IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO.

Y 0

L IST OF F IGURES
1 Short circuit windings . . . . . . . . . . . . . . . 2
2 A short circuit model of induction machine . . . 3
3 A global stator faulty model in dq frame . . . . 3
4 Broken rotor bar representation . . . . . . . . . . 4
5 Broken rotor bars model . . . . . . . . . . . . . 5
6 Stator and rotor faulty model of induction motors 5
7 Motor experimental setup . . . . . . . . . . . . . 7
8 Estimation of electrical and faulty parameters at
stator and rotor faulty case . . . . . . . . . . . . 8
9 Estimation of electrical and faulty parameters at
faulty case . . . . . . . . . . . . . . . . . . . . . 8
10 Spectral density (dB) of identification residuals . 8

L IST OF TABLES
I Estimation results of stator and rotor faults . . . 8
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y
Diagnosis by parameter estimation of stator and
rotor faults occurring in induction machines
Smail Bachir, Slim Tnani, Jean Claude Trigeassou and Gérard Champenois
Abstract—In this paper, authors give a new model of squirrel
cage induction motors under stator and rotor faults. Firstly, we
study an original model which takes into account the effects of
inter turn faults resulting in the shorting of one or more circuits
of stator phase winding. We introduce thus additional parameters
to explain the fault in the three stator phases. Then, we propose
a new faulty model dedicated to broken rotor bars detection.
The corresponding diagnosis procedure based on parameter
estimation of the stator and rotor faulty model is proposed.
The estimation technique is performed by taking into account
prior information available on the safe system operating in
nominal conditions. A special three-phases induction machine has
been designed and constructed in order to simulate true faulty
experiments. Experimental test results show good agreement
and demonstrate the possibility of detection and localization of
previous failures.
Index Terms—Induction motors, modeling, short circuit, bro-
ken rotor bars, faults diagnosis, parameter estimation, prior
information.
I. I NTRODUCTION
Condition monitoring of electric motors has attracted in-
creasing attention during the last decade, during which a
considerable number of studies on the detection of stator
and rotor faults have been made [15][19][23][26]. Thus, in
literature, many proposed techniques are based on spectral
analysis of stator currents, stator voltage and electromagnetic
torque and it has been shown that currents monitoring can be
used to estimate stator insulation degradation and to detect
broken rotor bars [1][12][15].
These methods are based on detection of sidebands at
certain frequencies using Fourier’s analysis (FFT software).
For these latter techniques, it is assumed that the current mea-
surements can be modelized like multi-component mixtures
whose magnitudes change when a failure occurs. Moreover,
even in safe conditions, the frequency depends on speed
and power supply frequency. So, the usual techniques, based
on spectral analysis, are not well-adapted and only para-
metric methods tackle faults detection for adjustable speed
motor drive. Recently, continuous identification has been used
to perform the diagnosis procedure [6][19][23][26]. These
techniques study the deviation of parameters to detect and
localize faults. In this paper, we present our results for a new
diagnosis technique of squirrel cage induction motors by off-
line parameter identification using real data.
Because it requires a model suited for fault modeling, a
new stator and rotor faulty model is proposed. Firstly, the
short circuit model proposed by [23] has been extended to
the general case. This very simple model makes it possible
1
to explain the fault in one phase with a simple short cir-
cuit element. On the other hand, it is inappropriate in the
case of simultaneous faults in several stator phases. Thus
in this paper, we propose the generalization of this model
with the help of a short circuit element dedicated to each
phase [4][6]. In presence of defects in several phases, each
short circuit element allows the detection and localization
of inter turn short circuits in the corresponding phase. To
take into account broken rotor bars, a new faulty model is
developed [7]. Electrical parameters and faulty parameters of
this model have been identified by Output-Error technique
[4][19][25]. In practice, identification technique should be
adapted to the use objectives. In our case, diagnosis is realized
by a parameters follow-up. Then, it is necessary to work in
continuous time representation because all parameters have
physical significance (resistance, inductance,..). Besides, pa-
rameters monitoring supposes a preliminary learning to know
exactly the electrical parameter values in the healthy case.
So it is important to introduce this physical knowledge to
perform parameter estimation for diagnosis purpose. Thus,
parameter estimation with prior information offers an elegant
solution [19][26]. It is in this spirit that a new methodology
of identification has been adapted recently to introduce in a
realistic way this prior information.
A special 1.1 kWatt squirrel cage induction motor has been
designed and constructed in order to simulate true inter turn
short circuits at several levels. Different rotors, with broken
rotor bars, are used to simulate a bar breakage occurring during
operation. Experimental results exhibit the good agreement
and confirm the possibility to diagnosis simultaneous stator
and rotor faults.
II. A NEW INDUCTION MOTOR MODEL FOR FAULTS
DETECTION
References [18][19][23] show that it is useless to consider
an unbalanced two axis Park’s model for diagnosis of induc-
tion motors. The deviation of their electrical parameters is
certainly an indication of a new situation in the machine, but
this evolution can be due to heating or an eventual change
in magnetic state of the motor [19]. On the other hand, it is
very difficult to distinguish stator faults from rotor ones. The
use of Fast Fourier’s analysis of identification residuals is an
original method to localize a fault, but estimation of electrical
parameters is unable to obtain the fault level [18].
A good solution is the introduction of an additional model
to explain the faults [7][6]. The parameters of this differential
model allow detection and localization of the faulty windings.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 2

