333A1 , 2008.
9. T Doguer and J Strackeljan, ’New Time Domain Method for the Detection of
Roller Bearing Defects’. Proceedings of International Conference on Condition
Monitoring & Machinery Failure Prevention Technologies CM 2008, Edin-
burgh, July 2008.
10. S Lahdelma and E Juuso, ‘Vibration Analysis of Cavitation in Kaplan Water
Turbines’. Proceedings of the 17 World Congress, The International Federation
of Automatic Control, Seoul, Korea, July 6-11, pp 13420-13425, 2008.
11. S Lahdelma and E Juuso, ’Signal Processing in Vibration Analysis’. Proceedings
of Fifth International Conference on Condition Monitoring & Machinery Failure
Prevention Technologies CM 2008 – MFPT 2008, Edinburgh, 15-18 July, pp
867-878, 2008.
12. S Lahdelma and E Juuso, ’Advanced Signal Processing in Mechanical Fault Di-
agnosis’. Proceedings of CM 2008 – MFPT 2008, Edinburgh, pp 879-889, 2008.
13. E Juuso and S Lahdelma, ’Intelligent Condition Indices in Fault Diagnosis’. Pro-
ceedings of CM 2008 – MFPT 2008, Edinburgh, pp 698-708, 2008.
14. J Handl and J Knowles, ‘Feature Subset Selection in Unsupervised Learning via
Multiobjective Optimization’. International Journal of Computational Intelli-
gence Research, Vol 2, No 3, pp 217-238, 2006.
15. G Jennifer and C Dy, ‘Feature Selection for Unsupervised Learning’. Journal of
Machine Learning Research, Vol 5, pp 845–889, 2004.
16. J Strackeljan, ‘Feature selection methods- an application oriented overview’.
TOOLMET‘01 Symposium, Oulu, Finland, pp 29–49, 2001.
17. Nauck, D., ‘Neuro-Fuzzy Systems for Explaining Data Sets’, In: Gabrys, B.
(eds.) ‘Do smart adaptive Systems exist?’, pp. 305-319, Springer 2005.
18. L Nunes de Castro and F J von Zuben, ‘Artificial Immune Systems: Part 1 - Ba-
sic Theory and Applications’. Technical report TR-DCA 1/1999, 96 p.
19. J Strackeljan and K Leiviskä, ’Artificial Immune System approach for the Fault
Detection in Rotating Machinery’. Proceedings of CM 2008 – MFPT 2008, Ed-
inburgh, 2008.
20. R Smith, ‘Best Maintenance Practices’. http://www.mt-online.com, 2002.
21. C D Begg, C S Byington and K P Maynard, ‘Dynamic Simulation of mechanical
fault transition’. Proceedings of the 54th Meeting of the Society for Machinery
Failure Prevention, pp. 203-212, 2000.
22. S Dunn, ‘Condition Monitoring in the 21stCentury’. http://www.plantmainte-
nance.com/articles/ConMon21stCentury.shtml.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1376
18
23. J Strackeljan and S Lahdelma, ‘Condition Monitoring of Rotating Machines
Based on Fuzzy Logic’. Proceedings of the International Seminar on Mainte-
nance, Condition Monitoring and Diagnostics, Oulu, Finland, 2-3 October, pp
53-67, 2000.
24. J Strackeljan and S Lahdelma, ‘Smart Adaptive Monitoring and Diagnostic Sys-
tems’. Proceedings of the 2nd International Seminar on Maintenance. Condition
Monitoring and Diagnostics, Oulu, Finland, 28-29 September, pp 47-61, 2005.
25. S Lahdelma, E Juuso and J Strackeljan. ‘Vibration Analysis with Generalised
Norms in Condition Monitoring’. AKIDA, Tagungsband zum 7. Aachener
Kolloquium für Instandhaltung, Diagnose und Anlagenüberwachung, 18./19.
November 2008, Aachen, Germany, A. Seeliger & P. Burgwinkel. Aachener
Schriften zur Rohstoff- und Entsorgungstechnik, Band 70, Verlag R. Zillekens,
pp 583-593, 2008.
The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention Technologies
1377
