Вы находитесь на странице: 1из 11

994 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 47, NO.

5, OCTOBER 2000

Recent Developments of Induction Motor Drives


Fault Diagnosis Using AI Techniques
Fiorenzo Filippetti, Member, IEEE, Giovanni Franceschini, Carla Tassoni, Senior Member, IEEE, and
Peter Vas, Member, IEEE

Abstract—This paper presents a review of the most recent present in the drive considered. Induction motors play an impor-
developments in the field of diagnosis of electrical machines and tant role in manufacturing environments, therefore, this type of
drives based on artificial intelligence (AI). It covers the application machine is mainly considered and many diagnostic procedures
of expert systems, artificial neural networks (ANNs), and fuzzy
logic systems that can be integrated into each other and also with are proposed both from industry and from academia [1]–[4].
more traditional techniques. The application of genetic algorithms Some work has also been done on converter-fed induction motor
is considered as well. In general, a diagnostic procedure starts drives in order to realize a fault-tolerant drive avoiding shut-
from a fault tree developed on the basis of the physical behavior down if the load conditions permit this [5]–[7].
of the electrical system under consideration. In this phase, the
The main steps of a diagnostic procedure can be classified as
knowledge of well-tested models able to simulate the electrical
machine in different fault conditions is fundamental to obtain follows:
the patterns characterizing the faults. The fault tree navigation 1) signature extraction;
performed by an expert system inference engine leads to the choice 2) fault identification;
of suitable diagnostic indexes, referred to a particular fault, and
relevant to build an input data set for specific AI (NNs, fuzzy logic, 3) fault severity evaluation.
or neuro-fuzzy) systems. The discussed methodologies, that play This paper describes the various steps and highlights the func-
a general role in the diagnostic field, are applied to an induction tions that can be accomplished by using AI techniques, un-
machine, utilizing as input signals the instantaneous voltages and
currents. In addition, the supply converter is also considered to derlying the advantages and limitations. Different faults of an
incorporate in the diagnostic procedure the most typical failures induction machine are first considered. Then, the paper con-
of power electronic components. A brief description of the various siders some typical faults of the more usual converters which
AI techniques is also given; this highlights the advantages and the can supply an induction machine.
limitations of using AI techniques. Some applications examples
are also discussed and areas for future research are also indicated.
Index Terms—Artificial intelligence, drives diagnostics, elec-
trical machines. II. AI-BASED TECHNIQUES

It should be noted that there are many types of AI-based tech-


I. INTRODUCTION niques [8]–[12]. Some of these use expert systems [8], artificial
neural networks (ANNs) [9], fuzzy logic [10], fuzzy-NNs [11],
T HE monitoring and fault detection of electrical machines
have moved in recent years from traditional techniques to
artificial intelligence (AI) techniques. Such techniques require
genetic algorithms (GAs) [12], etc. The various techniques have
been summarized recently in [13].
a “minimum configuration intelligence” since no detailed anal- The essence of an expert system is the ability to manage
ysis of the fault mechanism is necessary, nor is any modeling knowledge-based production rules that model the physical
of the system required. When an AI technique is used, fault system, while it is a main feature of NNs that they are general
detection and evaluation can be accomplished without an ex- nonlinear function approximators. This function approximation
pert. In the AI-based systems, several quantities are utilized as is achieved by using an appropriate network built up from
input signals: stator currents and voltages, magnetic fields and artificial neurons, which are connected by appropriate weights.
frame vibrations, etc. In general, stator currents and voltages However, the exact architecture of a NN is not known in
are preferred because they allow for the realization of nonin- advance; it is usually obtained after a trial-and-error procedure.
vasive diagnostic systems and the sensors required are usually Fuzzy logic systems are expert, rule-based systems, but they
can also be considered to be general nonlinear function ap-
proximators. In contrast to NNs, they give a very clear physical
Manuscript received January 15, 1999; revised July 6, 2000. Abstract pub-
lished on the Internet July 1, 2000. An earlier version of this paper was presented description of how the function approximation is performed
at IEEE IECON’98, Aachen Germany, August 31–September 4, 1998. (since the rules show clearly the function approximation
F. Filippetti is with the Department of Electric Engineering, University of mechanism). On the other hand, fuzzy-NNs are basically NNs
Bologna, 40136 Bologna, Italy (e-mail: fiorenzo.filippetti@mail.ing.unibo.it).
G. Franceschini and C. Tassoni are with the University of Parma, 43100 with fuzzy features, and it is one main advantage over “pure”
Parma, Italy (e-mail: g-franceschini@cassiopeia.eng.unipr.it; tassoni@cas- NNs that their architecture is well defined.
siopeia.eng.unipr.it). Finally, GAs are not general function approximators, but they
P. Vas is with the Department of Engineering, Aberdeen University, Aberdeen
AB9 2UE, U.K. (e-mail: p.vas@eng.abdn.ac.uk). are stochastic optimization techniques. However, they can be
Publisher Item Identifier S 0278-0046(00)08830-4. used together with neural or fuzzy-NNs, e.g., to obtain optimal
0278–0046/00$10.00 © 2000 IEEE
FILIPPETTI et al.: INDUCTION MOTOR DRIVES FAULT DIAGNOSIS USING AI TECHNIQUES 995

weights in a “pure” NN, or to obtain the membership func-


tions in fuzzy logic systems, etc. Further details of the various
AI-based techniques will be described in Sections VI and VII.

III. FAULT SIGNATURES OF AN INDUCTION MACHINE


It is well known that the electrical and mechanical faults of a
three- phase induction machine supplied by sinusoidal voltages
are linked with the harmonic content of the stator currents, i.e.,
each fault is associated with the presence of specific harmonic
components. Some of these fault frequencies are now shown
below. If there is rotor asymmetry (e.g., broken bars) in a cage
induction machine, or if there is dynamic eccentricity, then the
following spectrum lines are introduced:
Fig. 1. Input current variation for a 5.5-kW machine with a load torque of 30
(1) N starting at 0.5 s.