A. Stator faults modeling in induction motor Where

us : stator voltage
In order to take into account the presence of inter turns
is , ir and icc resp. stator, rotor and short circuit currents
short circuit windings in the stator of an induction motor, an
φs , φr and φcc resp. stator, rotor and short circuit flux
original model was proposed in reference [23]. It is composed
of an additional shorted winding in three phases axis. Fig. 1
shows a three phases, 2-poles, induction machine in case of [Rs ] = Rs · Identity(3) : stator resistance
short circuit winding at phase b. This faults induces in stator [Rr ] = Rr · Identity(3) : rotor resistance
a new windings Bcc short circuited and localized according to Rcc = ηcc Rs : short circuit resistance
first phase by the angle θ cc = 23π rad. ⎡ ⎤
L L
Lp + Lf − 2p − 2p
⎢ ⎥
[Ls ] = ⎣ − 2
Lp
Lp + Lf
L
− 2p ⎦,
S ta to r fa u lts (in ter tu rn s sh o rt circu it
Lp L
w in d in g s a t p h a se b ) F irst p o le o f
⎡ −⎤2 − 2p Lp + Lf
p h a se a
L L
Lp − 2p − 2p
a
⎢ L ⎥ 2
[Lr ] = ⎣ − L2p Lp − 2p ⎦, Lcc = ηcc (Lp + Lf )
L L
c − 2p − 2p Lp
⎡ ⎤
cos(θ) cos(θ + 2π 2π
3 ) cos(θ − 3 )
[Msr ] = Lp ⎣ cos(θ − 2π 3 ) cos(θ) cos(θ + 2π3 )
⎦,
2π 2π
θ cc S h o rt circu it cos(θ + 3 ) cos(θ
⎡ ⎤ − 3) cos(θ)
Bcc
w in d in g
cos(θcc )
[Mscc ] = ηcc Lp ⎣ cos(θcc − 2π 3 )
⎦,
2π
cos(θcc + 3 )
S eco n d p o le o f ⎡ ⎤
p h a se a cos(θcc − θ)
[Mrcc] = ηcc Lp ⎣ cos(θcc − θ − 2π3 )
⎦
b
cos(θcc − θ + 2π3 )

Fig. 1. Short circuit windings [Mrs ] = [Msr ]T , [Mccr ] = [Mrcc]T , [Mccs ] = [Mscc ]T

Two parameters are introduced to define the stator faults ncc Number of inter turns short circuit windings
ηcc = =
• The localization parameter θ cc which is a real angle ns Total number of inter turns in healthy phase
between the short circuit inter turn stator winding and the (2)
first stator phase axis (phase a). This parameter allows the θ is rotor angular position. L p and Lf are respectively princi-
localization of the faulty winding and can take only three pal and global leakage inductance referred to the stator.
values 0, 2π 4π
3 or 3 , corresponding respectively to a short 2) Two-phases stator faulty induction model: To minimize
circuit on the stator phases a, b or c. the number of model variables, we use Concordia transforma-
• The detection parameter η cc equal to the ratio between tion which gives αβ values of same amplitude as abc ones.
the number of inter turn short circuit windings and the Thus, we define three to two axis transformation T 23 as:
total number of inter turns in one healthy phase. This
parameter allows to quantify the unbalance and to obtain xαβs = T23 xs : stator variables
(3)
the number of inter turns in short circuit. xαβr = P (θ) T23 xr : rotor variables
1) short circuit model: In the stator faulty case, an ad-
ditional shorted circuit winding B cc appears in stator. This where xαβ is projection of x following α and β axis. Matrix
winding creates a stationary magnetic field H cc oriented transformations are defined as:
according to the faulty winding [4][23]. Thus, we define the  
short circuit current i cc into inter turn short circuit winding at 2 cos(0) cos( 2π ) cos( 4π )
[T23 ] = 3 3 3
origin of a short circuit flux φ cc . Voltage and flux equations sin(0) sin( 2π3 ) sin( 4π
3 )
for faulty model of induction machine with global leakage 
inductance referred to the stator can be written as: cos(θ) cos(θ + π2 )
P (θ) = : rotational matrix
⎧ sin(θ) sin(θ + π2 )
⎪ us = [Rs ] is + dt d
φs
⎪
⎪
⎪
⎪ 0 = [Rr ] ir + dt φr d
⎪
⎨ The short circuit variables are localized on one axis, these
0 = Rcc .icc + dt d
φcc projection on the two Concordia axis α and β is defined as:
(1)
⎪
⎪ φ = [L ] i + [M ] ir + [Mscc ] icc
⎪
⎪ s s s sr
 
⎪
⎪ φ = [Mrs ] is + [Lr ] ir + [Mrcc ] icc cos(θcc ) cos(θcc )
⎩ r iαβcc = · icc , φαβ = · φcc (4)
φcc = [Mccs ] is + [Mccr ] ir + Lcc icc sin(θcc ) cc sin(θcc )
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 3

Thus, (1) becomes: It is much simpler to work in the rotor reference frame
⎧ because we have only two stator variables to transform. There-
⎪ U = Rs iαβs + dtd
φαβ
⎪ αβs
⎪ fore, in state operation, all the variables have their pulsations
⎪
s
⎪
⎪ 0 = Rr iαβr + dt φαβ − ω P ( π2 ) φαβ
d
equals to s ωs (where s is the slip and ω s is stator pulsation).
⎪
⎪ r r
⎪
⎪ 0 = ηcc Rs iαβcc + dt
d
φαβ
⎪
⎪ 
We define Park’s transformation as:
⎪
⎪
cc
⎨ φ = Lm (i +i + 2 ηcc i )
αβs αβs αβr 3 αβcc
(5) xdq = P (−θ) xαβ (9)
⎪
⎪ +Lf iαβs 
⎪
⎪
⎪
⎪ = Lm (iαβs + iαβr ) + 23 ηcc Lm iαβcc
Afterward, the faulty model will be expressed under Park’s
⎪
⎪ φαβ

