MITSUBISHI ELECTRIC RESEARCH LABORATORIES

http://www.merl.com

Off-the-grid compressive sensing for broken-rotor-bar fault

detection in squirrel-cage induction motors
Liu, D.; Lu, D.
TR2015-096 September 2015

Abstract
In this paper, we propose an off-the-grid compressive sensing based method to detect broken-
bar fault in squirrel-cage induction motors. To validate our method, we first build a dynamic
model of squirrel-cage induction motor using multi-loop equivalent circuit to simulate motor
current under fault conditions. We then develop an off-the-grid compressive sensing algorithm
to extract the fault characteristic frequency from the simulated motor current by solving an
atomic norm minimization problem. Comparing to other fault detection methods via motor
current signature analysis, our method yields high resolution in extracting low-magnitude
fault characteristic frequency with only 0.7 second measurements. Simulation results validate
our proposed method.

2015 IFAC Symposium on Fault Detection, Supervision and Safety for Technical Pro-
cesses (SAFEPROCESS)

This work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in
whole or in part without payment of fee is granted for nonprofit educational and research purposes provided that all
such whole or partial copies include the following: a notice that such copying is by permission of Mitsubishi Electric
Research Laboratories, Inc.; an acknowledgment of the authors and individual contributions to the work; and all
applicable portions of the copyright notice. Copying, reproduction, or republishing for any other purpose shall require
a license with payment of fee to Mitsubishi Electric Research Laboratories, Inc. All rights reserved.

Copyright
c Mitsubishi Electric Research Laboratories, Inc., 2015
201 Broadway, Cambridge, Massachusetts 02139
 Off-the-Grid Compressive Sensing for
Broken-Rotor-Bar Fault Detection in
Squirrel-Cage Induction Motors
Dehong Liu ∗ Dingguo Lu ∗∗,1
∗
Mitsubishi Electric Research Laboratories, Cambridge, MA 02139
USA (e-mail: [email protected])
∗∗
University of Nebraska-Lincoln, Lincoln, NE 68588, USA (email:
[email protected])

Abstract: In this paper, we propose an off-the-grid compressive sensing based method to

detect broken-bar fault in squirrel-cage induction motors. To validate our method, we first build
a dynamic model of squirrel-cage induction motor using multi-loop equivalent circuit to simulate
motor current under fault conditions. We then develop an off-the-grid compressive sensing
algorithm to extract the fault characteristic frequency from the simulated motor current by
solving an atomic norm minimization problem. Comparing to other fault detection methods via
motor current signature analysis, our method yields high resolution in extracting low-magnitude
fault characteristic frequency with only 0.7 second measurements. Simulation results validate
our proposed method.

Keywords: Fault detection, broken bar, squirrel-case induction motor, compressive sensing

