**
I. NOMENCLATURE
Fs
s
Supply frequency
Per unit slip
Is
iD , iQ
v D , vQ
Vd , I d
P, Q, S
x
xj
,
N
m
2
Classes
Number of samples in the training data set
Center of gravity vector
Variance vector
Matrix representing the N sample patterns
Variance-covariance matrix
Eigenvectors matrix
Eigenvalues diagonal matrix
II. INTRODUCTION
(1)
A.
x j x d
1,
2 , ...
I s = i D 2 + iQ 2
(2)
Let ms be the I s
( ( ) )
1
2
(3)
(4)
(4.1)
2
1 M
I (k )
M k =1 d
(5)
Vd ( f = Fs )
I d ( f = Fs )
(6)
S = P + j Q
x1
TABLE I
PARAMETERS SUMMARY
x2
md
x3
mq
x4
E
x5
Zd
states of the system are represented by classes in a 5dimensional space and each pattern vector x k (k = 1N;
and N = 100) is represented by a point.
As the 5-dimensional representation space is determined, a
procedure called Principal Components Analysis [7] (PCA)
is applied in order to project and visualize the different
classes in a 2 or 3-dimensional subspace.
The principle of the PCA is developed in the following
section.
1)
...
(7.6)
d'
...
X proj = X1proj
X 2proj
Xd'proj = X t V '
(7.7)
with,
m=
N k =1
xk
(7.1)
j =
1 N
x kj m j
N k =1
(7.2)
xk - m
1
xk x kt
N k =1
(7.4)
jI
V = V1 V2
(7.5)
(7.8)
(7.3)
V2 ... Vd'
2)
where j =1 d
Then the normalized vector components are:
x k norm =
V ' = V1
j=1
d
j=1
j
(7.9)
TABLE II
CONTRIBUTION OF THE EIGENVALUES
eigenvalues
percentage
4.4414
89.73
89.73
0.4652
9.4
99.13
0.0243
0.49
99.61
0.0140
0.28
99.89
Cumulated percentage :
0.0052
B.
0.11
-0.42
0.59
-0.07
0.65
-0.16
0.46
-0.28
-0.19
0.65
0.49
-0.45
-0.41
-0.79
-0.033
-0.073
V=
(7.10)
1/ 2
dE ( xu , xk ) =
100
( xu
xk ) ( x u x k )
( xu
xk )
(8.1)
1/ 2
d M ( xu , x k ) =
-1
( xu x k )
(8.2)
-0.42
-0.62
0.56
0.31
-0.15
0.47
-0.13
-0.13
0.20
-0.84
75 %
3
3
25 %
4
100 %
50 %
2
1
5
0 %
Fig. 1b PCA result: the different classes, and their evolutions according to
the level of load, represented in the three-dimensional subspace (healthy
rotor , 3 broken bars o).
1)
Fig. 1a PCA result: the different classes, and their evolutions according to
the level of load, represented in the two-dimensional subspace (healthy rotor
, 3 broken bars o).
According to fig. 1a and 1.b, one can see that the classes
(representing the two states) move with the level of load.
Therefore, the first conclusion is that the different operating
conditions of the induction can be identified.
During the training step, each pattern has been affected to
one of the known classes. A decision rule based on the k
nearest neighbors is developed in order to affect unknown
patterns in the classes.
d E x 2 , m1
> d E x1 , m1
and
TABLE III
DISTANCE MEASUREMENTS RESULT
m1
0.0073
( )
(x ,m )
(x ,m )
(x ,m )
d M x1 , m i
dM
dE
dE
a)
decision
m2
0.0219
First class
only if
0.0183
0.0074
0.0099
0.0104
0.0099
0.0088
Second
class
First class
Error risk
Second
class
, else
and let
, of
l,
l.
( x k c , k = 1Nc) and
are
calculated (9.1)
m2
x2
) (
d x k c ,m l = x k c m l
xk c m l
12
(9.1)
is either the identity matrix or the inverse of the variancecovariance matrix (8.1) (8.2).
Then the distance reject rule is:
x1
m1
k = 1,Nc
- Else xu belongs to l .
Figure 3 is an illustration of these two kinds of reject.
Fig. 2 Classification
measurements.
2)
after
Euclidean
and
Mahalanobis
distances
mc
1,
2,
...
ml
Fig. 3. Decision step leading to the classification (a), the ambiguity reject
(with almost the same number of neighbors in both classes) (b), and the
distance reject (c)
Number of samples
misclassified Before
reject options
4
Number of
samples rejected
in ambiguity
0
Number of
samples rejected
in distance
1
10
VI. CONCLUSION
A decision system to detect broken bars in induction
motors has been presented. With the help of the signature
(pattern vector), the procedure, based on pattern recognition
approach, allows the identification of the different operating
conditions. Different levels of load are studied in order to
follow the classes evolution. In the training step the principal
component analysis has allowed to see that the classes were
separated, which means that the defined pattern vector is a
relevant signature. The k-nearest neighbors rule, including
ambiguity reject and distance reject, is widely applied in
statistical pattern recognition. For our study, it offers
encouraging results, especially when it is based on the
Mahalanobis distance measurement. With this training set,
other decision rules, such as accurate decision boundaries
between classes, can be applied in order to improve the
decision step.
Number of samples
misclassified Before
reject options
5
Number of
samples rejected
in ambiguity
0
Number of
samples rejected
in distance
1
10
10
10
10
10
10
VII. REFERENCES
[1] R. Yacamini, S. C Chang, "Noise and vibration from induction
machines fed from harmonic sources ", IEEE Trans. Energy
Conversion, Vol. 10 no 2, June 1995.
[3]
[4]
[5]
[6]
[7]
VIII. BIOGRAPHIES
Roland Casimir was born in Saint-Rose
(Guadeloupe), France, on September 14,
1975. He obtained in 2000 the Master Degree
in Electrical Engineering from the "Institut
National Polytechnique" of Toulouse. He is
currently engaged in Ph.D. research in the
department of Electrical engineering of the
Ecole Centrale de Lyon. At the laboratory
CEGELY, his research interests are in fault
diagnosis of electrical machines by means of
pattern recognition methods
Emmanuel Boutleux was born on july 1970
in Friville-Escarbotin (Picardie), France.
Within
the
laboratory
HEUDIASYC
(Compigne France) he became PhD in
System Control in 1996. He is currently an
Assistant Professor in Automatica at Ecole
Centrale de Lyon. His research at the
laboratory CEGELY includes diagnosis of
complex systems by means of pattern
recognition methods.
Guy Clerc was born in Libourne, France, on
November 30, 1960. He received the
Enginer's degree and the PhD in Electrical
engineering from the Ecole Centrale de Lyon,
France, in 1984 and 1989, respectively. From
2000, he was Professor of University at
UCBL. He taught "numerical control of
electrical
machines"
and
electrical
engineering. He carried out research on
control and diagnosis of induction machines.
He has written several papers on robust control and fault prediction.
He has written a book with G. Grellet entitled "Actionneurs
lectriques : principes et commandes" (C) Eyrolles.