Accepted Manuscript - Universidade Federal de Minas Gerais › documentos ›...
Transcript of Accepted Manuscript - Universidade Federal de Minas Gerais › documentos ›...
Accepted Manuscript
Title: Incipient fault detection in induction machinestator-winding using a fuzzy-Bayesian change point detectionapproach
Authors: Marcos F.S.V. D’Angelo, Reinaldo M. Palhares,Ricardo H.C. Takahashi, Rosangela H. Loschi, Lane M.R.Baccarini, Walmir M. Caminhas
PII: S1568-4946(09)00221-XDOI: doi:10.1016/j.asoc.2009.11.008Reference: ASOC 723
To appear in: Applied Soft Computing
Received date: 20-7-2008Revised date: 27-5-2009Accepted date: 15-11-2009
Please cite this article as: M.F.S.V. D’Angelo, R.M. Palhares, R.H.C. Takahashi, R.H.Loschi, L.M.R. Baccarini, W.M. Caminhas, Incipient fault detection in inductionmachine stator-winding using a fuzzy-Bayesian change point detection approach,Applied Soft Computing Journal (2008), doi:10.1016/j.asoc.2009.11.008
This is a PDF file of an unedited manuscript that has been accepted for publication.As a service to our customers we are providing this early version of the manuscript.The manuscript will undergo copyediting, typesetting, and review of the resulting proofbefore it is published in its final form. Please note that during the production processerrors may be discovered which could affect the content, and all legal disclaimers thatapply to the journal pertain.
Page 1 of 44
Accep
ted
Man
uscr
ipt
Incipient fault detection in induction machine
stator-winding using a fuzzy-Bayesian change
point detection approach
Marcos F. S. V. D’Angelo a, Reinaldo M. Palhares b,
Ricardo H. C. Takahashi c,∗, Rosangela H. Loschi d,
Lane M. R. Baccarini e, Walmir M. Caminhas f
aDepartment of Computer Science - UNIMONTES,
Av. Rui Braga, sn, Vila Mauriceia, 39401-089, Montes Claros - MG - Brasil
bDepartment of Electronics Engineering, Universidade Federal de Minas Gerais,
Av. Antonio Carlos 6627 - 31270-901, Belo Horizonte - MG - Brazil
cDepartment of Mathematics, Universidade Federal de Minas Gerais
dDepartment of Statistics, Universidade Federal de Minas Gerais
eDepartment of Electrical Engineering, Universidade Federal de Sao Joao del-Rei,
Praca Frei Orlando, 170 - Centro - 36307-352, Sao Joao del-Rei - MG - Brazil
fDepartment of Electrical Engineering, Universidade Federal de Minas Gerais
Abstract
In this paper the incipient fault detection problem in induction machine stator-
winding is considered. The problem is solved using a new technique of change point
detection in time series, based on a two-step formulation. The first step consists of a
fuzzy clustering to transform the initial data, with arbitrary distribution, into a new
one that can be approximated by a beta distribution. The fuzzy cluster centers are
Preprint submitted to Applied Soft Computing 27 May 2009
* Manuscript
Page 2 of 44
Accep
ted
Man
uscr
ipt
determined by using a Kohonen neural network. The second step consists in using
the Metropolis-Hastings algorithm for performing the change point detection in the
transformed time series generated by the first step with that known distribution.
The incipient faults are detected as long as they characterize change points in such
transformed time series. The main contribution of the proposed approach is the
enhanced resilience of the new failure detection procedure against false alarms,
combined with a good sensitivity that allows the detection of rather small fault
signals. Simulation and practical results are presented to illustrate the proposed
methodology.
Key words: Incipient fault detection, induction machine stator-winding, fuzzy
clusters, Bayesian analysis, Metropolis-Hastings algorithm
1 Introduction
Induction motors are the most important electric machinery for different in-
dustrial applications. Faults in the stator windings of three-phase induction
motor represent a significant part of the failures that arise during the motor
lifetime. When these motors are fed through an inverter, the situation tends to
become even worse due to the voltage stresses imposed by the fast switching
of the inverter [1]. From a number of surveys, it can be realized that, for the
induction motors, stator winding failures account for approximately 30% of
all failures [2] [3].
The stator winding of induction machine is subject to stress induced by a
variety of factors, which include thermal overload, mechanical vibrations and
∗ Corresponding author. email:[email protected].
2
Page 3 of 44
Accep
ted
Man
uscr
ipt
voltage spikes. Deterioration of winding insulation usually begins as an inter-
turn short circuit in one of the stator coils. The increased heating due to this
short circuit will eventually cause turn to turn and turn to ground faults which
finally lead the stator to break down [4].
Although there is no experimental data that indicate the time delay between
inter-turn and ground insulation failure, it is believed that the transition be-
tween the two states is not instantaneous. Therefore, early detection of inter
turn short circuit during motor operation can be of great significance as it
would eliminate subsequent damage to adjacent coils and the stator core, re-
ducing repairing cost and motor outage time [5] [6].
However, early stages of deterioration are difficult to detect. In general, most
of the previous references present approaches for dealing with abrupt faults
in the stator winding, which are easier to be detected than incipient faults. In
spite of these difficulties, a great deal of progress has been made on induction
machine stator-winding incipient fault detection. Methods that use voltage
and current measurements offer several advantages over test procedures that
require machine to be taken off line or techniques that require special sensors
to be mounted on the motor [7]. Other methods, in the context of abrupt
fault detection related to the stator-winding, can be found in [8], [9], [10], [11],
[12]. Other types of faults in induction machines, such as dynamic eccentricity,
unbalanced rotors, bearing defects and broken rotor bars have been tackled
via other approaches of fault detection that are specific for each case (see [13]
for details and further references).