⎪
⎪
r reference frame. So, short circuit current (8) becomes:
⎪
⎪ = 2
⎪
⎪ φ 3 ηcc Lm Q(θcc ) (iαβs + iαβr ) 2 ηcc
⎩ αβcc
+( 23 Lm + Lf ) ηcc
2
iαβcc idqcc = P (−θ) Q(θcc ) P (θ) U dqs (10)
3 Rs
where 3) Global Stator faulty model: Fundamentally, we show
ω is rotor electrical frequency that in faulty case, an induction machine can be characterized
Lm = 32 Lp : magnetizing inductance by two equivalents modes. The common mode model corre-
 sponds to the healthy dynamics of the machine (Park’s model)
cos(θcc )2 cos(θcc ) sin(θcc )
Q(θcc ) = whereas the differential mode model explains the faults. This
cos(θcc ) sin(θcc ) sin(θcc )2
model, very simple to implement because expressed in Park’s
frame, offers the advantage to explain the defect through a
If we neglect L f according to L m in short circuit flux short circuit element dedicated to the faulty winding. On the
expression (5), we can write new flux equations as: other hand, it is unsuitable in case of simultaneous defects on
⎧ several phases. Indeed, this representation is only adapted in
⎪ φαβ = φαβ + φαβ
⎪
⎪ case of single phase defect. In the presence of short circuits
⎨
s f m
= Lf iαβs + Lm (iαβs + iαβr − ĩαβcc ) on several phases, this model translates the defect by aberrant
(6)
⎪ φαβr = φαβm = Lm (iαβs + iαβr − ĩαβcc )
⎪ parameters values, because it takes into account only a single
⎪
⎩ φ̃
αβ
= ηcc Q(θcc ) φαβ winding.
cc m
To remedy it, we generalize this model by dedicating to each
where phase of the stator a short circuit element Q cck to explain a
 
2 3 possible faulty winding [4][6]. So, in presence of several short
ĩαβcc = − ηcc iαβcc , φ̃αβ = φ (7) circuits, each faulty element allows the diagnosis of a phase
3 cc 2 αβcc
by watching the value of the parameter. This simple deviation
φαβ , φαβ are respectively magnetizing and leakage flux. allows to indicate the presence of unbalance in the stator. Fig.
m f
3 shows the global stator faulty model in dq Park’s axis with
Then, short circuit current equation becomes: global leakage referred to the stator.
2 ηcc d
ĩαβcc = Q(θcc ) φαβ (8) ω .P ( π / 2 ).φ dq
3 Rs dt m i dq
s i'dq
s
Rs Lf
s
i dq
According to this equation, the faulty winding B cc be- i dq
cc1
i dq
cc2
i dq
cc3
r

comes a simple unbalanced resistance element in parallel with U dq Qcc2 Qcc3

s Qcc1 Lm Rr
magnetizing inductance. The existence of localization matrix
Q(θcc ) in (8) makes complex the state space representation i dq
m
in Concordia’s axis. In a large range of industrial application,
voltage drop in R s and Lf is neglected according to stator
voltage U αβs then, we can put a short circuit element Q cc in Fig. 3. A global stator faulty model in dq frame
input voltage border (Fig. 2). Line currents iαβ s become the
sum of short circuit current ĩαβcc and usual current iαβs in Voltage and flux equations for faulty model with global
classical Concordia model. leakage inductance referred to the stator can be written as:
Park’s model (stator and rotor):
i'αβ Rs Lf ω .P ( π / 2 ).φ αβ
Rs idqs + dt
i αβ
s s m
U dqs = d
φdq + ω P ( π2 ) φdq
s s
i αβ
cc
i αβ
r φdq = Lf idqs + Lm (idqs + idqr )

s
(11)
0 = Rr idqr + dt
d
φdq
U αβ
s Q cc Lm Rr

r
φdq = Lm (idqs + idqr )
r
i αβ
m
Differential mode model (short circuit currents):
2 ηcck
Fig. 2. A short circuit model of induction machine icck = P (−θ) Q(θcck ) P (θ) U dqs , k = 1, 3 (12)
3 Rs
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y
Resultant dq stator currents become:
3
 For nb rotor bars, if we assume that the rotor cage can
idqs = idqs + icck
k=1
Each stator phase is characterized by its faulty parameters
(ηcck , θcck ), where k indicates one of the three stator phases.
B. Rotor faults modeling
Recently, rotor faults occurring in induction motors have
been investigated. Various methods have been used, including
measurement of rotor speed indicating speed ripple, in the
same way as spectral analysis of line current [9][12][16]. The
main problem concerning these monitoring methods is that
4
The number of turns in one rotor phase indeed fictitious.
(13) be replaced by a set of n b mutually coupled loops, each
loop is composed by two rotor bars and end ring portions
[1][5]; then the total number of rotor turns in one phase
for three-phases representation is equal to n3b . For nbb
broken rotor bars, faulty parameter η 0 becomes:
3 nbb
η0 = (15)
nb
1) Model of broken rotor bars: As stator faults modeling
(see equation 5), we can write voltage and flux equations of
new faulty winding B 0 in dq Park’s frame [7]:
⎧
⎨ 0 = η0Rr io + dt
dφ0
they are essentially invasive, requiring obvious interruption ⎪
of operation. Moreover, they are inappropriate under varying φ0 = 23 η0 Lm [ cos(θ0 ) sin(θ0 )] (idqs + idqr ) (16)
speed. For these reasons, parameter estimation is preferred for ⎪
⎩ 2 2
fault detection and diagnosis of induction motors [18][19]. + 3 η0 Lm i0
Parameter estimation is based on the simulation of a con- The current i 0 in the faulty winding B 0 creates a stationary
tinuous state-space model of induction motor. This model magnetic field H0 being directed according to broken rotor bar
assumes sinusoidal magnetomotive forces, non-saturation of axis. This additional magnetic field is at origin of faulty flux
magnetic circuit and negligible skin effect. Under these as- φ0 . By throwing i 0 and φ0 on dq Park axis, one associates the
sumptions, stator in dq Park’s axis and squirrel cage rotor stationary vectors:
made of nb bars can be modeled by an equivalent circuit. As
for stator fault, rotor fault is modeled by a new axis B 0 referred