where is the slip, and is the supply frequency [4]. In order The problem of individualizing and separating the various
to have a detailed analysis of these types of faults, a dynamic fault effects relies on the interpretation of the frequency compo-
model of the induction machine must be used [14], [15]. This nents contained in the current spectrum together with the con-
is because a dynamic model allows the computation of both the sideration of all the machine operating conditions. The task is
sideband currents at frequencies characterizing a very difficult to solve since, in addition to the specific fault con-
rotor fault. ditions, the current spectrum is influenced by other factors, such
The main effect produced by a stator asymmetry, e.g., short as electric supply, noise, and motor geometry; moreover, the ma-
circuit or a static eccentricity, is a negative-sequence component chine can also be under nonstationary operating conditions.
in the input currents [3] It is natural to consider an expert system based on objects and
production rules as a manager of all the diagnostic procedures.
(2) Obviously, the crucial point is the development of a knowledge
base that generally requires both deterministic and heuristic rea-
soning [26]–[29]. The production rules model the logical infer-
In [16], a model suitable for the detailed analysis of stator
ential concepts and allow the system to draw conclusions from
winding short circuits is presented; this model is based on the
the data and to take direct action. The objects model the data
space-vector theory [17]. This includes the effects of space
structures suitably designed to retrieve and incorporate into the
harmonics, which must be included in the model, since there is
knowledge base the basic data of the machine under test (i.e.,
stator winding asymmetry. The amplitude of the negative-se-
type, power, pole number, rotor bar number, etc.) and all the
quence component computed through this model can be related
other data necessary for the diagnostic aims.
to the percentage of the short-circuited turns and to a restricted
number of machine parameters, while it is quite independent of
the machine operating conditions. Therefore, it constitutes the IV. DATA RETRIEVING STRATEGIES
reference variable to inter-turn failure diagnosis.
The first decision is to state the stationary condition of the
Various models referred to dynamic or static eccentricity can
machine. To this aim, a computer program that performs a
be found in the literature [18], [19], leading to anomalous cur-
suitable data acquisition and a short-time Fourier transform can
rent spectrum lines as well. However, these models allow one
be activated. This transform works by dividing the signal into
to consider a higher order effect of the asymmetry than the first
consecutive or overlapping portions and returns the sequence
one. Therefore, the experimental procedures based on the vari-
of the corresponding fast Fourier transform (FFT). Note that
ation of the slot lines are preferred for the eccentricity detection
this condition allows one to include speed oscillations as well.
[20], [21].
By elaborating this sequence or by visual inspection as well,
The mechanical damages referred to bearing failures intro-
it is possible to state a time after the machine is in the steady
duce harmonic components in the current spectrum at the fre-
state [25], [30]. As an example, Figs. 1 and 2 show, for a
quencies
5.5-kW healthy machine, the stator current variation and the
(3) corresponding short-time Fourier transform. It can be seen that
the FFT is steady after approximately 1 s, as the input current
where is a mechanical frequency depending on the type of profile confirms this.
fault and on the bearing characteristics [4], [22]. [23]. Other Some components of current spectrum depend on the ma-
components at different frequencies can be introduced by load chine speed or slip, therefore, to avoid the need to measure the
anomalies and it should be noted that they can cause misunder- machine speed using conventional transducers (observing, con-
standings, because these lines can be confused with the ones due sequently, the noninvasive criteria), the machine diagnostic ex-
to machine failures [24], [25]. pert system has to follow the task architecture shown in Fig. 3.
996 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

Fig. 2. Two-dimensional view of the corresponding short-time Fourier Fig. 4. Trend of the input current sideband sum of a 900-kW 6-kV eight-pole
transform. induction machine.

level. Its primary task is to eliminate those components that pro-


vide no useful failure information and perform a component
classification. There are various methods to perform this task
[34]–[36]. The best way is to use a data base which contains the
machine history in terms of trends of well-stated failure com-
ponents, i.e., those related to air-gap eccentricities, stator short
circuits, broken rotor bars, and bearing failures. Moreover, the
expert system has to consider those components which are of
potential interest because of their power and repetition.
Fig. 3. Expert system fragment scheme. The knowledge of a component trend is a crucial point be-
cause it makes the expert system a robust threshold handler
which decides to consider or ignore a particular failure com-
Starting from the time after the machine can be considered to ponent. Fig. 4 shows a typical example related to a 900-kW in-
be operating in the steady state, two acquisitions of the instan- duction machine extensively tested. For simplicity, the figure
taneous signals over different time intervals and with different is referred to the last group of more than 1000 tests performed.
sampling rates are activated. The first one is related to a small The trend of the sum of the two fault components (in percentage
observation interval and to a high sampling frequency. Using the with respect to the component at the supply frequency) at fre-
procedure described in [31] and [32], the slot harmonic lines can quencies is shown [15]. A rotor bar gets effectively
be searched and the slip value can be computed utilizing the first broken during the test reported as the 50th case. It can be seen
slot harmonic line at frequency as that, performing a trend analysis it is easy to determine an at-
tention threshold and to evaluate, consequently, the weight of
a component (in the case of Fig. 4 the determined threshold is
(4) 0.5%).
Obviously, it is difficult for an expert system to manage a
large amount of data. To reduce the computational requirements,
where
the expert system could become a simple threshold handler, by
pole pair number;
managing only the thresholds related to the various failure com-
supply frequency;
ponents. These thresholds can be directly computed by using
rotor slot number.
simplified models or formulas or can be previously evaluated
The second acquisition is related to a large observation interval
by performing separately the trend analysis [31], [32]
and to a low sampling frequency. The choice of these param-
By combining the information derived from the trends of
eters is strictly linked with the slip values, i.e., with the har-
the various harmonic components, from the attention thresholds
monic components to be searched. The criteria to make a suit-
and considering the machine operating conditions as well, the
able choice and to obtain a good frequency resolution are dis-
system can determine directly a possible fault situation, by per-
cussed in detail in [33], where an alternative method for the
forming steps 1) and 2) of Section I. For illustration purposes,
speed determination starting from current spectrum is proposed.
Fig. 5 shows the knowledge base fragment referred to the stator
fault identification presented in [16].
V. SPECTRUM LINE SEARCH AND FAULT CLASSIFICATION
However, an expert system structure can also be considered
At this point, the role of the expert system inference engine more simple than the one already described. For example, this
is to filter the harmonic components and to perform a reduc- is the case when the trends or the attention thresholds are not
tion of the large amount of spectral information to a suitable known or there is reduced computing power. For this purpose,
FILIPPETTI et al.: INDUCTION MOTOR DRIVES FAULT DIAGNOSIS USING AI TECHNIQUES 997

Fig. 6. Fault map (example to distinguish stator, rotor, and bearing damage).