In this paper, a new two-step formulation for incipient fault detection in the
stator windings of induction machines is proposed. The proposed methodology
3
Page 4 of 44
Accep
ted
Man
uscr
ipt
deals with the fault detection problem as a change point detection problem
over the time series of the rms (root mean square) values of stator currents.
The change point detection algorithm is based on a fuzzy set technique and
a Markov Chain Monte Carlo (MCMC ) method. The proposed method, dif-
ferently from former techniques, does not require any prior knowledge about
statistical properties of the time series before the application of the MCMC
procedure. This is made possible by the first step, in which a fuzzy set tech-
nique is applied in order to cluster and to transform the initial time series
(about which there is no a priori knowledge of its distribution) into a time
series whose probability distribution can be approximated by a beta distribu-
tion. Specifically in the first step, a Kohonen network is used to find centers of
the clusters, and in the sequel the fuzzy membership degree is computed for
each point of the initial time series, generating a time series with beta distribu-
tion. This new time series, generated in the first step, allows to systematically
apply the same strategy to detect the change point via a MCMC method with
a fixed reference distribution: the beta distribution. The Metropolis-Hastings
algorithm [14] is used to perform the change point detection. The main idea in
this paper is to apply the change point detection strategy in a data sequence
that carries information of relevant physical variables of the dynamic system.
A change point detection gives support to the hypothesis of fault occurrence.
The research on the theme of change point detection in time series has been
performed in the context of several applications, such as financial series [15],
ecological series [16], hydrometeorological time series [17], etc. The main tech-
niques presented in the literature are statistical tests and Bayesian analysis. In
the change point detection problem the standard statistical test is the CUSUM
(Cumulative Sum). The CUSUM test proposed by [18] is widely used in the
4
Page 5 of 44
Accep
ted
Man
uscr
ipt
change point detection, and applications of this method can be seen in [19],
[20], and [21], as well as its modifications and extensions. However other type
of statistical approaches can be considered as the two step presented in [22]
which is based on learning the statistical properties of the process. In the con-
text of Bayesian analysis different MCMC methods may be used as, for exam-
ple, the Metropolis-Hastings, Gibbs sampling (see [14] and references therein),
and reversible jump MCMC (see [23] and references therein). In the Bayesian
analysis context, the product partition model (PPM) proposed by [24] may be
used to model uncertainties that exist in a sequence of random quantities. The
PPM has also been applied to the identification of multiple change points in
the mean of data modeled by Gaussian distribution, as presented in [25] and
[26]. In [27] the PPM has been extended to identify multiple change points
both in the mean and variance of Gaussian-distribution data. However, all
those previous approaches necessarily demand some type of prior knowledge
about the time series, namely the type of distribution that models the data
set. An important contribution of the approach proposed in this paper is the
possibility of dealing with data with unknown probability distributions.
The main contribution of the proposed approach, however, is related to the
enhanced resilience of the new failure detection procedure against false alarms,
combined with a good sensitivity that allows the detection of rather small fault
signals. This property comes from the adopted PPM model, which assumes
explicitly that just one change can occur within the time window under anal-
ysis, performing a search for the most provable change point, and calculating
the probability of such point being effectively a change point. The behavior of
the resulting failure detection system contrasts with the outcome of other ap-
proaches, which search for any change point, assuming implicitly that several
5
Page 6 of 44
Accep
ted
Man
uscr
ipt
change points can occur: such approaches lead either to too sensitive systems
(which cause false alarms) or to too insensitive systems (which will not detect
several faults). Simulation and experimental results illustrate such comparison
of the proposed method with other approaches.
The paper is organized as follows. Section II presents and analyzes the in-
duction machine simulation considering the case of incipient fault on stator-
winding. Section III describes the methodology used for change point detec-
tion. Section IV shows the simulations and experimental results for on–line
incipient fault detection in induction machine stator-winding. Finally, section
V presents the concluding remarks.
2 Induction machine modeling and simulation with turn-to-turn
short-circuit in stator winding
Many studies have shown that a large proportion of induction machine faults
are related to the stator-winding [8], [9], [10], [11]. The induction machine
stator-winding is subject to stress due to many factors, which include thermal
overload, mechanical vibration and peak voltage caused by a speed controller.
The deterioration of insulation usually begins as a short-circuit fault of the
stator-winding. This section describes the model that is employed here for
the simulation of inter-turn short-circuits in the stator windings of induction
machines.
This work employs a generic model for the machine [12], valid for any dq
(direct and quadrature) axis speed obtained by the Park’s transformation [28].
Representing the currents, voltages and electromagnetic flows by i, v and λ,
6
Page 7 of 44
Accep
ted
Man
uscr
ipt
the resistance, leakage and mutual inductance by r, Ll and Lm, the phases a,
b and c by indexes a, b and c, the windings of the stator and rotor by indexes
s and r, the stator and rotor voltages equations become:
[vs] = [rs][is] +d[λs]
dt(1)
[vr] = [rr][ir] +d[λr]
dt(2)
where
[vs] = [ vas1 vas2 vbs vcs]T
[vr] = [ var vbr vcr]T
[is] = [ ias ias − if ibs ics]T
[ir] = [ iar ibr icr]T
[λs] = [ λas1 λas2 λbs λcs]T
[λr] = [ λar λbr λcr]T
In the above, the index as2 represents the shorted turns and if is the current
in the short-circuit. Figure 1 represents the schematic diagram of a motor with
an inter-turn short-circuit.
In the model proposed in reference [12], the stator windings voltages are given
7
Page 8 of 44
Accep
ted
Man
uscr
ipt
Fig. 1. Representation of stator windings of a motor with inter-turn short circuit.
by:
Vds +2
3μrsifcosθ = rsids +
dλds
dt+ ωλqs (3)
Vqs +2
3μrsifsinθ = rsiqs +
dλqs
dt+ ωλds (4)
V0s +1
3μrsif = rsi0s +
dλ0s
dt(5)
The rotor circuit equations are the same as for traditional symmetrical model.