to the first rotor axis a r by the angle θ 0 [7]. This additional cos(θ0 ) cos(θ0 )
idq0 = i0 , φdq = φ0
short circuited winding is at the origin of a stationary rotor sin(θ0 ) 0 sin(θ0 )
field H0 (t) steered according to rotor fault axis (Fig. 4). Equations (16) become relations between stationary vectors
according to rotor frame. So, voltage and flux equations of
ar
stator, rotor and faulty winding of induction motor are given
B ro k en ro to r θ0 by:
b a r a x is
bar 1
B0 bar 2 bar nb
U dqs = Rs idqs + + ω P ( π2 ) φdq
d
dt φdqs
bar nb - 1 s
φdq = Lf idqs + Lm (idqs + idqr + 23 η0 idq0 )
s
0 = Rr idqr + dt
d
φdq
r 
φdq = Lm (idqs + idqr ) + 23 η0 Lm idq0
r
E n d -rin g dφ
0 =  η0 Rr idqo + dtdq0 
2
φdq = 3 η0 Lm Q(θ0 ) (idqs + idqr + 23 η0 idq0 )
0
(17)
By using same transformation as to obtain primary trans-
br cr lation of an equivalent scheme in power transformer, we can
write global flux equations as:

Fig. 4. Broken rotor bar representation φdq = φdq + φdq = Lf idqs + Lm (idqs + idqr − ĩdq0 )
s f m
φdq = φdq = Lm (idqs + idqr − ĩdq0 )
r m
So, two additional parameters are introduced to explain rotor φ̃dq = η0 Q(θ0 ) φdq
faults: 0 m
(18)
• The angle θ 0 between fault axis (broken rotor bar axis)
with  
and the first rotor phase. This parameter allows the 2 3
localization of the broken rotor bar. ĩdq0 = − η0 idq0 , φ̃dq = φ (19)
3 0 2 dq0
• To quantify the rotor fault, we introduce a parameter η 0
equal to the ratio between the number of equivalent inter Also, current equation of faulty winding is given by:
turns in defect and the total number of inter turns in one dφdq dφdq
2 η0
healthy phase: ĩdq0 = R0 m
= Q(θ0 ) m
(20)
dt 3 Rr dt
Number of inter turns in defect
η0 = (14)
Total number of inter turns in one phase where Q(θ0 ) is localization matrix.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 5

2) Equivalent electrical schemes: According to equation i dq i'dq Rs Lf ω .P ( π / 2 ).φ dq

s s s
(20), faulty winding is a simple resistance element in parallel i dq
r
i dq i dq
cc2
i dq
cc1 cc3
with magnetizing inductance and rotor resistance. Because, Rr
U dq Qcc1 Qcc2 Qcc3
the reference frame is chosen according to rotor speed, it is s Lm
impossible to translate this element in stator border U dqs . So- Rdefect
i dq
lution consists in establishing equivalent scheme of induction m

machine with adding Park’s rotor resistance R r to faulty one

R0 . Thus, equivalent resistance R eq referred to rotor is the Fig. 6. Stator and rotor faulty model of induction motors
stake in parallel of rotor resistance and faulty resistance as:
−1
Req = Rr−1 + R0−1
(21) to be quasi stationary with respect to the dynamics of the
= Rr−1 + 23 η0 Rr−1 Q(θ0 )
electric variables, the model becomes linear but not stationary
By inversion, we obtain expression of an equivalent resis- with fourth order differential equations [5][19]. For simplicity,
tance matrix: the state vector is chosen composed of two-phase components
Req = Rr + Rdef ect of the dq stator currents i dqs and the rotor flux φdq . Then, the
(22) r
= Rr − 1+α α
Q(θ0 ) Rr continuous time model of the faulty induction motor, expressed
in the mechanical reference frame, is given by:
with α = 23 η0 .
ẋ(t) = A(ω) x(t) + B u(t) (24)
ω .P ( π / 2 ).φ dq
i dq
s
Rs Lf
s y(t) = C x(t) + D u(t) (25)
i dq
r
with
Rr
U dq
s Lm  T
x= ids iqs φdr φqr : state vector
R defect
i dq
m
 
Uds ids
u= , y= : input-output of system
Uqs iqs
Fig. 5. Broken rotor bars model

A11 A12
A(ω) = where
Thus, equivalent rotor resistance in broken rotor bars case is A21 A22
⎧
a series connection of a healthy rotor resistance R r and faulty ⎪
⎪ A11 = −(Rs + Req ) L−1 − ω P (π/2)
⎨ f
resistance Rdef ect . Fig. 5 is the resulting rotor fault circuit A12 = (Req L−1
m − ω P (π/2)) L−1
f
diagram in induction machines. ⎪
⎪ A = Req
⎩ 21
The angle θ0 allows an absolute localization of the faulty A22 = −Req L−1
m
winding according to the first rotor phase. Indeed, induced
bars currents composed a n b -phases system and faulty angle  T 
1
0 0 0 1 0 0 0
θ0 is fixed by initial rotor position according to stator one. On B = Lf
1 , C = ,
the other hand, when two broken rotor bars occur in machine, 0 Lf 0 0 0 1 0 0
3 2 ηcck
estimation of faulty angles θ 01 and θ02 allows to obtain a gap D= k=1 3 Rs P (−θ) Q(θcck ) P (θ)
angular Δθ between broken bars [7]:
Δθ = θ02 − θ01 (23) ncck
Req = Rr + α
1−α Q(θ0 ) Rr and ηcck = ns