tion between an inference engine, an unsupervised NN and a


conventional technique. While the attention thresholds are com-
puted via simplified models, two AI technologies and one deter-
ministic technology are utilized. In the second case, the expert
system performs some evaluation functions providing the input
FS for the NN that completes the task providing the final fault
Fig. 5. Expert system knowledge base fragment. identification.
It should be noted that, with regard to the training phase of the
the role of the expert system can be restricted to the search network, an unsupervised learning does not require a teacher.
of the harmonic components and the fault classification task is Therefore, all the FSs obtained from online monitoring sys-
omitted. The goal of the inference engine is, then, just the re- tems, (from the expert system inference or, finally, if available,
trieving of a suitable set of harmonic components that can cor- from model-based approaches), can constitute components of
rectly label the so-called failure set (FS). Once the FS is known, the learning data set [31]. The result justifies the effort to col-
at the end of the expert system inference, another AI method- lect these data. In fact, for this purpose, once the learning data
ology can be used for the fault classification task. In fact, an set is built and the learning phase is performed, the NN will con-
unsupervised NN, (which has been previously trained) can be tain and generalize the prior knowledge of the expert system. It
introduced to finally classify the single type of fault. also provides a very fast response in terms of fault classification
Different FSs containing current spectrum lines related to var- and requires a minimum configuration intelligence to perform
ious machine conditions are presented. An unsupervised NN its features [23]. Moreover, if a suitable normalization is per-
clusters the available FS into characteristic regions of an output formed on the input FS, then the trained NN can be applied to a
node lattice. The more the input vectors are different, the more family of similar induction machines.
effective the clustering becomes, and the various regions are Recalling the structure of the expert system production rules
well separated, realizing a fault topographic map that can also (see Section III), it should be noticed that a particular assump-
be utilized immediately by visual inspection. Considering for tion cannot always be declared true in exact terms. Therefore,
simplicity only the rotor broken bars (fault condition 1), stator it is necessary to consider not only “true” or “false” but a range
short circuits (fault condition 2), and the bearing damages (fault between them. This is because the reasoning of the experts is
condition 3), there are three indicators as inputs. Restricting the often expressed not by numbers (crisp values) but by linguistic
attention to a single fault, it is obvious that the input FS, for terms as “small,” “large,” “low,” etc., taken from the natural lan-
bearing damages, should contain a consistent value in the third guage. Another AI methodology, fuzzy logic, gives a complete
position, while the other values should be negligible. On the solution to these problems [13], [39]–[41].
contrary, the FS related to the rotor bar breakage and stator short The essential parts of a fuzzy logic system are a set of lin-
circuits should contain consistent values, respectively, in the guistic rules of the form IF A THEN B, where A and B are lin-
first and second positions. Finally, the FS related to the healthy guistic variables characterized by appropriate membership func-
conditions contains all negligible values. Therefore, the various tions, and an algorithm which can convert a linguistic strategy
categories of input FS have well-defined similarity relations. based on expert knowledge into a fuzzy inference. Generally, an
Thus, the unsupervised NN gets a diagnostic preprocessor for expert system may contain fuzzy and crisp rules or, more specif-
classification purposes [23], [34], [35], [37], [38]. ically, some fuzzy knowledge islands and some crisp knowl-
As an example, Fig. 6 shows the fault map trained by rotor edge islands. Finally, when the investigated problem is too com-
asymmetries (1), stator asymmetries (2), and bearing damages plex to be analyzed by quantitative techniques and the available
(3) together with healthy machine cases ( ) [23]. sources of information are only interpreted qualitatively, the ex-
By using the procedure previously discussed, steps 1) and pert system can be completely fuzzy, and becomes a fuzzy in-
2) are obtained utilizing an inference engine or the combina- ference system.
998 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

It is possible to improve the expert system described above


by performing steps 1) and 2) through the insertion of the fuzzy
logic concepts. However, with regard to step 3), i.e., the fault
severity evaluation, this will be discussed in detail in the next
section.

VI. FAULT SEVERITY EVALUATION


A. General
From a theoretical point of view, the conventional general
method to make the fault severity evaluation is the objective
knowledge based on well-tested mathematical models that have
the capability of considering the various machine faults. Unfor-
tunately, a heavy and time-consuming computational task must
be performed due to the models’ complexity and, as discussed
above, a large number of machine design parameters is required
by the models [14], [16]. Fig. 7. ANN architecture for stator short-circuit diagnosis.
Different algorithms can also be utilized, e.g., parameter
variation or state variables variation consequent to the faults learning procedure. However, the ANN sensitivity to synthesize
[42], but, parameters and state variables determined by con- the system depends on the chosen input variables which must be
ventional models depend on model accuracy and sensitivity. optimal fault indicators. In order to fulfill this condition and to
Consequently, fault-parameter variation or fault-state variable obtain an effective method, a preventive-model-based study can
variation correlation is often difficult to evaluate [43]. be performed and this justifies the previously mentioned impor-
By remaining in the field of the objective knowledge, an in- tance of the model approaches. Thus, the input–output data set
teresting possibility is the development of simplified faulty ma- can be obtained, as well as from experimental data, and also
chine models. These models have a low level of requirements in from model-based approaches [46], [50].
terms of design parameters and computational features; more- As an example, Fig. 7 shows the ANN architecture to quan-
over, being oriented to only one fault effect, they can be devel- tify a stator short-circuit condition [51]. Here, and de-
oped to obtain for the specific fault a sufficient accuracy. By note the negative- and positive-sequence stator currents, respec-
adopting suitable assumptions, it is also possible to derive basic tively, and and are the slip and rated slip, respectively. The
models which are characterized by a very simple link between negative-sequence component is quite independent of the
the main indicators and the fault severity (often the link can be operating conditions and it constitutes a reference variable to
reduced to only a formula) [44], [45]. inter-turn failure diagnosis, while the positive component is
Obviously, the simplified models have to be validated exper- dependent both on the short-circuit percentage and slip value,
imentally to state the validity limits, but when confined into the but the utilization of , , and does not lead to the NN con-
limits, they can be used (as mentioned in the previous sections) vergence. However, a suitable process of leads to another ef-
for a first specific fault quantization or a threshold overcoming fective variable, and the diagnostic index is the modulus differ-
evaluation. As an example, the simplified approximated formula ence between the phasors and (positive-sequence com-
which links the main diagnostic index related to rotor asymme- ponent of the healthy machine in the same operating condition)
tries and labeled as with the broken bar number , being divided by that is ( ). This variable is inde-
the total bar number, is given by [46] pendent of the operating conditions and its utilization improves
the network convergence. Therefore, Fig. 7 (where is the slip,
(5) is the rated slip, is the rated current, and is the output
fault percentage). Furthermore, if a suitable normalization of the
At this point, a perplexity arises. Why are we scanning the diagnostic indexes is provided and simplifying assumptions are
conventional methods based on mathematical models, while our introduced, an ANN can be obtained which can diagnose a class
aim is to illustrate innovative AI methods? The reason is that a of similar machines.
model-based approach is sometimes required to obtain effective
C. Fuzzy-Logic-Based Diagnosis Examples
AI methods, as will be discussed below.
The application of an ANN leads to a problem similar to the
B. NN-Based Diagnosis Examples one discussed for the expert system production rules in Sec-
A very new and attractive solution for the fault severity eval- tion V. In fact, the concrete reasoning is often expressed not
uation can be found applying again one of the AI techniques, by numbers (crisp values) but by linguistic terms as “small,”
i.e., the supervised neural approach [13], [46]–[49]. A super- “large,” “low,” etc., taken from the natural language. The crisp
vised ANN is a general function approximator which can syn- values of the fault indicators obtained by fault models, may be
thesize the relationships between the different variables consti- seen as a drawback of the diagnostic system, due to the general
tuting the input vectors and the output diagnostic indexes that requirement to obtain heuristic or qualitative information on the
indicate the fault severity, starting from examples utilized in the system. To approach the performance of a diagnostic system to
FILIPPETTI et al.: INDUCTION MOTOR DRIVES FAULT DIAGNOSIS USING AI TECHNIQUES 999