The stator and the rotor electromagnetic flows of stator in dq axis, are given
by:
λds = Lsids + Lmidr − 2
3μLsifcosθ (6)
8
Page 9 of 44
Accep
ted
Man
uscr
ipt
λqs = Lsiqs + Lmiqr − 2
3μLsifsinθ (7)
λ0s = Llsi0s +μ
3Llsifsinθ (8)
λdr = Lridr + Lmids − 2
3μLmifcosθ (9)
λqr = Lriqr + Lmiqs − 2
3μLmifsinθ (10)
The voltage and the induced electromagnetic flow in the short-circuit turns
are given by:
vas2 = μrs(idscosθ + iqssinθ − if ) +dλas2
dt(11)
λas2 = μLls(iqssinθ + idscosθ − if ) + μLm(iqssinθ +
+idscosθ + iqrsinθ + idrcosθ − 2
3μif) (12)
The electromagnetic torque is given by:
T =3
2
p
2Lm(iqsidr − idsiqr) − p
2μLmif iqr (13)
The induction machine stator current simulation results for a fault in which
5% of turns in the phase a become in short-circuit after the time 0.72s are
shown in the Figures 2–4. The root mean square (rms) current values are
illustrated in Figures 5–7. Notice that when a short-circuit occurs in phase
9
Page 10 of 44
Accep
ted
Man
uscr
ipt
a, the increase in the current of phase a is greater than that of phases b and
c. The fault detection methodology proposed here is based on finding such
events of non-balanced changes in the current rms values.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2−30
−20
−10
0
10
20
30
40
Fault
time (s)
i as
Fig. 2. Current of phase a.
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2−30
−20
−10
0
10
20
30
40
Fault
time (s)
i bs
Fig. 3. Current of phase b.
10
Page 11 of 44
Accep
ted
Man
uscr
ipt
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2−40
−30
−20
−10
0
10
20
30
Fault
time (s)
i cs
Fig. 4. Current of phase c.
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 24
4.5
5
5.5
6
6.5
time (s)
i as
Fig. 5. rms current of phase a.
3 Change point detection methodology
In this section, the proposed two-step formulation to the change point detec-
tion problem is detailed. Consider a time series signal in which a change point
is to be detected. The first step consists in transforming the given time series
11
Page 12 of 44
Accep
ted
Man
uscr
ipt
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 24
4.5
5
5.5
6
6.5
time (s)
i bs
Fig. 6. rms current of phase b.
0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 24
4.5
5
5.5
6
6.5
time (s)
i cs
Fig. 7. rms current of phase c.
into another one with beta distribution using a fuzzy set technique [29]. In
order to illustrate how this is done, the following time series is used:
12
Page 13 of 44
Accep
ted
Man
uscr
ipt
y(t) =
⎧⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎩
p1 + 0.1 ∗ ε(t) − 0.1 ∗ ε(t − 1), if t � m,
p2 + 0.1 ∗ ε(t) − 0.1 ∗ ε(t − 1), if t > m
(14)
where p1 is the first operation point (the mean value before the change point),
p2 is the second operation point (the mean value after the change point), ε(t)
is a noise signal with π(·) distribution and m is the change point. The figure
8 shows the time series y(t) with fixed p1 = 1, and p2 = 2, ε(t) ∼ U(0, 1)
(uniform distribution in the interval [0, 1]), m = 30 and 60 samples.
0 10 20 30 40 50 600.8
1
1.2
1.4
1.6
1.8
2
2.2
Samples
y(t
)
Fig. 8. Time series with fixed p1 = 1, and p2 = 2, ε(t) ∼ U(0, 1), m = 30 and 60
samples.
Definition (Fuzzy Cluster Centers) Let y(t) be a time series, and consider
a positive integer k. Define the set
C = {Ci | min{y(t)} ≤ Ci ≤ max{y(t)}, i = 1, 2, . . . , k}
13
Page 14 of 44
Accep
ted
Man
uscr
ipt
such that it solves the minimization problem:
mink∑
i=1
∑y(t)∈Ci
‖ y(t) − Ci ‖2 . (15)
The set C = {Ci, i = 1, 2, . . . , k} which minimizes (15) is called the cluster
center set for the time series y(t). The fuzzy membership degree (or fuzzy
membership function value) of the fuzzy relation y(t) ∈ Ci (which means y(t)
belongs to each cluster Ci) is given by:
μi(t) �⎡⎣ k∑
j=1
‖y(t) − Ci‖2
‖y(t) − Cj‖2
⎤⎦−1
(16)
�
Notice that, given a set C of cluster centers, it is an easy task to measure the
distance of each point in the time series y(t) to each center Ci. The problem
of finding the centers can be solved, for instance, via K-means [30], C-means
[31], and Kohonen network [32]. In this paper, the Kohonen network technique
is used.
The proposed fuzzy clustering to transform a given time series into a new one
is described below:
(1) Input the time series y(t);
(2) Find Ci, i = 1, 2, the elements of the cluster center set for y(t) using the
Kohonen network (as illustrated in figure 9 considering, for example, the
time series in (14)).
(3) Compute the fuzzy membership degree given in (16), for each sample
of the time series, y(t), with respect to each center Ci (as illustrated in
14
Page 15 of 44
Accep
ted
Man
uscr
ipt
figure 10 considering, for example, the time series in (14)).
0 10 20 30 40 50 600.8
1
1.2
1.4
1.6
1.8
2
2.2
Samples
y(t
)
Fig. 9. Time serie centers: ‘×’ denotes both the min{y(t)} and max{y(t)} and ‘◦’denotes the centers found using the Kohonen network.
0 10 20 30 40 50 60
0
0.2
0.4
0.6
0.8
1
Samples
μi(
t)
Fig. 10. Membership function, μ1(t)(· · · ) and μ2(t)(− − −).