C. Global stator and rotor faulty model

In previous sections, two models of stator and rotor faults
were presented. For a global simulation and detection of
simultaneous stator and rotor faults, we propose the global
faulty model including:
• Park’s model with the electrical parameters (R s Rr Lm
Lf )
• Stator faulty model with the three additional parameters
(ηcck , k = 1 − 3)
• Rotor faulty model with broken rotor bars parameters
(η0 , θ0 )
Fig. 6 shows a global electrical model of squirrel cage
induction motors for stator and rotor faults detection.
For simulation, it is necessary to write this faulty model in
state space representation. If mechanical speed ω is assumed
where ncck and ns are respectively the number of inter turns
short windings at k th phase and the total number of turns in
one healthy phase.
Q(θcck ) is a matrix depending on short circuit angle θ cck (if
the inter turn short circuit is at the phase a (resp. b and c),
then the angle θ cck is 0 rad (resp. 23π and 43π ).
The discrete-time model is deduced from the continuous one
by second order serie expansion of the transition matrix [19].
By using a second order serie expansion and the mechanical
reference frame, a sampling period T e around 1 ms can be
used. The usual first order serie expansion (Euler approxima-
tion) requires very short sampling period to give a stable and
accurate model. These approximation by serie expansion are
more precise with low frequency signals. Thus, discrete-time
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 6

model is given by: we consider the previous mathematical model (Eqs. 24-25)
and we define the parameter vector:
xk+1 = Φk xk + Bdk uk (26)
T
yk = C xk + D uk (27) θ = [Rs Rr Lm Lf ηcc1 ηcc2 ηcc3 η0 θ0 ] (30)
Assume that we have measured K values of input-output
where
(u(t), y ∗ (t) with t = k · Te ), the identification problem is then
Te T2 to estimate the values of the parameters θ. Then, we define
Φk = eATe = I + A + A2 e (28)
1! 2! the output prediction error:
Te2
Bd k = (I · Te + A )B (29) εk = y∗k − ŷk (θ̂, u) (31)
2 · 1!
and xk = x(tk ) and y k = y(tk ). The components of the where predicted output ŷ k is obtained by numerical simulation
known input vector u k are the average of the stator voltage of the state space faulty model (Eq. 27) and θ̂ is an estimation
between tk and tk+1 . of true parameter vector θ.
As a general rule, parameter estimation with OE technique
III. D IAGNOSIS PROCEDURE is based on minimization of a quadratic criterion defined as :

K K 
 
A. Introduction
J= εTk εk = (i∗dsk − îdsk )2 + (i∗qsk − îqsk )2 (32)
As soon as a fault occurs, the machine is no longer k=1 k=1
electrically balanced. Using previous faulty modes , electrical Usually, for induction motors, one has good knowledge on
parameters (Rs , Rr , Lm and Lf ) does not change and only electrical induction motors parameters, so it is very interesting
the faulty parameters (η cck and η0 ) vary to indicate a fault to introduce this information in the estimation process to
level. provide more certainty on the uniqueness of the optimum. For
Thus, during industrial operation, diagnosis procedure by this, we have applied the modification of the classical quadratic
parameter estimation of induction machines requires sequen- criterion [19][25], in order to incorporate physical knowledge.
tial electrical data acquisitions. Using each set of datas, identi- 1) Compound quadratic criterion: In order to incorpo-
fication algorithm computes a new set of electrical parameters rate physical knowledge or prior information, the classical
to know the magnetic state of the machine and new faulty quadratic criterion has been modified. The solution is to
parameters to have an approximation of the number of inter consider a compound criterion J c mixing prior estimation
turns short circuit windings and broken rotor bars. θ0 (weighted by its covariance matrix M 0 ) and the classical
criterion J (weighted by the variance of output noise δ̂ 2 ).
B. Identification algorithm Then, the compound criterion is usually defined as:
Parameter estimation is the procedure that allows the deter- J
Jc = (θ̂ − θ 0 )T M0−1 (θ̂ − θ 0 ) + (33)
mination of the mathematical representation of a real system δ̂ 2
from experimental data. Two classes of identification tech- Thus, the optimal parameter vector minimizing J c is
niques can be used to estimate the parameters of continuous the mean of prior knowledge and experimental estimation
time systems: Equation Error and Output Error [13][19][26] weighted by their respective covariance matrix.
• Equation Error techniques are based on the minimization In real case, we have no knowledge of the fault; indeed,
of quadratic criterion by ordinary least-squares [13][25]. no prior information is introduced on faulty parameter. Only
The advantage of these techniques is that they are simple electrical parameters (R s , Rr , Lm and Lf ) are weighted in the
and require few computations. However, there are severe compound criterion. Thus, covariance matrix is defined as:
drawbacks, especially for the identification of physical 1 1 1 1
parameters, not acceptable in diagnosis, such as the bias M0−1 = diag( 2 , σ 2 , σ 2 , σ 2 , 0, 0, 0, 0, 0) (34)
σR Rr Lm Lf
caused by the output noise and the modeling errors. s

• Output Error (OE) techniques are based on iterative min- 2 2 2 2

σR s
, σR r
, σL m
and σL f
are respectively the variance of
imization of an output error quadratic criterion by a Non parameters with prior information R s , Rr , Lm and Lf .
Linear Programming (NLP) algorithm. These techniques 2) Minimization of the compound criterion: We obtain the
require much more computation and do not converge optimal values of θ by Non Linear Programming techniques.
to an unique optimum. But, OE methods present very Practically, we use Marquardt’s algorithm [17][19] for off-line
attractive features, because the simulation of the output estimation:
model is based only on the knowledge of the input, so the

parameter estimates are unbiased [25][26]. Moreover, OE θ̂ i+1 = θ̂i − {[Jθθ + λ · I]−1 .Jθ }θ̂=θ (35)
i
methods can be used to identify non linear systems. For with  K 
these advantages, the OE methods are more appropriate k ·σ k,θ
εT
for diagnosis of induction motors [19] . Jc θ =2· M0−1 (θ̂
− θ0) − δ̂ 2
k=1
: gradient.
 K 
Parameter identification is based on the definition of a σk,θ ·σk,θ
T