Fig. 9. 3-D map of the input–output relationships between the sideband


Fig. 8. Input variables fuzzy sets for I . components I and I .

the real world and to obtain a more direct intelligence, again, 6) If ( is medium) and ( is large) then (one_or_two).
an AI technique, i.e., fuzzy logic can provide the solution [9], 7) If ( is large) and ( is small) then (one).
[40], [41]. A system based on fuzzy logic allows the transforma- 8) If ( is large) and ( is medium) then (one_or_two).
tion of heuristic and linguistic terms into numerical values via 9) If ( is large) and ( is large) then (two).
fuzzy rules and membership functions and is able to approxi- Fig. 9 provides a three-dimensional (3-D) map of the input
mate the complex relationships related to a diagnostic task. To and output variables. The details of the fuzzy diagnostic im-
develop a specific example, the detection of broken rotor bars provement for rotor fault detection are given in [13] and [41].
fault severity is considered by utilizing Mamdani-type fuzzy in- By comparing the NN-based approach with the fuzzy ap-
ference and using as input variables the fault components and proach, it is possible to draw the following conclusions. A fuzzy
at frequencies provided into the FS by the expert approach models a complex problem employing an IF–THEN
system [41]. type of expert rules and linguistic variables to capture directly
A model-based analysis, previously performed on the the qualitative aspects of the human reasoning process involved.
500-W test motor, gives the guidelines for the choice of the However, the problem is shifted to the membership function and
linguistic variables to evaluate the fault components severity. rule tuning. On the contrary, an NN-based technique allows an
For simplicity, we consider only the labels “small,” “medium,” accurate solution to a particular fault problem, without the need
and “large” are considered. The simulated data suggest to of accurate knowledge of the faulty system. However, the main
consider “small” to an amplitude up to 50 dB, both “small” drawback is related to the fact that, in general, the exact archi-
and “medium” up to 40 dB, etc., until 30 dB that can tecture of the NN to be used is not known in advance.
be considered “large.” Obviously, the amplitude of the fault
components increases with the number of broken bars. By
D. Fuzzy-NN-Based Diagnosis Examples
selecting generalized bell-shaped membership functions with
suitable parameters, it is possible to obtain the membership To overcome the already mentioned problem and also to ob-
functions for the two fault components and . The shape of tain further possibilities, it is necessary to join the features of
the fuzzy sets is then shown in Fig. 8. NNs and fuzzy techniques, and this can be performed by using
At this stage, the output fuzzy set related to the broken bars a neuro-fuzzy (fuzzy-neural) approach [52]. One possibility for
must be defined. Following again the simulation results and this is to use an (adaptive-network-based fuzzy inference system
recalling that a diagnostic system must be able to detect the (ANFIS). This NN is an adaptive network which implements,
threshold between healthy and faulty conditions, it is sufficient as overall input–output function, a fuzzy inference [11]. In this
to consider, in terms of linguistic variables, the following way, it is possible to develop a fuzzy system embedded in an
five conditions: “no broken bars” no; “an incipient fault” adaptive NN, without the extra effort to build and tune the mem-
if; “one broken bar” one; “one_or_two broken bars” bership functions and the rule parameters. The architecture of
one_or_two; and, finally, “two broken bars” two. For the the ANFIS is well defined.
output fuzzy sets, five Gaussian functions centered, respec- An example of ANFIS application can be derived directly
tively, in suitable points have been chosen. By considering the from the previous example related to the induction motor rotor
two-dimensional matrix of the input variables, each subspace faults [53]. For simplicity, a first order Sugeno-type-inference
is associated with a fuzzy output situation. Consequently, the system is used, which is more suitable for the ANFIS applica-
fuzzy rules are obtained as follows. tion. Such reasoning requires rules with the IF part fuzzy and the
THEN part crisp. Assuming again the inputs to be and , the
1) If ( is small) and ( is small) then (no). first rule of the previous fuzzy system becomes the following:
2) If ( is small) and ( is medium) then (if). If ( is small) and ( is small) then
3) If ( is small) and ( is large) then (one). where , , and are crisp parameters of the THEN part. The
4) If ( is medium) and ( is small) then (if). other rules are similar (but contain different , , parame-
5) If ( is medium) and ( is medium) then (one). ters). The adaptive network structure consists of a multilayer
1000 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