Notice that, since the idea is to find just one change point, only two centers
are to be found. The function μ1(t) defines a new time series, in which the
change point is to be detected.
15
Page 16 of 44
Accep
ted
Man
uscr
ipt
Further, it is clear that the distributions of μ1(t), according to (16), are con-
fined in the interval [0, 1]:
• if y(t) −→ C1 then μ1(t) −→ 1−
• if y(t) −→ C2 then μ1(t) −→ 0+
• if C1 −→ C2 then μ1(t) −→ 12
• if y(t) ⊂ [C1, C2] then μ1(t) ⊂ [0, 1]
and using the Kullback-Leibler divergence [33] one may conclude that the
distributions of μ1(t) shape a family of beta distribution with different input
parameters: for μ1(t), t ≤ m, one obtains a beta(a, b) distribution, or beta(c, d)
distribution, if t > m. For example, in the case in which there is a change point
in the time series, the parameter a is greater than the parameter b in the beta
distribution of μ1(t) if t ≤ m and the parameter c is smaller than the parameter
d in the beta distribution of μ1(t) if t > m. Figure 11 presents an illustration
of the distributions of μ1(t) for the time series (14). This empirical test has
been performed for several time series with different probability distributions,
always leading to the same family of beta distributions after the clustering
technique.
Since the clustering technique maps the original time series, with arbitrary
probability distribution, into a new time series μ1(t) with a beta probabil-
ity distribution function, this fixed statistical model can be assumed in the
Bayesian formulation to detect the change point in the transformed time se-
ries (second step). In this paper, the Metropolis-Hastings algorithm is used,
since it is a powerful and simple strategy. The goal of the Metropolis-Hastings
algorithm [14] is to construct a Markov chain that has a specified equilibrium
distribution π.
16
Page 17 of 44
Accep
ted
Man
uscr
ipt
0 0.2 0.4 0.6 0.8 10
2
4
6
8
10
12
14
0 0.2 0.4 0.6 0.8 10
5
10
15
a
b
Fig. 11. Distributions of: (a) μ1(t), t ≤ m and; (b) μ1(t), t > m.
Define a Markov chain as follows. If Xi = xi, then draw a candidate value Y
from a distribution with density fY |X(y) = q(xi, y). The q function is known as
the transition kernel of the chain. It is a function of two variables, the current
state of the chain xi and the candidate value y. For each xi, the function
q(xi, y) is a density which is a function of y.
The candidate value Y is then accepted or rejected. The probability of accep-
tance is
α(x, y) = min
(1,
π(y)
π(xi)
q(y, xi)
q(xi, y)
)(17)
If the candidate value is accepted, then set Xi+1 = Y , otherwise set Xi+1 = Xi.
Thus, if the candidate value is rejected, the Markov chain has a repeat in the
sequence. It is possible to show that under general conditions the sequence
X0, X1, X2, ... is a Markov chain with equilibrium distribution π.
In practical terms, the Metropolis-Hastings algorithm can be specified by the
following steps:
17
Page 18 of 44
Accep
ted
Man
uscr
ipt
Metropolis-Hastings Algorithm
(1) Choose a starting value x0, the number of iterations, R, and set the
iteration counter r = 0;
(2) Generate a candidate value y using the reference distribution given by
q(xr, y);
(3) Calculate the acceptance probability in (17) and generate u ∼ U(0, 1);
(4) Compute the new value for the current state:
xt+1 =
⎧⎪⎪⎪⎪⎪⎪⎨⎪⎪⎪⎪⎪⎪⎩
y, if α(x, y) ≥ u,
xt, otherwise
(5) If r < R, return to step 2. Otherwise stop.
Notice that, as discussed previously, the clustering technique generates a trans-
formed time series with the following distribution:
y(t) ∼ beta(a, b), for t = 1, ..., m
y(t) ∼ beta(c, d), for t = m + 1, ..., n
The parameters to be estimated for the Metropolis-Hastings algorithm are a,
b, c, d and the change point m. In this type of algorithm, the choice of non-
informative priors is performed usually from “flat” distributions, for example:
a ∼ gamma(0.1, 0.1)
b ∼ gamma(0.1, 0.1)
18
Page 19 of 44
Accep
ted
Man
uscr
ipt
c ∼ gamma(0.1, 0.1)
d ∼ gamma(0.1, 0.1)
m ∼ U{1, 2, ..., n}, with p(m) =1
n
These distributions, with parameters 0.1, have been chosen for the purpose of
spreading the whole parametric space.
The reference distribution, used in Step (2) of the Metropolis-Hastings algo-
rithm to generate the candidate value to the change point m, is computed
as:
q(m | a, b, c, d)∝ q(m, a, b, c, d) (18)
∼m∏
i=1
G(a + b)
G(a)G(b)ya−1
i (1 − yi)b−1
n∏i=m+1
G(c + d)
G(c)G(d)yc−1
i (1 − yi)d−1
where a, b, c and d are generated by priors distribution and G(k) ∼ gamma(k, 1).
The final analysis is performed as: the change point, m, is obtained by checking
where the maximum of q(m | a, b, c, d) occurs, with the exception of the border
points of the distribution (if the maximum occurs on such points, then there
is no change point). Figure 12 shows the result when applying the proposed
methodology for p1 = 1, p2 = 2, e(t) ∼ U(0, 1), and m = 30. The function q
can be interpreted as a probability of change in the time series at the instant
m.
This methodology is applied, in the next section, in the problem of incipient
fault detection problem in induction machine stator winding – which is stated
as a change point detection problem.
19
Page 20 of 44
Accep
ted
Man
uscr
ipt
0 10 20 30 40 50 600
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Samples
q(m
|a,b
,c,d
)
Fig. 12. Result of the methodology proposed for p1 = 1, p2 = 2, e(t) ∼ U(0, 1), and
m = 30.