1. INTRODUCTION is also termed the characteristic frequency. BRB fault

detection via motor current signature analysis(MCSA) is
The squirrel-cage induction motors suffer from a variety basically detecting the characteristic frequency component
of mechanical and electrical faults, of which about 10% (1 − 2s)fs .
are related to broken-rotor-bar(BRB) fault Zhang et al.
However, it is challenging to detect the characteristic
(2011). The BRB fault can be caused by either imperfec-
component due to the following factors. First, the magni-
tions in the manufacturing process, or non-smooth opera-
tude of characteristic frequency is relatively small, typical
tions such as direct-on-line starting duty cycles and pul-
30∼40dB lower than that of the fundamental power supply
sating mechanical loads. Although the BRB fault generally
frequency. Second, the characteristic frequency (1 − 2s)fs
does not lead to instant failures to the induction motors,
is very close the power supply frequency fs . Under steady
it causes serious secondary effects, such as excessive vibra-
operating condition, the frequency distance between the
tions, poor starting performance, and torque fluctuation,
characteristic frequency and the fundamental frequency fs
etc. Even worse, the broken part may hit stator windings
can be as small as 0.01fs . It is generally difficult to distin-
at high velocity, causing catastrophic failures on winding
guish the characteristic frequency from the fundamental
insulation. Therefore, it is of great importance to detect
frequency using the traditional Fourier spectral analysis.
the BRB fault in time so as to reduce the cost of mainte-
Although a big extension of measurement time may be
nance and repair Santos et al. (2006).
helpful, it requires near constant load to ensure both the
In order to detect the BRB fault, signatures are extracted slip and motor current remain stable during the whole
from motor current, air-gap flux, torque, and vibration, measurement period, otherwise the load fluctuation will in-
etc, for further analysis. Among these fault signatures, terfere the accuracy of fault detection. This constant load
motor current signature is gaining more attention for its requirement in many cases can be troublesome because
non-invasiveness and low cost. When there exists a BRB the unavoidable load fluctuation in reality especially over
fault in the squirrel-cage induction motor, the rotor oper- a long measurement period. Therefore, it is necessary to
ates asymmetrically, inducing extra frequency components develop a high frequency resolution analysis method using
fsb = (1 ± 2κs)fs in the stator current Filippetti et al. very short time measurements to meet the real situations
(1998), where s is the speed slip, typically ranges from in BRB detection.
0.005 to 0.05 under steady operating condition; fs is the
In the past decades, researchers have developed different
power supply frequency; and κ = 1, 2, 3, ... is the harmonic
MCSA methods such as Fourier transform spectral analy-
frequency index. Among these extra components, the (1 −
sis, short-time Fourier transform Bellini et al. (2001); Zhao
2s)fs component is the strongest one and typically treated
and Lipo (1996); Toliyat and Lipo (1995) and subsequent
as the indicator of a BRB fault. Thus frequency (1 − 2s)fs
high resolution spectral analysis using ESPRIT Xu et al.
1 Initial work was done while Dingguo Lu was an intern with Mit-
(2012) and MUSIC Kim et al. (2013). Although existing
subishi Electric Research Laboratories, 201 Broadway, Cambridge, methods can achieve high resolution in BRB detection,
MA 02139, USA.
they have several limits. First, they still require seconds neglect magnetic saturation and assume linear magnetic
of measurements under constant load. Any load fluctua- characteristics.
tion within the measurement period could interfere the
The equivalent circuit of squirrel-cage induction motor
accuracy of fault detection. Second, they are not capable
is shown in Fig.1. Assuming there are n rotor bars, the
of detecting early-stage fault, of which the characteristic
squirrel-cage rotor can then be modeled as n + 1 indepen-
frequency component is very weak.
dent current loops, where n of them are identical circuit
In recent years, the development of compressive sensing loops under ideal condition, with each loop consisting
(CS) provides us a feasible solution to analyze high resolu- of two adjacent rotor bars connected by two end ring
tion frequency components even with few measurements. portions. The remaining circuit loop is formed by one of
Compressive sensing is an innovative method to capture the end rings. So, the current distribution in rotor can be
and represent sparse or compressible signals at a rate specified in terms of (n + 1) independent loop currents,
well below its Nyquist sampling rate Baraniuk (2007); i.e., n rotor-bar loop currents ij (1 ≤ j ≤ n) plus one end
Candès and Wakin (2008). This sampling rate reduc- ring loop current ie .
tion is achieved by measuring uncorrelated or random-
a
ized projections of the sparse signals and reconstructing Re Le Re Le Re Le
ua ia
the sparse signal using improved signal models and non-
Rb Rb Rb Rb
linear reconstruction algorithms. In spectral analysis this 0 … in i1 i2 …
sampling rate reduction means with a fewer amount of ub ib Lb Lb Lb Lb
b
measurements, we are able to reconstruct the same res- u ic Re Le Re Le Re Le
c c
olution frequency spectrum, or with the same amount of (b)
measurements but higher resolution than the traditional (a)
methods. In MCSA based fault detection, the motor cur- Fig. 1. Equivalent circuit of (a)stator windings, and
rent spectrum with fault characteristic frequency exhibits (b)rotor in squirrel-cage induction motor
sparse characteristics in the frequency domain. Therefore,
the characteristic frequency component can be resolved
with high resolution using compressive sensing based tech- 2.1 Stator Voltage and Flux Equations
niques. Considering the fact that the characteristic fre-
quency is distributed in the continuous frequency domain, Based on the equivalent circuit, the voltage and flux
we consider the off-the-grid compressive sensing technique linkage equations for the stator windings can be written
Tang et al. (2013). as:
dΨs
In order to verify our method, we build a dynamic model U s = Rs Is + , (1)
dt
using multi-loop equivalent circuit to simulate stator cur- Ψs = Ls Is + Msr Ir , (2)
rent, in which a broken bar fault is modeled by an open
where the stator voltage
circuit while an early stage broken-bar fault is simulated
T
using an increased resistance of the fault branch. We Us = [ua ub uc ] , (3)
then develop an off-the-grid compressive sensing algorithm with
to extract the characteristic frequency component in the ua = U0 cos(2πfs t + φ0 ), (4)
simulated stator current.
the stator current
This paper is organized as follows. In Section 2, we describe T
Is = [ia ib ic ] , (5)
the dynamic model of induction machine for stator current
simulation under healthy and fault conditions. In Section the stator winding flux
T
3, we introduce compressive sensing fundamentals and our Ψs = [ψa ψb ψc ] , (6)
off-the-grid compressive sensing algorithm. Fault detection the stator resistance
results on simulated data are presented in Section 4. "
Rs 0 0
#
Finally, we draw conclusions in Section 5.
Rs = 0 R s 0 , (7)
0 0 Rs
the stator inductance
2. DYNAMIC MODEL OF INDUCTION MOTOR
Ls Mab Mac
" #
Ls = Mba Ls Mbc , (8)
In squirrel-cage induction motors, the stator consists of Mca Mcb Ls
three sinusoidally distributed windings, displaced by 120◦
spatial angle. The rotor contains longitudinal conductive the stator-rotor mutual inductance
ma1 ma2 · · · man mae
" #
bars connected at both ends by shorting rings, forming
a squirrel-cage like shape. As the induction motor is Msr = mb1 mb2 · · · mbn mbe , (9)
operating, the stator windings set up a rotating magnetic mc1 mc2 · · · mcn mce
field through the rotor, inducing electrical current in the and the rotor current
rotor bars, producing force acting at a tangent orthogonal T
Ir = [i1 i2 · · · in ie ] . (10)
to the rotor, and resulting in torque to turn the shaft.
Here we use bold capital letters for matrices, regular
In the following part of this section, we first develop a capital letters for constant parameters and small letters
dynamic model for motors in normal healthy condition, for time-variant parameters. It is important to note that
then extend it to fault conditions. For simplicity, we the stator winding resistance in (7) and the inductance of
stators in (8) are constant under our assumption, while for solving differential equations to simulate the stator
the stator-rotor mutual inductance in (9) varies with the current during dynamic operation. In this paper, we use
angular position of rotor. This is because the mutual the fourth-order Runge-Kutta method Butcher (1987).
inductance is related to the relative position between the
stator windings and the rotor bars, which changes during 2.4 Motor Parameters
operation.
Under normal healthy condition, the inductances and
2.2 Rotor Voltage and Flux Equations resistances in (7), (8), (9), (15) and (16) can be calculated
Luo et al. (1993); Boucherma et al. (2006). To save the
The voltage and flux linkage equations for the rotor loops space of this paper, we skip the details of parameter
can be written as: calculations.
dΨr
Ur = Rr Ir + , (11) When one bar is fully broken, the related branch becomes
dt open circuit. Then the totally number of circuit loops is
Ψr = Lr Ir + Mrs Is , (12) reduced by one since the related two loops are replaced by
where the rotor voltage a new loop with doubled end-ring segments, as shown in
T
Ur = [u1 u2 · · · un ue ] = [0](n+1)×1 , (13) Fig. 2(a). Consequently, in (11)∼(16) the corresponding
loops should be removed or rebuilt with the equivalent
the rotor flux parameters.
T
Ψr = [ψ1 ψ2 · · · ψn ψe ] , (14)
Re Le Re Le Re Le Re Le Re Le Re Le
the rotor resistance Rb Rb Rb Rb R’b Rb Rb
 R −R 0 · · · 0 −R … ij ij+2 … … ij ij+1 ij+2 …
Re