VII. FAULT DIAGNOSIS OF DRIVES


A. Open-Loop Drives
If the induction machine is supplied by a frequency converter,
the machine diagnostic system has to be combined with the fault
possibilities of the converter. For this purpose, an integrated di-
agnostic system associated with the monitoring of an open-loop
electric drive will be considered.
As far as power converters for commercial applications of
electric drives are concerned, a relatively small number of con-
figurations exists. Therefore, similar circuit topologies should
produce similar results for typical faults and, hence, it is pos-
sible to study a diagnostic system able to operate regardless
of different manufacturers and ratings. Voltage-source-inverter
(VSI)-type converters are now considered. To investigate pos-
Fig. 10. Adaptive ANFIS architecture for rotor fault diagnostics based on the sible fault operations of such a converter, much work has been
sideband components I and I . carried out in recent years. The literature concerns mostly the
classification and the analysis of true fault operations of single
feedforward network in which each node performs a specific components of electric systems [54]. The analytical approach to
function (node function). To indicate the different adaptive ca- reliability prediction, i.e., the preventive diagnostic, on the con-
pabilities, usually an adaptive node is indicated by a square (in trary, is developed only regarding subsystems [7], [55].
Fig. 10). The adaptive NN is trained by a supervised learning. The results of the analysis are considered as the starting point
However, in this case, (different from the classical feedforward to study the monitoring and diagnostic system that will follow
NN architecture), the links only indicate the flow direction of the system operating conditions detecting online abnormal con-
signals between nodes and no weights are associated with the ditions. Obviously, in a multistage system, the nodes chosen to
links. During the learning phase the updating quantities are the be monitored in parallel should be limited, trading off between
parameters of the adaptive nodes. cost and expected results. Therefore, the monitoring strategy
To perform the features of the already defined Sugeno-type must be suitably designed and, to this aim, correlation among
fuzzy system, the adaptive NN can be organized as shown in the practical fault set and the different variables should be found.
Fig. 10. In Fig. 10, there are several layers. For example, layer For a VSI, the drive scheme contains three cascaded subsystems:
1 is associated with the membership function of the linguistic an ac–dc converter, a dc–ac converter, and the induction motor.
variable (small, medium, large) for and for . The cor- The effects produced by typical faults on the input–output
responding output specifies the membership grade of the input. variables of the subsystems must be analyzed in order to con-
The parameters of the membership function pertain to the pa- struct the rules for the knowledge base of the diagnostic system
rameter set. As the values of these parameters change, the shape [5]–[7], [56], [57]. Typical faults of a converter cause changes
of the membership functions varies accordingly. in the supply waveforms and, consequently, in the spectrum of
As discussed above, in the conventional fuzzy approach the the input–output currents and voltages. Also, in this case, the
parameters must be chosen by examining directly the fault cases spectrum analysis is a fundamental tool for diagnostic purposes
or the model results. This means that the fuzzy-NN starts with a and every noncharacteristic component becomes a possible di-
first attempt of membership functions, then automatically con- agnostic index. Therefore, the methodologies already discussed
structs the required membership functions by using an example for the various motor faults can be extended to the supply con-
set. The other layers are devoted to the remaining functions of verters.
the fuzzy set. Finally, the last layer that computes the overall Considering the usual drive protections and, consequently,
output as the summation of all incoming signals. the available variables, the subset of monitoring nodes is con-
Thus, an adaptive network has been constructed which is stituted by the input–output variables of the subsystems, i.e.,
functionally equivalent to a fuzzy inference system, but without • main supply currents;
the effort to tune the rule parameters and the membership • dc-link current and dc-link voltage;
function parameters. This task is accomplished by the NN • dc–ac converter output currents (input motor currents).
during the learning phase. Since the parameter set characterizes Some typical true fault events that allow reduced operating
the fuzzy system features, it can be concluded that, in this conditions of the drive without involving the short circuits pro-
case, the adaptive fuzzy-NN is able to construct autonomously tections are:
the embedded fuzzy system. The authors have already shown • transistor base drive open fault;
in [51] the utilization of this interesting technique for stator • phase open at machine terminal;
short-circuit detection as well. It has to be emphasized that this • any motor asymmetry due to an internal fault.
technique is very interesting because starting from examples, it As an example, Fig. 11 shows the experimental results re-
requires a minimum a priori knowledge and, consequently, can ferred to a machine open phase in comparison with healthy con-
be utilized in order to design a diagnostic system respecting the ditions. Fig. 11(a) and (b) shows the spectrum and instantaneous
minimum configuration AI (MCAI) principle [13], [23]. values of main current and motor current related to the no-fault
FILIPPETTI et al.: INDUCTION MOTOR DRIVES FAULT DIAGNOSIS USING AI TECHNIQUES 1001

Fig. 11. Experimental spectra and instantaneous supply current and output converter current in (a), (b) healthy condition and (c), (d) fault condition.

condition. Fig. 11(c) and (d) shows the corresponding results in chine condition. The typical spectrum lines produced by faults
the case of an open phase. in the machine input currents are present in the voltage spec-
Significant changes for the two situations are clearly visible trum as well. The manipulated variables are available beyond
in both the instantaneous shapes and the spectra. Following the usual motor terminal and can be sensed for diagnostic pur-
the diagnostic philosophy already utilized for the machine poses. Since the regulators in the presence of faults will still
faults, the current spectrum anomalies introduced by the con- force the controlled variables to the reference values, the con-
verter faults can be assumed as robust indicators. Obviously, troller outputs change, reflecting fault entity and type. There-
an efficient diagnostic system should give information about fore, new diagnostic indexes can be retrieved for manipulated
the component’s incipient failures, whose early detection can variables, too.
avoid catastrophic events. To this aim, the parametric failures An unlike philosophy has been followed in order to design
of power switching components should be considered and the and to realize the control of fault-tolerant systems [61], [62].
goal of the diagnostic system becomes the detection of pa- Also, in this field, advanced AI techniques start to be applied
rameter variations effects. However, this task is difficult. In [13]. In the following, a genetic-based technique is described.
fact, experimental and model-based investigations show that 1) GA-Based Self-Repairing Drives: As discussed in
if the more usual long-term failures are considered in the Section II, GAs are stochastic optimization techniques. More
switching components, i.e., the related parameter variations, precisely, they are simple, powerful, general purpose, deriva-
the input–output current, and voltages spectra are not effective tive-free, stochastic global optimization methods (search
fault indicators. No significant changes can be distinguished algorithms) inspired by the laws of natural selection and
between the healthy condition and the parametric faulty con- genetics. Thus, they use the concept of Darwins’s theory of
ditions in the spectrum line amplitude without considering its evolution, which is based on the rule of the survival of the
time variations [7]. This problem could be solved by applying fittest. These algorithms are derivative free, which means that
trend analysis techniques. they do not need functional derivative information to search
for a set of parameters that minimize (or maximize) a given
B. Closed-Loop Drives objective function. Instead, they exclusively rely on repeated
evaluations of the objective function, and the subsequent search
For closed-loop drives, only recently have some papers ap- direction after each evaluation follows certain heuristic guide-
peared in the literature about the impact of control loops on lines. Since the search procedure does not use any derivative,
the machine fault effects [58]–[60]. Obviously, the control itself in general, this technique is slower than derivative-based opti-
modifies the behavior of machine supply variables and a more mization methods. However, the freedom from the dependence
sophisticated procedure must be adopted in order to assess ma- of functional derivatives means that there is no requirement for
1002 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