4 Implementation and results of the proposed methodology
In this section, the implementation details are described, and both simulation
results and experimental results are presented. A block diagram of the system
can be seen in Figure 13. The system monitors the instantaneous values of
the motor currents ias, ibs, ics and the rotor speed ω. Firstly, the analogical
measurements are converted in digital data through an A/D converter. Then,
the root mean square (rms) value of each phase current is calculated over a
period of time.
Let i denote the rms value associated to the current whose instantaneous
value is i. In the case of the induction machine with delta connection, the
fault detection is performed by the following rule set, and considering the
Figure 1:
IF m(ias) ∼= m(ibs) > m(ics) THEN fault in phase b;
IF m(ias) ∼= m(ics) > m(ibs) THEN fault in phase a;
20
Page 21 of 44
Accep
ted
Man
uscr
ipt
Fig. 13. Block diagram of the proposed methodology
IF m(ibs) ∼= m(ics) > m(ias) THEN fault in phase c;
IF m(ibs) ∼= m(ics) ∼= m(ias) THEN fault free.
In the case of the induction machine with star connection, the rule set for
fault detection should be adapted as follows:
IF m(ias) > m(ibs) ∼= m(ics) THEN fault in phase a;
IF m(ibs) > m(ias) ∼= m(ics) THEN fault in phase b;
IF m(ics) > m(ias) ∼= m(ibs) THEN fault in phase c;
IF m(ibs) ∼= m(ics) ∼= m(ias) THEN fault free.
where m(ν) indicates the probability of change in the time series of variable ν.
This probability of change is given by the change point detection algorithm, as
21
Page 22 of 44
Accep
ted
Man
uscr
ipt
described in the former section. It should be noticed that the rms values of the
stator currents in an induction motor change their values several times within
a normal operation cycle. Under normal conditions, these changes should be
balanced, i.e., similar changes should occur in the three phases. The idea
behind these rule sets is to detect changes in the rms values of the currents
that are not balanced between the phases: these changes are related to faults.
4.1 Simulation Results
The simulation results of incipient fault detection in the stator winding of
an induction machine with star connection are shown in Figures 14–19, con-
sidering the rule set for star connection and 0.1% of turns in short-circuit in
phase a. These results have been obtained by simulation of the induction ma-
chine using the model described in Section 2 and the change point detection
methodology presented in Section 3.
As can be seen from Figures 15, 17 and 19, the fault detection has been per-
formed in the correct time and in the correct phase, with the system indicating
that the probability of change in the current of phase a is about 60%, while
the probability of change in the currents of phases b and c are below 10%,
with a clear unbalancing of the changes between the phases. The system has
also correctly indicated that the fault has occurred in phase a. For the case of
fault-free and noise-free simulation, the probability of change in each current
phase results null when using the proposed approach.
22
Page 23 of 44
Accep
ted
Man
uscr
ipt
0 50 100 150 200 2504.16
4.165
4.17
4.175
4.18
4.185
4.19
4.195
4.2
4.205
4.21Ia
Fig. 14. rms current of phase a: ias with 0.1% of turns in short-circuit in phase a
0 50 100 150 200 2500
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Fig. 15. Probability of indication of change point detection in current ias
23
Page 24 of 44
Accep
ted
Man
uscr
ipt
0 50 100 150 200 250
4.086
4.088
4.09
4.092
4.094
4.096
4.098
4.1
4.102Ib
Fig. 16. rms current of phase b: ibs with 0.1% of turns in short-circuit in phase a
0 50 100 150 200 2500
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Fig. 17. Probability of indication of change detection in current ibs
24
Page 25 of 44
Accep
ted
Man
uscr
ipt
0 50 100 150 200 250
4.19
4.195
4.2
4.205
4.21
4.215Ic
Fig. 18. rms current of phase c: ics with 0.1% of turns in short-circuit in phase a
0 50 100 150 200 2500
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Fig. 19. Probability of indication of change detection in current ics
25
Page 26 of 44
Accep
ted
Man
uscr
ipt
4.2 Practical Results
Figure 20 illustrates the experimental setup. It consists of a 3HP , 220/380V ,
60Hz, 4 poles, squirrel-cage induction machine.
Phase windings are composed by two phase groups with three coils in each
one. Each coil is composed of 33 turns. Figure 21 shows details of the connec-
tion diagram of the stator windings. The induction motor is specially wound
with tapping that allows a turn-to-turn fault insertion in one of the phases
(Figure 22).
Fig. 20. Experimental test bed
A mechanical load is provided by a separate dc generator feeding a variable
resistor. In order to allow tests to be performed at different load levels, the dc
excitation current and load resistor are both controllable. The data acquisition
system consists of:
• three hall effect current sensors (LEM, LTA50P);
26
Page 27 of 44
Accep
ted
Man
uscr
ipt
Fig. 21. Connection diagram of stator winding
Fig. 22. Photograph showing the stator terminals
• three hall effect voltage sensors (LEM, LV 100-300);
• analog input board (National Instruments PCI 6013).
In all experimental tests, the stator windings of the motor have delta con-
nection. When short-circuits were introduced, a shorting resistor was used in
order to limit the value of the short-circuit current, thus protecting the motor
windings from complete failure.
The Figures 23 – 25 show the instantaneous currents of phases a (ias), b (ibs)
27
Page 28 of 44
Accep
ted
Man
uscr
ipt
and c (ics) with ∼ 1.5% of turns in short-circuit in phase b. The short-circuit
has been established at the time t ∼= 2s. Figures 26 – 28 show the rms value
currents of phases a (ias), b (ibs) and c (ics), for the same experiment.