0 b b Lb Lb Lb Lb Lb Lb Lb
..
−Rb R0 −Rb . 0

0 Re

 Re Le Re Le Re Le Re Le Re Le Re Le
. (a) (b)
 
 0 −Rb R0 . . 0 0 Re
 
Rr =  .
, (15)
 .. .. .. .. .. .. .. 
Fig. 2. Equivalent circuit of fault conditions: (a)Fully
. . . . . .

broken bar, and (b)Early-stage broken bar
 
 0
 0 0 · · · R 0 −R b Re 

−Rb 0 0 · · · −Rb R0 Re
Re Re Re · · · Re Re nRe If the bar is not fully broken, but in an early stage of
broken fault condition, we can simulate this intermediate
with R0 = 2(Rb + Re ), the rotor inductance damage state by increasing the corresponding bar resis-
tance to a certain value Rb0 , as shown in Fig. 2(b). The
 
Lr + L0 M12 − Lb · · · M1n − Lb
M13 Le
. more severe the damage, the larger the resistance Rb0 . As a
 M21 − Lb Lr + L0 M23 − Lb . . M2n
 
Le  result, the related entries in our model should be modified.
 
 ..  For example, if the bar shared by the j th and the (j + 1)th
.
 M.31 M32 − Lb Lr + L0
Lr =  M3n Le , (16)
. . . .. ..
 loops is damaged, then the corresponding resistance sub-
 .. .. .. .. 