differentiable objective functions, so the objective function can [3] S. Farag, R. G. Bartheld, and T. G. Habetler, “An integrated on-line
be as complex as required. motor protection system,” in Conf. Rec. IEEE-IAS Annu. Meeting,
Denver, CO, Oct. 1994, pp. 117–122.
The data processed by a GA is a set (population) of binary bit [4] P. Vas, Condition Monitoring, Diagnosis and Parameter Estimation of
strings (chromosomes) which represent multiple points in the Electrical Machines. Oxford, U.K.: Oxford Univ. Press, 1993.
search space. A bit string is a binary figure with a finite length, [5] V. Rajogopalan, K. Debebe, and T. S. Sankar, “Expert system for fault
diagnosis of VSI fed AC drives,” in Conf. Rec. IEEE-IAS Annu. Meeting,
(bit strings are combinations of 0s and 1s, which represent the Dearborn, MI, Oct. 1991, pp. 368–373.
value of a number in binary form); each bit is called an allele [6] D. Kastha and B. K. Bose, “Investigation of fault modes of voltage-fed
(gene) and is decoded by an evaluator to obtain the objective inverter system for induction motor drives,” IEEE Trans. Ind. Applicat.,
vol. 30, pp. 1028–1038, July/Aug. 1994.
function value of the individual point in the search space. This [7] F. Filippetti, G. Franceschini, and C. Tassoni, “Integrated diagonstic sys-
function is minimized (or maximized) by the GA and is then tems for failure identification in power converter,” in Proc. EPE’95, vol.
converted into a fitness value, a value proportional to its near- 3, Seville, Spain, Sept. 1995, pp. 270–274.
[8] F. Hayes-Roth, D. A. Waterman, and D. B. Lenat, Building Expert Sys-
ness to its solution. tems. Reading, MA: Addison-Wesley, 1983.
The idea of GA-based self-repairing drives was first dis- [9] B. Kosko, Neural Networks for Signal Processing. Englewood Cliffs,
cussed in [13]. According to this, it is possible to implement NJ: Prentice-Hall, 1991.
[10] L. Zadeh, “Fuzzy sets,” Inform. Contr., no. 8, pp. 338–353, 1965.
AI-based self-repairing and self-constructive controllers and [11] J. R. Jang, “ANFIS: Adaptive network based fuzzy inference system,”
estimators, e.g., in variable-speed “sensorless” high-perfor- IEEE Trans. Syst., Man, Cybern., vol. 23, pp. 665–685, May/June 1993.
mance drives. To obtain an extremely flexible self-repairing [12] P. E. Goldberg, Genetic Algorithm in Search Optimization and Machine
Learning. New York: Addison-Wesley, 1989.
and self-constructive (self-reproducing, self-configuring) con- [13] P. Vas, Artificial-Intelligent-Based Electrical Machines and Drives. Ap-
troller/estimator, a field-programmable gate array (FPGA) can plication of Fuzzy, Neural, Fuzzy-Neural and Genetic-Algorithm-Based
be used, which contains simple logic cells, each of which can Techniques. Oxford, U.K.: Oxford Univ. Press, 1999.
[14] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “Transient model
perform simple logical operations. The operation performed oriented to diagnostics of induction machine with rotor asymmetry,” in
by each cell and the way its inputs and outputs are connected Proc. ICEM’94, vol. 3, Paris, Sept. 1994, pp. 62–67.
with the other logic cells of the FPGA are determined by an [15] , “AI Techniques in induction machines diagnosis considering
speed ripple effect,” IEEE Trans. Ind. Applicat., vol. 34, pp. 98–108,
appropriate binary bit string, which is stored in the master Jan./Feb. 1998.
memory of the FPGA. This bit string describes completely [16] G. Gentile, N. Rotondale, F. Filippetti, G. Franceschini, M. Martelli,
the behavior of the FPGA and changing the bit string leads and C. Tassoni, “Analysis approach of induction motor stator faults to
on-line diagnostics,” in Proc. ICEM’90, Cambridge, MA, Aug. 1990,
to another circuit architecture. Some commercially available pp. 491–497.
FPGA chips are extremely flexible: different bit strings cor- [17] J. Stepina, “Non transformational Matrix Analysis of electrical Ma-
respond to different architectures and a new bit string can be chinery,” Elect. Mach. Electromech., no. 4, pp. 255–268, 1979.
[18] A. Stravou and J. Penman, “The on-line quantification of air-gap eccen-
downloaded onto the chip from a host PC at any time. To obtain tricity in induction machines,” in Proc. ICEM’94, vol. 2, Paris, France,
the optimal solution (optimal chip architecture), a GA is used. Sept. 1994, pp. 261–266.
The completely automated self-repairing feature also allows [19] T. J. Sobczyk, C. Tassoni, and P. Vas, “Models for induction motors with
air-gap asymmetry for diagnostic purposes,” in Proc. ICEM’96, vol. 2,
automated diagnosis. Such self-repairing and self-constructing Vigo, Spain, Sept. 1996, pp. 79–84.
systems could revolutionize the design of controllers and [20] J. R. Cameron, W. T. Thomson, and A. B. Dow, “On line current moni-
estimators used in all types of drive systems. toring and analysis to quantify the level of air-gap eccentricity in I. mo-
tors,” in Proc. ICEM’88, vol. 3, Pisa, Italy, Sept. 1988, pp. 205–210.
[21] D. G. Dorrel, W. T. Thomson, and S. Roach, “Analysis of airgap flux
current and vibration signals as a function of static and dynamic airgap
VIII. CONCLUSIONS eccentricity in 3-phase induction motors,” in Conf. Rec. IEEE-IAS Annu.
Meeting, Lake Buena Vista, FL, Oct. 1995, pp. 563– 570.
This paper has given a general review of the recent devel- [22] R. R. Schoen, T. G. Habetler, F. Kamran, and R. G. Bartheld, “Motor
opments in the field of AI-based diagnostic systems in elec- bearing damage detection using stator current monitoring,” in Conf. Rec.
trical machines and drives. It has covered the application of ex- IEEE-IAS Annu. Meeting, Denver, CO, Oct. 1994, pp. 110–116.
[23] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “Integrated con-
pert systems, fuzzy logic, ANNs, and combined systems, e.g. dition monitoring and diagnosis of electrical machines using minimum
fuzzy-NNs. The advantages of using AI were shown and some configuration intelligence,” in Proc. EPE’97, vol. 2, Tronheim, Norway,
aspects related to the application of minimum a priori knowl- Sept. 1997, pp. 983–988.
[24] R. R. Schoen and T. G. Habetler, “A new method of current based con-
edge were also discussed. dition monitoring in induction machines operating under arbitrary load
Although at present they are limited to only a number of prac- conditions,” in Proc. ICEM’94, vol. 2, Paris, France, Sept. 1994, pp.
tical implementations, it is believed that these techniques will 282–287.
[25] F. Filippetti, G. Franceschini, G. Grellet, G. Salles, and C. Tassoni,
have a significant role in electrical drive diagnostic systems; in “Monitoring of an actuator’s load by neural network,” in SDEMPED’97,
particular, fuzzy-neural diagnostic systems will become exten- Carry le Rouet, France, Sept. 1997, pp. 296–301.
sively used. In the future, GA-assisted neural- and fuzzy-neural [26] F. Filippetti, G. Franceschini, M. Martelli, and C. Tassoni, “An approach
to knowledge representation about induction machine diagnostics in ex-
systems may also be used more extensively for machine condi- pert systems,” in Proc. ICEM’88, vol. 3, Pisa, Italy, Sept. 1988, pp.
tion monitoring, e.g., in self-repairing electrical drives. 289–296.
[27] D. Leith, N. D. Deans, and L. I. D. Stewart, “Condition monitoring of
electrical machine using real time expert systems,” in Proc. ICEM’88,
REFERENCES vol. 3, Pisa, Italy, Sept. 1988, pp. 297–302.
[28] P. Calonnec, T. Derrey, E. Destobbeler, and L. Protin, “Induction motor
[1] J. R. Cameron, W. T. Thomson, and A. B. Dow, “On-line current moni- reliability use of fault tree,” in Proc. EPE’95, vol. 3, Seville, Spain, Sept.
toring of induction motors,” in Proc. IEE EMD’87, 1987, pp. 117–122. 1995, pp. 281–285.
[2] J. B. Kliman and J. Stein, “Methods of motor current signature analysis,” [29] R. Isermann, “On the design and control of mechatronic systems—A
Elect. Mach. Power Syst., no. 20, pp. 455–463, 1992. survey,” IEEE Trans. Ind. Electron., vol. 43, pp. 4–15, Feb. 1996.
FILIPPETTI et al.: INDUCTION MOTOR DRIVES FAULT DIAGNOSIS USING AI TECHNIQUES 1003