The results of incipient fault detection in induction machine stator winding are
shown in Figures 29–31, considering the rule set for delta connection and 1.5%
of turns in short-circuit in phase b. To illustrate a change point detection using
the proposed approach, the fault-free and incipient fault cases are considered.
For the fault-free case (Figures 29a, 30a and 31a) Figures 29c, 30c and 31c
show that the probability of change in each current phase is null when using
the proposed approach. For the incipient fault case (Figures 29b, 30b and
31b), the detection of the fault has been performed in the correct time instant
and in the correct phase, as shown in Figures 29d, 30d and 31d. In this case,
the system indicates that the probability of change in the currents of phases
a and b is approximately 10%, while the probability of change in the current
of phase c is null, with a clear unbalancing of the changes between the phases.
These informations lead the system to indicate (correctly) that the fault has
occurred in phase b.
For comparison purposes, two other methods have been considered:
• The classical CUSUM statistical method [19].
• The direct use of a fuzzy clustering technique as the one presented in Sec-
tion 3.
To illustrate how a change point detection using the classical CUSUM works,
the fault-free and incipient fault cases are considered. For the fault-free case,
Figures 32 c, 33 c and 34 c show that the CUSUM generates several “false
alarm” indications. For the case of incipient fault with 1.5% of turns in
28
Page 29 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 2 2.5 3 3.5−15
−10
−5
0
5
10
15
i as
t(s)
Fig. 23. Instantaneous current of phase a: ias with ∼ 1.5% of turns in short-circuit
in phase b
0 0.5 1 1.5 2 2.5 3 3.5−15
−10
−5
0
5
10
15
i bs
t(s)
Fig. 24. Instantaneous current of phase b: ibs with ∼ 1.5% of turns in short-circuit
in phase b
short-circuit in phase b, the CUSUM method detects nearly the same change
point that has been indicated by the proposed methodology, as shown in Fig-
ures 32 d, 33 d and 34 d.
29
Page 30 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 2 2.5 3 3.5−15
−10
−5
0
5
10
15
i cs
t(s)
Fig. 25. Instantaneous current of phase c: ics with ∼ 1.5% of turns in short-circuit
in phase b
0 0.5 1 1.5 2 2.5 3 3.58.2
8.25
8.3
8.35
8.4
8.45
8.5
8.55
8.6
8.65
i as
t(s)
Fig. 26. rms current of phase a: ias with ∼ 1.5% of turns in short-circuit in phase b
A possible approach for detecting faults could be the direct examination of
the fuzzy set membership transformed variable from step 1 of the proposed
algorithm to indicate the position of the change-point. This approach has been
30
Page 31 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 2 2.5 3 3.57.7
7.8
7.9
8
8.1
8.2
8.3
8.4
i bs
t(s)
Fig. 27. rms current of phase b: ibs with ∼ 1.5% of turns in short-circuit in phase b
0 0.5 1 1.5 2 2.5 3 3.57.7
7.75
7.8
7.85
7.9
7.95
8
i cs
t(s)
Fig. 28. rms current of phase c: ics with ∼ 1.5% of turns in short-circuit in phase b
evaluated, with the results shown in Figures 35, 36 and 37, where the fault-free
case is presented. It should be noticed that the membership functions indicate
several false alarms, as in the case of the CUSUM.
31
Page 32 of 44
Accep
ted
Man
uscr
ipt0 0.5 1 1.5 2
8.2
8.25
8.3
8.35
8.4
8.45
8.5
8.55a
t(s)
i as
0 0.5 1 1.5 20
0.2
0.4
0.6
0.8c
t(s)
q(m
|a,b
,c,d
)
0 1 2 3 48.2
8.3
8.4
8.5
8.6
8.7b
t(s)
i as
0 1 2 3 40
0.02
0.04
0.06
0.08
0.1
0.12d
t(s)
q(m
|a,b
,c,d
)
Fig. 29. a - ias in free-fault case; b - ias with ∼ 1.5% of turns in short-circuit in
phase b; c - Change detection in current ias for fault-free case; d - Change detection
in current ias with ∼ 1.5% of turns in short-circuit in phase b
It should be also noticed that the same membership function value signal,
combined with the Metropolis-Hastings algorithm in the proposed new algo-
rithm, prevents the detection of false alarms. This enhanced behavior is due
to the explicit quantification of the “probability of change” that is performed
by the Metropolis-Hastings algorithm, with the explicit assumption that just
one change can occur. This endows the proposed algorithm with the ability
to distinguish effective changes in a signal from mere stochastic fluctuations
in the raw fuzzy membership signal.
Other methods could be considered for the same fault detection problem, such
32
Page 33 of 44
Accep
ted
Man
uscr
ipt0 0.5 1 1.5 2
7.7
7.8
7.9
8
8.1
8.2a
t(s)
i bs
0 0.5 1 1.5 20
0.1
0.2
0.3
0.4
0.5c
t(s)
q(m
|a,b
,c,d
)
0 1 2 3 47.7
7.8
7.9
8
8.1
8.2
8.3
8.4b
t(s)
i bs
0 1 2 3 40
0.02
0.04
0.06
0.08
0.1d
t(s)
q(m
|a,b
,c,d
)
Fig. 30. a - ibs in fault-free case; b - ibs with ∼ 1.5% of turns in short-circuit in
phase b; c - Change detection in current ibs for fault-free case; d - Change detection
in current ibs with ∼ 1.5% of turns in short-circuit in phase b
as recursive estimation and identification schemes. However, these methods
require the specification of a model for the residual generation process as well
as the definition of decision thresholds. The proposed methodology does not
depend on such kind of information.