Jcθθ ≈ 2 · M0−1 + k=1 δ̂2 : hessian.

model. For the case of fault diagnosis in induction machines,
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 7

λ : monitoring parameter. A. Motor experimental setup

∂ ŷ
σ k,θ = ∂θ : output sensitivity function. The motor used in the experimental investigation is a three
phases, 1.1 kWatt, 4-poles squirrel cage induction machine
In the proposed formulation, it is necessary to distinguish (Fig. 7). Stator windings were modified by addition of a
two kinds of sensitivities: number of tappings connected to the stator coils in the 1 st
 and 2nd phases (464 turns by phase). The other end of
σ y,θ = ∂y
∂θ : output sensitivity vector. theses external wires is connected to a terminal box, allowing
∂x (36)
σ x,θ = ∂θ : state sensitivity matrix. introduction of shorted turns at several locations and levels in
the stator winding. Different rotors, with broken bars, are used
For each parameter θ i , we determinate σ x,θi by numerical to simulate a bar breakage occurring during operation. The
integration of the differential system, obtained by the partial induction machine is driven by field oriented vector algorithm
differentiation of equation (24). Then σ x,θi is the solution of included in a speed control closed-loop and run under different
the state-space system: loads with the help of a DC generator mechanically coupled to
the motor. The data acquisition was done at a sampling period
∂A(θ) ∂B(θ) equal to 0.7 ms. Before identification, measured variables
σ̇ x,θi = A(θ) · σ x,θi + ·x+ ·u (37)
∂θi ∂θi are passed through a 4 th order butterworth anti-aliasing filter
whose cut-off frequency is 500 Hz.
Then by partial differentiation of equation (25), we get:
T P o sitio n
∂C(θ) ∂D(θ)
σ̇ y,θi = C(θ)T · σ x,θi + ·x+ ·u (38) Speed
∂θi ∂θi A C G rid C ontroller
~ 380 v CEGELEC E lectrica ls
3) Implementation: In this paper, we are only interested V F T V 4003B S en so rs
by prior information when performing parameter estimation
of physical systems. Prior information is mainly used to avoid B utterw orth
S p eed E x cita tio n F iters
aberrant estimates given by minimization of classical criterion. referen ce + P R B S
As a consequence, our interest is focused on the optimal choice
of θ 0 , M0 and δ̂ 2 . ia ib ic va vb vc

Prior information can result from two origin: D a ta

A cq u isition Shaft encoder
• Experiments or motor information given by industrials. FASTLAB 1024 pts/r
In this case, θ0 and M0 are obtained by electrotechnical 12 b its
tests performed on induction machines (locked rotor, load
shedding, ...) and all material characteristics. Fig. 7. Motor experimental setup
• Practically, prior information is given by physical knowl-
edge and partial estimation. Firstly, a set of experiment Identification algorithm needs persistent excitation to pro-
and identification of only electrical parameters with clas- vide appropriate estimation. This excitation is realized with a
sical criterion J is used in order to constitute an electrical P.R.B.Sequence equal to 90 rpm added to the reference of the
reference value data base, their pseudo-covariance matrix speed loop equal to 750 rpm.
M and the noise pseudo-variance are σ̂ 2 defined as:

 B. Experimental results
σ̂ 2 = 2(K−N
Jopt
) 1) Detection and localization of experimental faults: Dif-
M = σ̂ 2 (φTd φd + φTq φq )−1 (φd + φq )T (φd + φq )(φTd φd + φTq ferent
φq )−1 tests (10 realizations by experiment) have been per-
(39) formed (Table I) :
where K, N and J opt are respectively the number of datas, the
1) Healthy motor.
number of parameters and the optimal value of experimental
2) A short circuit of 18 inter turns in 1 st phase with 1
criterion. The matrix φ d and φq are matrix of output sensitivity
broken bar.
functions according to dq current axis.
3) A short circuit of 58 inter turns in 2 nd phase with 2
Thus, the covariance matrix M 0 is obtained by diagonal broken bars with Δθ = 2,82π
values of M . To evaluate the noise variance, it is necessary
4) A short circuit of 18 inter turns in 1 st phase and 58 in
to use δ̂ 2 > σ̂ 2 to take into account the effect of modeling
2nd phase with 2 broken bars with Δθ = 2π 28 .
errors.
5) A short circuit of 58 inter turns in 1 st phase and 29 in
2π
2nd phase with 2 broken bars with Δθ = 2,8 .
IV. E XPERIMENTAL TESTS Table I shows the mean of faulty parameter estimates for
10 acquisitions.
In order to validate this diagnosis procedure, a series of As observed in table I, there is good agreement between
experimental tests has been carried out. a real fault and its estimation. All faulty parameters vary to
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 8

(Ω) (Ω) (windings)

indicate the values of inter turn short circuit in the three-stator 11 3.86 60

windings and the number of broken rotor bars.

10.5
3.85 40
TABLE I 10
nc
E STIMATION RESULTS OF STATOR AND ROTOR FAULTS Rs Rr
3.84 20
9.5
Estimation results
Experiments (mean of 10 realizations)
9 3.83 0
n̂cc1 n̂cc2 n̂cc3 n̂bb Δθ 0 1 2 3 4 0 1 2 3 4 0 1 2
1. Healthy motor 5.57 3.52 −0.03 0.08 – (windings)
(H) (H)
2. case 2 17.86 −1.11 2.51 0.94 – 0.445 0.08 6
3. case 3 3.11 54.52 0.28 1.86 π/2.5
4. case 4 16.05 53.31 −2.54 1.88 2π/25 Lm Lf
0.435 0.077 4
5. case 5 53.69 26.87 −2.46 1.82 π/2.6

0.425 0.074 2

Indeed, parametric approach gives good estimations of short

0.415 0.071 0
circuit turns number n̂ cck . The estimation error is negligible 0 1 2 3 4 0 1 2 3 4 0 1 2

and does not exceed five turns in each situation of defect.