[30] B. Yazici, G. B. Kliman, W. J. Premerlani, R. A. Koegl, G. B. Robinson, [51] F. Filippetti, G. Franceschini, A. Ometto, and S. Meo, “A survey of
and A. Abdel-Malek, “An adaptive on line, statistical methos for neural network approach for induction machine on-line diagnosis,” in
bearing fault detection using stator current,” in Conf. Rec. IEEE-IAS Proc. UPEC’96, vol. 1, Iraklio, Greece, Sept. 1996, pp. 17–20.
Annu. Meeting, New Orleans, LA, Oct. 1997, pp. 213– 220. [52] P. V. Goode and M. Y. Chow, “Using a neural/fuzzy system to extract
[31] A. Cavallini, F. Filippetti, G. Franceschini, S. Pirani, and C. Tassoni, knowledge of incipient fault in induction motors (Part I and Part II),”
“A global instrument for on-line induction motor integrity assessment,” IEEE Trans. Ind. Electron., vol. 42, pp. 131–146, Apr. 1995.
in Proc. Int. Symp. Stockholm Power Technology—Electrical Machines [53] F. Filippetti, G. Franceschini, C. Tassoni, and L. Collamati, “NN fuzzy
and Drives, Stockholm, Sweden, June 1995, pp. 86–90. logic synthesis in diagnostic systems for electric machines,” in Proc. Int.
[32] A. Cavallini, F. Filippetti, G. Franceschini, and S. Pirani, “A labview Symp. Electrical Machines (SME’96), Krakow, Poland, June 1996, pp.
based virtual instrument for on-line induction motor condition moni- 107–112.
toring,” in Proc. Int. Aegean Conf. Electrical Machines and Power Elec- [54] P. A. Aloisi, “Failure diagnosis in medium power semiconductors,” in
tronics ACEMP’95, Kusadasi, Turkey, June 1995, pp. 278–282. Proc. EPE’91, vol. 3, Florence, Italy, Sept. 1991, pp. 117–119.
[33] L. Collamati, F. Filippetti, G. Franceschini, and C. Tassoni, “A tech- [55] P. Venet, G. Grellet, and G. Rojat, “Identification of significant parame-
nique for induction machine speed detection based on rotor electrical ters relative to failure during operation of a switch-mode power supply
asymmetry frequencies identification,” in Proc. IEEE ICHQP’96, Las application to preventive diagnostic,” in Proc. EPE ’91, vol. 3, Florence,
Vegas, NV, Oct. 1996, pp. 381–385. Italy, Sept. 1991, pp. 112–116.
[34] R. R. Schoen, T. G. Habetler, B. K. Lin, J. H. Schlag, and S. Farag, [56] A. C. Renfrew and J. X. Tian, “The use of knowledge-based system
“An unsupervised on-line system for induction motor fault detection in power ectronic circuit fault diagnosis,” in Proc. EPE’93, vol. 7,
using stator current monitoring,” in Conf. Rec. IEEE-IAS Annu. Meeting, Brighton, U.K., Sept. 1993, pp. 57–62.
Denver, CO, Oct. 94, pp. 103–109. [57] G. Franceschini, G. Gentile, N. Rotondale, C. Tassoni, and M. Tursini,
[35] J. H. Schlang, T. G. Habetler, and B. K. Lin, “An unsupervised neural “An approach to knowledge-base representation in electric drive fault
network fault discriminating system implementation for on line con- diagnosis,” in Proc. ICEM’94, vol. 2, Paris, France, Sept. 1994, pp.
dition monitoring of induction machine using stator current,” in Proc. 358–362.
SDEMPED’97, Carry le Rouet, France, Sept. 1997, pp. 285–289. [58] R. S. Wiser, M. Shagginger, C. Kral, and F. Pirker, “The integration of
[36] W. T. Thomson, “A review of on-line condition monitoring techniques machine fault detection into an indirect field oriented induction machine
for induction motors—Past, present and future,” in Proc. SDEMPED’99, drive control scheme,” in Conf. Rec. IEEE-IAS Annu. Meeting, St. Louis,
Gijon, Spain, Sept. 1999, pp. 3–17. MO, Oct. 1998, pp. 278–285.
[37] J. Penman and C. M. Yin, “Feasibility of using unsupervised learning [59] , “The Vienna I.M. monitoring method; On the impact of the field
NN for the condition monitoring of electrical machines,” Proc. oriented control on real operational behavior of a faulty machine,” in
IEE—Elect. Power Applicat., vol. 141, no. 6, pp. 317–322, 1994. Proc. IEEE IECON’98, Aachen, Germany, Sept. 1998, pp. 1544–1549.
[38] F. Filippetti, G. Franceschini, and C. Tassoni, “A survey of AI tech- [60] A. Bellini, F. Filippetti, G. Franceschini, and C. Tassoni, “Closed loop
niques approach for induction machines on-line diagnostics,” in Proc. control impact on the diagnosis of induction motor faults,” in Conf. Rec.
PEMC’96, vol. 2, Budapest, Hunghary, 1996, pp. 314–318. IEEE-IAS Annu. Meeting, Phoenix, AZ, Oct. 1999, pp. 1913–1921.
[39] T. Terano, K. Asai, and M. Sugeno, Applied Fuzzy Systems. Boston, [61] R. Spee and A. K. Wallace, “Remedial strategies for brushless DC drives
MA: AP Professional, 1994. failures,” IEEE Trans. Ind. Applicat., vol. 27, pp. 259–266, Mar./Apr.
[40] E. Ritchie, E. Xiaolan Deng, and T. Jokinen, “Diagnosis of rotor faults 1990.
in squirrel cage ind. motors using a fuzzy login approach,” in Proc. [62] D. Kastha and B. K. Bose, “Fault mode single-phase operation of a
ICEM’94, vol. 2, Paris, France, Sept. 1994, pp. 348–352. variable frequency induction motor drive and improvement of pulsating
[41] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “A fuzzy logic torque characteristics,” IEEE Trans. Ind. Electron., vol. 41, pp. 426–433,
approach to on-line induction motor diagnostics based on stator current Aug. 1994.
monitoring,” in Proc. Int. Symp. Stockholm Power Technology—Elec-
trical Machines and Drives, Stockholm, Sweden, June 1995, pp.
270–274.
[42] K. R. Cho, J. H. Lang, and S. D. Umans, “Detection of broken rotor bars
in Ind. motors using state and parameter estimation,” IEEE Trans. Ind. Fiorenzo Filippetti (M’00) was born in Fano,
Applicat., vol. 28, pp. 702–709, May/June 1992. Italy, in 1945. He received the Master’s degree
[43] F. Filippetti, G. Franceschini, C. Tassoni, and P. Vas, “Broken bar de- in electrical engineering from the University of
tection in induction machines: Comparison between current spectrum Bologna, Bologna, Italy, in 1970.
approach and parameter estimation approach,” in Conf. Rec. IEEE-IAS He joined the University of Bologna in 1976 as an
Annu. Meeting, Denver, CO, Oct. 1994, pp. 95–102. Assistant Professor and is currently an Associate Pro-
[44] F. Filippetti, G. Franceschini, C. Tassoni, A. Ometto, and S. Meo, “A fessor of Electrotechnics in the Department of Elec-
simplified model of induction motor with stator shorted turns oriented tric Engineering. He is the author or coauthor of more
to diagnostics,” in Proc. ICEM’96, vol. 3, Vigo, Spain, Sept. 1996, pp. than 90 scientific papers. His main research interests
410–413. include the simulation and modeling of electric cir-
[45] F. Filippetti, G. Franceschini, M. Martelli, and C. Tassoni, “Develop- cuits and systems and the application of diagnostic
ment of expert system knowledge base to on-line diagnosis of rotor elec- techniques to circuits and systems.
trical faults in induction motors,” in Conf. Rec. IEEE-IAS Annu. Meeting, Prof. Filippetti is a member of the Italian Association of Electric and Elec-
Houston, TX, Oct. 1992, pp. 92–99. tronics Engineers (AEI) and the European Consortium for Research on Condi-
[46] F. Filippetti, G. Franceschini, and C. Tassoni, “Neural network aided tion Monitoring of Electric Systems and Drives (CRCM).
on-line diagnostics of induction motor rotor faults,” IEEE Trans. Ind.
Applicat., vol. 31, pp. 892–899, July/Aug. 1995.
[47] M.-Y. Chow and C. Mangum, “A NN approach to real time condition
monitoring of induction machines,” IEEE Trans. Ind. Electron., vol. 38,
pp. 449–454, Dec. 1991. Giovanni Franceschini was born in Castelnovo ne’
[48] J. Penman and C. M. Yin, “The application of NN in identifying Monti, Italy, in 1960. He received the Master’s de-
faults in induction machines,” in Proc. Int. Conf. Electrical Machines gree in electronic engineering from the University of
(ICEM’92), vol. 3, Manchester, U.K., Sept. 1992, pp. 1256–1260. Bologna, Bologna, Italy.
[49] M. Y. Chow, R. N. Sharpe, and J. C. Hung, “On the application and He is currently an Associate Professor of Electric
design of artificial neural networks for motor fault detection (Part I and Drives at the University of Parma, Parma, Italy. His
Part II),” IEEE Trans. Ind. Electron., vol. 40, pp. 181–197, Apr. 1993. current research interests include high-performance
[50] F. Filippetti, G. Franceschini, C. Tassoni, S. Meo, and A. Ometto, “NN electric drives and diagnostic techniques for indus-
aided on-line diagnostics of induction machine stator faults,” in Proc. trial electric systems. He is the author or coauthor of
UPEC’95, vol. 1, London, U.K., Sept. 1995, pp. 148–151. more than 70 technical papers in these fields.
1004 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 47, NO. 5, OCTOBER 2000