33
Page 34 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 27.7
7.75
7.8
7.85
7.9
7.95
8a
t(s)
i cs
0 0.5 1 1.5 20
0.05
0.1
0.15
0.2
0.25
0.3
0.35c
t(s)
q(m
|a,b
,c,d
)
0 1 2 3 47.7
7.75
7.8
7.85
7.9
7.95
8b
t(s)i cs
0 1 2 3 40
0.1
0.2
0.3
0.4
0.5d
t(s)
q(m
|a,b
,c,d
)
Fig. 31. a - ics in fault-free case; b - ics with ∼ 1.5% of turns in short-circuit in
phase b; c - Change detection in current ics for fault-free case; d - Change detection
in current ics with ∼ 1.5% of turns in short-circuit in phase b
34
Page 35 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 28.2
8.25
8.3
8.35
8.4
8.45
8.5
8.55a
t(s)
i as
0 0.5 1 1.5 20
0.02
0.04
0.06
0.08
0.1
0.12
0.14c
t(s)
CU
SU
M
0 1 2 3 48.2
8.3
8.4
8.5
8.6
8.7b
t(s)i a
s
0 1 2 3 40
0.05
0.1
0.15
0.2d
t(s)
CU
SU
M
Fig. 32. a - ias in fault-free case; b - ias with ∼ 1.5% of turns in short-circuit in
phase b; c - Change detection in current ias for fault-free case; d - Change detection
in current ias with ∼ 1.5% of turns in short-circuit in phase b
35
Page 36 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 27.7
7.8
7.9
8
8.1
8.2a
t(s)
i bs
0 0.5 1 1.5 20
0.05
0.1
0.15
0.2
0.25c
t(s)
CU
SU
M
0 1 2 3 47.7
7.8
7.9
8
8.1
8.2
8.3
8.4b
t(s)i b
s
0 1 2 3 40
0.05
0.1
0.15
0.2d
t(s)
CU
SU
M
Fig. 33. a - ibs in fault-free case; b - ibs with ∼ 1.5% of turns in short-circuit in
phase b; c - Change detection in current ibs for fault-free case; d - Change detection
in current ibs with ∼ 1.5% of turns in short-circuit in phase b
36
Page 37 of 44
Accep
ted
Man
uscr
ipt
0 0.5 1 1.5 27.7
7.75
7.8
7.85
7.9
7.95
8a
t(s)
i cs
0 0.5 1 1.5 20
0.05
0.1
0.15
0.2
0.25c
t(s)
CU
SU
M
0 1 2 3 47.7
7.75
7.8
7.85
7.9
7.95
8b
t(s)i cs
0 1 2 3 40
0.02
0.04
0.06
0.08
0.1
0.12d
t(s)
CU
SU
M
Fig. 34. a - ics in fault-free case; b - ics with ∼ 1.5% of turns in short-circuit in
phase b; c - Change detection in current ics for fault-free case; d - Change detection
in current ics with ∼ 1.5% of turns in short-circuit in phase b
37
Page 38 of 44
Accep
ted
Man
uscr
ipt0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
8.2
8.25
8.3
8.35
8.4
8.45
8.5
8.55a
t(s)
i as
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.80
0.2
0.4
0.6
0.8
1b
t(s)
μ 1t a
nd μ 2t
Fig. 35. a - ias in fault-free case; b - Intersection between μ1t and μ2t indicating
change in current ias for fault-free case
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.87.7
7.8
7.9
8
8.1
8.2a
t(s)
i bs
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.80
0.2
0.4
0.6
0.8
1b
t(s)
μ 1t a
nd μ 2t
Fig. 36. a - ibs in fault-free case; b - Intersection between μ1t and μ2t indicating
change in current ibs for fault-free case
38
Page 39 of 44
Accep
ted
Man
uscr
ipt
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.87.7
7.75
7.8
7.85
7.9
7.95
8a
t(s)
i cs
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.80
0.2
0.4
0.6
0.8
1b
t(s)
μ 1t a
nd μ 2t
Fig. 37. a - ics in fault-free case; b Intersection between μ1t and μ2t indicating change
in current ics for fault-free case
39
Page 40 of 44
Accep
ted
Man
uscr
ipt
5 Conclusions
In this paper a novel fuzzy/Bayesian methodology for change point detection
in time series has been used to treat the on-line fault detection problem in
induction machine stator-winding. The methodology is based on a two-step
formulation: Firstly, a fuzzy clustering generates a transformed time series
with beta distribution. In the second step, a Metropolis-Hastings algorithm
is used to detect the probability of the occurrence of a change point in the
transformed time series. This two-step formulation allows a systematic efficient
way to solve a change point detection problem, which is employed for detecting
incipient faults as change points that occur in some system signal.
The proposed methodology has as advantages, compared to other techniques
for the FDI problem, the fact that no mathematical models of the motor and
no previous knowledge of signal statistical distributions are necessary, and also
an enhanced resilience against false alarms, combined with a good sensitivity
that allows the detection of rather small fault signals.
This methodology has been successfully applied: Simulation and experimental
results have been presented as evidences of the effectiveness of the proposed
methodology, even in the case of faults that cause very low level disturbances.
Acknowledgments
This work has been supported in part by the Brazilian agencies CNPq and
FAPEMIG.
40
Page 41 of 44
Accep
ted
Man
uscr
ipt
References
[1] S. M. A. Cruz, J. Cardoso, Diagnosis of stator interturn short circuits in dtc
induction motor drives, IEEE Transactions on Industry Applications 40 (5)
(2004) 1349–1360.
[2] M. Arkan, D. K. Perovic, P. Unsworth, Online stator fault diagnosis in induction
motors, IEE Proceedings Electronic Power Application 148 (6) (2001) 537–547.
[3] M. R. Iravani, M. Karimi-Ghartemani, Online estimation of steady state
and instantaneous symmetrical components, IEE Proceedings Generation,
Transmission and Distribution 150 (5) (2003) 616–622.