Fig. 9. Estimation of electrical and faulty parameters at faulty case
At simultaneous faults in several phases (case 4 and 5), we
observe that the estimates of the faulty parameters of each
phase is a realistic indication of the faults. This proves that This comparison is important because it is evident that only
each short circuit element explains the fault occurring at faulty parameters change when the faults occurs according
its phase and that no significant correlation exists between to prior information principle. Moreover, electrical parameter
these elements. Moreover, broken rotor bars estimation n̂ bb variations are function of the temperature and of the magnetic
gives a satisfactory indication from the fault. Angular distance state of the machine and are independent from the faults.
Estimation Δθ between the broken rotor bars (case 3, 4 and 3) Spectrum analysis: The residuals of the identification
5) allows to know rotor fault arrangement. algorithm of usual Park’s model give an indication of the
defect level [6][19]. Indeed, more the defect is important,
(windings) Bars
more the optimum criterion is high, indicating so an increase
Two broken
58 inter turns short
circuit windings
Estimation of prediction
rotor bars error. Then, in a faulty case, we observe an
60
Estimation 2
Real faults important modeling error due to the very restrictive condi-
Real faults
tions of Park’s model. Practically, Fourier’s analysis of these
45 1.5 identification residuals in faulty case using conventional Park’s
29 inter turns short One broken
circuit windings model
rotor bar is represented at Fig. 10.a. It exhibits spectrum lines at
30
18 inter turns short
1 low frequencies and around 50 Hz (corresponding respectively
circuit windings to rotor and stator faults).
15 0.5 Healthy
Healthy (dB) (dB)
machine machine 40 40
a. Usually Park’s b. Stator and rotor
0 0
10 Model 10 faulty model

−20 −20
Fig. 8. Estimation of electrical and faulty parameters at stator and rotor
faulty case −50 −50

−80 −80
Fig. 8 presents the evolution of inter turn short circuit
estimation in one phase for several experiments and the −110
0 50 100 150 200 (Hz)
−110
0 50 100 150 200 (Hz)