Carla Tassoni (A’90–SM’92) was born in Bologna, Peter Vas (M’89) received the Ph. D. and D. Sc. de-
Italy, in 1942. She received the Master’s degree grees in electrical engineering from the University of
in electrical engineering from the University of Newcastle-upon-Tyne, Newcastle-upon-Tyne, U.K.,
Bologna, Bologna, Italy, in 1966. in 1979 and 1988, respectively.
She joined the University of Bologna as an As- He has worked at various European universities
sistant Professor and then became an Associate Pro- and, in 1990, he joined Aberdeen University,
fessor of Electrical Machines in the Department of Aberdeen, U.K., as a Professor of Electrical Engi-
Electric Engineering. She is currently a full Professor neering in the Department of Engineering, where
of Electrotechnics at the University of Parma, Parma, he is the Head of the Intellient Motion Control
Italy. She is the author or coauthor of more than 100 Group. He is the author of more than 160 papers and
scientific papers. Her main research interests include five textbooks. His current research areas include
the simulation and modeling of electric circuits and systems and the application sensorless and quasi-sensorless vector and direct torque-controlled drives,
of diagnostic techniques. artificial intelligence applications to motion control, condition monitoring, and
Prof. Tassoni is a member of the European Consortium for Research on Con- parameter estimation.
dition Monitoring of Electric Systems and Drives (CRCM). Prof. Vas was the Laureate of the 1990 George Montefiore International
Award.

Вам также может понравиться