[4] R. Tallam, S. Lee, G. Stone, G. Kliman, J. Yoo, T. Habetler, R. Harley, A survey
of methods for detection of stator faults in induction machines, in: Proceedings
of SDEMPED, Diagnostics for Electric Machines, Power Eletronics and Drives,
2003, pp. 35–46.
[5] W.T.Thomson, M. Fenger, Current signature analysis to detect induction motor
faults, IEEE Industry Applications Magazine 7 (2001) 26–34.
[6] X. Boqiang, L. Heming, S. LilingJ, Apparent impedance angle based detection of
stator winding interturn short circuit fault in induction motors, in: Proceedings
of the Industry Application Conference, 2003, pp. 1118–1125.
[7] J. Sottile, F. Trutt, J. L. Kohler, Experimental investigation of on-line methods
for incipient fault detection in induction motors, in: Proceedings of the Industry
Application Conference, 2000, pp. 2682–2687.
[8] P. V. J. Rodriguez, A. Arkkio, Detection of stator winding fault in induction
motor using fuzzy logic, Applied Soft Computing 8 (2008) 1112–1120.
[9] M. S. Ballal, H. M. Suryawanshi, M. K. Mishra, Detection of incipient faults
41
Page 42 of 44
Accep
ted
Man
uscr
ipt
in induction motors using FIS, ANN and ANFIS techniques, Journal of Power
Electronics 8 (2) (2008) 181–191.
[10] A. M. da Silva, R. J. Povinelli, N. A. O. Demerdash, Induction machine broken
bar and stator short-circuit fault diagnostics based on three-phase stator current
envelopes, IEEE Transactions on Industrial Electronics 55 (3) (2008) 1310–1318.
[11] R. M. Tallam, S. B. Lee, G. C. Stone, G. B. Kliman, J. Yoo, T. G. Habetler,
R. G. Harley, A survey of methods for detection of stator-related faults in
induction machines, IEEE Transactions on Industry Applications 43 (4) (2007)
920–933.
[12] L. M. R. Baccarini, B. R. Menezes, H. N. Guimares, W. M. Caminhas, A method
for early detection of stator winding faults, in: Proceedings of VII International
Conference on Industrial Applications, Recife/Brazil, 2006, pp. 1–6.
[13] S. Rajagopalan, J. A. Restrepo, J. M. Aller, T. G. Habetler, R. G. Harley,
Nonstationary motor fault detection using recent quadratic time-frequency
representations, IEEE Transactions on Industry Applications 44 (3) (2008) 735–
744.
[14] D. Gamerman, Markov chain Monte Carlo: stochastic simulation for Bayesian
inference, Chapman & Hall, 1997.
[15] K. J. Oh, T. H. Roh, M. S. Moon, Developing time-based clustering neural
networks to use change-point detection: Application to financial time series,
Asia-Pacific Journal Of Operational Research 22 (1) (2005) 51–70.
[16] B. Beckage, L. Joseph, P. Belisle, D. B. Wolfson, W. J. Platt, Bayesian change-
point analyses in ecology, New Phytologist 174 (2) (2007) 456–467.
[17] L. Perreault, J. Bernier, B. Bobe, E. Parent, Bayesian change-point analysis in
hydrometeorological time series. part 1. the normal model revisited, Journal of
Hydrology 235 (3-4) (2000) 221–241.
42
Page 43 of 44
Accep
ted
Man
uscr
ipt
[18] D. V. Hinkey, Inference about the change point from cumulative sum test,
Biometria 26 (1971) 279–284.
[19] O. Hadjiliadis, V. Moustakides, Optimal and asymptotically optimal cusum
rules for change point detection in the brownian motion model with multiple
alternatives, Theory of Probability and its Applications 50 (1) (2006) 75–85.
[20] S. Lee, Y. Nishiyama, N. Yoshida, Test for parameter change in diffusion
processes by cusum statistics based on one-step estimators, Annals of the
Institute of Statistical Mathematics 58 (2) (2006) 211–222.
[21] S. Lee, S. Park, K. Maekawa, K. Kawai, Test for parameter change in arima
models, Communications in Statistics: Simulation and Computation 35 (2)
(2006) 429–439.
[22] J. Takeuchi, K. Yamanishi, A unifying framework for detecting outliers and
change points from time series, IEEE Transactions on Knowledge and Data
Engineering 18 (4) (2006) 482–492.
[23] K. Kyungduk, M. Vannucci, Bayesian wavelet-based methods for the detection
of multiple changes of the long memory parameter, IEEE Transactions on Signal
Processing 54 (11) (2006) 4461–4470.
[24] J. A. Hartigan, Partition models, Communication in Statistics-Theory and
Methods 19 (8) (1990) 2745–2756.
[25] D. Barry, J. A. Hartigan, A bayesian analysis for change point problems, Journal
of the American Statistical Association 88 (421) (1993) 309–319.
[26] E. M. Crowley, Product partition models for normal means, Journal of the
American Statistical Association 92 (437) (1997) 192–198.
[27] R. Loschi, F. Cruz, Extension to the product partition model: computing the
probability of a change, Computational Statistics and Data Analysis 48 (2)
(2005) 255–268.
43
Page 44 of 44
Accep
ted
Man
uscr
ipt
[28] P. C. Krause, Analysis of Electric Machinery, McGraw-Hill, 1986.
[29] L. A. Zadeh, Fuzzy sets, Information and Control 8 (3) (1965) 338–353.
[30] L. Kaufman, P. J. Rousseeuw, Finding groups in data: An introduction to
cluster analysis, John Wiley & Sons, 1990.
[31] J. C. Bezdek, Pattern recognition with fuzzy objective function algorithms,
Plenum Press, 1981.
[32] T. Kohonen, Self-organizing maps, Springer Series in Information Sciences,
Springer, 2001.
[33] S. Kullback, R. A. Leibler, On information and sufficiency, Annals of
Mathematical Statistics 22 (1) (1951) 79–86.
44