dispersion of the 10 estimations in different situations of rotor

faults. We observed that all the estimation results exhibit the
good approximation of the stator and rotor faults.
2) Parameters evolution: Fig. 9 gives, for one realization
in faulty situation (case 5), the evolution of electrical and
faulty parameters during estimation procedure. For electrical
state, It shows that their optimum values are achieved in only
four iterations. On the other hand, their variation according
to the initial values corresponding to prior information is
negligible. For faulty state, it is shown that their variations,
contrary from the electrical parameters, are very important.
Each faulty parameters varies to indicate stator and rotor
fault level occurring in the machine (example: n cc1 varies to
approach the 58 inter turns in defect presents on the 1 st phase
and nbb to approach 2 broken rotor bars).
Fig. 10. Spectral density (dB) of identification residuals
On the other hand, FFT analysis of the identification resid-
uals using the faulty model shows at Fig. 10.b that previous
spectrum lines have disappeared. This proves a posteriori that
this new faulty model explains correctly stator and rotor faults
occurring in induction machines.
For instance, we notice that stator shorted windings current
have a fundamental frequency equal to 2 F 0 , where F0 is
the power supply frequency; if we develop the expression
of the short circuit currents (equation 10), we find that the
frequency of their fundamental is equal to 2 F 0 . This explains
the absorption of the spectrum lines at this frequency by the
differential faulty model.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y
V. C ONCLUSION
A new model dedicated to squirrel cage induction machines
has been presented for the realistic identification and detection
of stator and rotor faults. Firstly, the inter turn short circuit
winding have been modeled by a short circuit element. Each
element has been dedicated to a stator phase in order to explain
the fault. Secondly, a new equivalent Park’s rotor resistance
has been expressed to allow the decreasing of the number of
rotor bars in faulty situation. Parameter estimation is used to
perform fault detection and localization. Prior information on
electrical parameters of Park’s model (common model) has
been introduced in the optimum search.
Experimental tests illustrate the efficiency of this technique
for use in off-line stator and rotor faults diagnosis of induction
machine under varying speed. The estimates of the number
of inter turns short circuit windings and broken rotor bars in
different realizations give a good approximation of the fault
level in the machine.
Our next objective will be to detect and localize the faults
without the use of a speed sensor. So, it will be necessary to
develop a speed estimator with only electrical signals.
R EFERENCES
[1] A. Abed, L. Baghli, H. Razik and A. Rezzoug, Modelling induction
motors for diagnosis purposes, EPE’99, Lausanne, Suisse, pp. 1-8,
September 1999.
[2] A. B ELLINI , F. F ILIPPETTI , G. F RANCESCHINI AND C. TASSONI , To-
wards a correct quntification of induction machines broken bars through
input electric signals, ICEM’00, Espoo, Finland, pp. 781-785, 28-30 aoùt
2000.
[3] R. I SERMANN , Fault diagnosis of machines via parameter estimation
end knewledge processig - Tutoriel paper, Automatica, VOL. 29, No . 04,
pp. 815-835, 1993.
[4] S. Bachir, S. Tnani, T. Poinot and J. C. Trigeassou, Stator fault diagnosis
in induction machines by parameter estimation, IEEE International
SDEMPED’01, Grado, Italie, pp. 235-239, September 2001.
[5] S. Bachir, S. Tnani, G. Champenois and J. C. Trigeassou, Induction
motor modeling of broken rotor bars and fault detection by parameter
estimation, IEEE International SDEMPED’01, Grado, Italie, pp. 145-149,
September 2001.
[6] S. Bachir, S. Tnani, J. C. Trigeassou and G. Champenois, Diagnosis
by parameter estimation of stator and rotor faults occuring in induction
machines, EPE’01, Graz, Autriche, August 2001.
[7] S. Bachir, S. Tnani, G. Champenois and J. Saint-Michel, Modélisation
et diagnostic des ruptures de barres rotoriques par identification
paramétrique, Electrotechnique du Futur’01, Nancy, France, pp. 165-
170, November 2001.
[8] S. Bachir, Contribution au diagnostic de la machine asynchrone par
estimation paramétrique, Thèse de doctorat, Université de Poitiers,
France, 2002.
[9] F. Fillippitti, G. Franceshini, C. Tassoni and P. Vas, Broken bar detection
in induction machine : comparaison between current spectrum approach
and parameter estimation approach, IAS’94, NewYork, USA, pp. 94-102,
1999.
[10] P. M. Frank, Fault diagnosis in dynamical systems using analytical and
knowledge based redundancy - A survey, Automatica, VOL. 26, No . 3,
pp. 459-474, 1990.
[11] G. Grellet and G. Clerc, Actionneurs électriques. Principes, modèles et
commande, Eyrolles, Paris, 1997.
[12] A. G. Innes and R. A. Langman, The detection of broken bars in variable
speed induction motors drives, ICEM’94, December 1999.
[13] L. Ljung, System identification: Theory for the user, Prentice Hall,
USA, 1987.
[14] A. Makki, A. Ah-jaco, H. Yahoui and G. Grellet, Modelling of capacitor
single-phase asynchronous motor under stator and rotor winding faults,
IEEE International SDEMPED’97, Carry-le-Rouet, France, pp. 191-197,
September 1997.
9
[15] M. G. Maléro and Al, Electromagnetic torque harmonics for on-
line interturn shortcircuits detection in squirrel cage induction motors,
EPE’99, Lausanne, Suisse, September, 1999.
[16] S. T. Manolas, J. Tegopoulos and M. Papadopoulos, Analysis of squirrel
cage induction motors with broken rotor bars, ICEM’96, Vigo, Spain,
pp. 19-23, 1996.
[17] D. W. Marquardt, An Algorithm for least-squares estimation of non-
linear parameters, Soc. Indust. Appl. Math, VOL. 11, NO. 2, pp. 431-
441, 1963.
[18] S. Moreau, J. C. Trigeassou, G. Champenois and J. P. Gaubert, Diagnosis
of Induction Machines: A procedure for electrical fault detection and
localization, IEEE International SDEMPED’99, Gijon, Espagne, pp. 225-
230, September 1999.
[19] S. Moreau, Contribution à la modélisation et à l’estimation
paramétrique des machines électriques à courant alternatif: Application
au diagnostic, Ph.D. Thesis, Université de Poitiers, France, 1999.
[20] J. Ragot, M. Daraouach, D. Maquin and G. Bloch, Validation de
données et diagnostic, Traité des nouvelles technologies, série diagnostic
et maintenance, Hermés, 1990.
[21] J. Richalet, A. Rault and R. Pouliquen, Identification des processus par
la mèthode du modèle, Gordon Breach, Thérie des systèmes, No . 04,
1971.
[22] M. Staroswiecki M. Hamad, Validation of measurements and detection
of sensors failures in control systems, Signal processing III: Theory and
applications. I.T. Young Editor, Elsevier, 1986.
[23] E. Schaeffer, Diagnostic des machines asynchrones: modèles et outils
paramètriques dédiés à la simulation et à la détection de défauts, Ph.D.
Thesis, Université de Nantes, France, 1999.
[24] G. Seguier and F. Notelet, Electrotechnique industrielle, Lavoisier Tec
et Doc, Paris 1994.
[25] J. C. Trigeassou, Recherche de modèles expérimentaux assistée par
ordinateur, Technique et Documentation Lavoisier, Paris, 1988.
[26] J. C. Trigeassou, J. P. Gaubert, S. Moreau and T. Poinot, Modélisation
et identification en génie électrique à partir de résultats expérimentaux,
Journées 3EI’99, Supelec Gif-sur-Yvette, March 1999.
[27] P. Vas, F. Filippetti, G. Franceschili and C. Tassoni, Transient modelling
oriented to diagnostics of induction machines with rotor asymmetries,
ICEM’94, December 1994.
[28] Gdr SDSE, Sûreté et disponibilité des Systèmes Electrotechniques,
Villeurbanne, du 20 au 21 janvier 2000.
Smail Bachir prepares the Ph. D. degree in Au-
tomatic and control from Poitiers University in the
department of Electrical engineering of LAII labo-
ratory (France). He was born in 1974 in Algeria.
He is a member of scientific department from Leroy
Somer Society in France. His research interest in-
cludes electrical machines, parameter estimation and
fault diagnosis.
Slim Tnani was born in Tunisia in 1967. He re-
ceived the Ph.D in 1995 from the university of
Franche-comté (France). Now, he is professor as-
sistant at the university of Poitiers (France). His
major interest in research are modelling of elctrical
machines and active filtering in power systems.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. XX, NO. Y 10

Jean-Claude Trigeassou was born in France in

1946. He received the Ph.D degree in 1980 from the
Ecole Nationale Supérieure de Mécanique de Nantes
(France) and the habilitation degree in 1987 from the
University of Poitiers (France). Now, he is profes-
sor at the University of Poitiers. His major fields
of interest in research are modelling, identification
and parameter estimation of physical systems, with
particular application to the diagnosis of electrical
machines.

Gérard Champenois was born in France in 1957.

He received the Ph.D. degree in 1984 and the ”ha-
bilitation degree” in 1992 from the Institut National
Polytechnique de Grenoble (France). Now, He his
professor at the University of Poitiers (France). His
major fields of interest in research are electrical
machines associated with static converter ;control,
modelling and diagnosis by parameter identification
techniques.