M − L
. .  block Rrj,(j+1) = [R0 , − Rb ; −Rb , R0 ] in (15) should be
Mn2 Mn3 · · · Lr + L0 Le modified to [R0 − Rb + Rb0 , − Rb0 ; −Rb0 , R0 − Rb + Rb0 ].

n1 b
Le Le Le ··· Le nLe
Comparing to other models such as the d-q model Santos
with L0 = 2(Lb + Le ), and the rotor-stator mutual
et al. (2006), this equivalent circuit model is straightfor-
inductance Mrs = MTsr .
ward to understand and flexible to simulate intermediate
fault state. It is important to note that this model may
2.3 Mechanical Equations become ill-conditioned when the ratio between Rb0 and Rb
is too large. A small step size inversely proportional to
The mechanical equation of the induction motor can be the ratio in the fourth-order Runge-Kutta method is then
expressed as: necessary for proper simulation results.
dωr
Te − Tl = J , (17)
dt 3. COMPRESSIVE SENSING BASED DETECTION
where Te represents the electromagnetic torque, Tl is the
load, J stands for the rotor inertia, and ωr is the angular
3.1 Background of Compressive Sensing
velocity. Using the basic principle of energy convertion, the
electromagnetic torque can be calculated as:
In the Shannon/Nyquist sampling theorem, the sampling
dMsr rate (the so-called Nyquist rate) should be at least twice
Te = ITs Ir . (18)
dθr the signal bandwidth in order to exactly recover the orig-
The angular velocity can be expressed in terms of change inal signal. Compressive sensing is an innovative method
rate of rotor’s angular position θr as follows: to capture and represent sparse or compressible signals at
dθr the rate well below Nyquist rate Candès and Wakin (2008);
ωr = . (19) Baraniuk (2007). This sampling rate reduction is achieved
dt
by using uncorrelated projection measurements, improved
In summary, equations (1) ∼ (19) form a dynamic model of signal models, and non-linear reconstruction algorithms.
induction motors with unknown stator and rotor currents. Assume the signal under observation x can be represented
Given the motor parameters, we can use standard methods by a sparse coefficient vector α with basis A as x = Aα.
Measurements are made on x through a projection matrix K
X wk
K
X
Φ as follows xj = ck ei2π 2W j = ck ei2πfk j , (28)
y = Φx = ΦAα = Σα. (20) k=1 k=1
wk
∈ − 21 , 21 , which is exactly the same as
 
According to the CS theory, if Σ satisfies the re- with fk = 2W
stricted isometry property, α can be reconstructed with (22) after a trivial translation of the frequency domain.
overwhelming probability close to 1 by solving an L1 - With off-the-grid compressive sensing, we are able to
minimization problem Candès and Wakin (2008) perform fault detection by extracting sparse characteristic
min ||α||1 s.t. y = Σα. (21) frequency component in the stator current signal with
improved performance.
In order to apply the CS theory to real applications,
researchers typically adopt a discretization procedure to 3.2 Off-the-Grid Compressive Sensing Algorithm
reduce the continuous parameter space to a finite set of
grid points. This strategy generally yields state-of-the-art There are several reconstruction algorithms to reconstruct
performance for problems where the true parameters lie the sparse signal, with different computational complexity
on the grid. However, in cases where the true parameters and accuracy Tang et al. (2013); Rao et al. (2013). These
do not fall into the finite grid, the signal cannot often algorithms minimize a least-square loss function that mea-
sparsely represented by the discrete basis. In this situation, sures the difference between the signal representation and
off-the-grid compressive sensing provides us a solution by the observations, subject to a constraint in terms of an
generalizing the discrete basis to a continuous one Tang atomic norm as in (25). In general, it takes long time
et al. (2013). for those algorithms to converge to the extract sparse fre-
quency. In our fault detection applications, since we have
Suppose we observe the signal some prior knowledge about the frequency distribution,
K
X i.e., the characteristic frequency, if there exists any, should
xj = ck ei2πfk j , j ∈ {0, .., n − 1}, (22) be close to the fundamental frequency, we can predefine
k=1 the atoms for fast reconstruction. For example, the stator
with unknown K(K  n) frequencies f1 , ..., fK ∈ [0, 1] on current consists fundamental power supply frequency and
an index set {j} of size m(K < m ≤ n) selected uniformly possible fault characteristic frequency which is close to
at random. the power supply frequency. In our simulation, we refine
frequency grid in the neighborhood of the fundamental
Let ∆f be the minimum wrap-around distance of frequen- frequency with neighborhood radius 5Hz while only con-
cies on the unit circle sider coarse grid elsewhere. Based on this idea, we propose
∆f = min |fk − fj | . (23) a conditional gradient method with predefined atoms for
k6=j efficient reconstruction.
4
If ∆f ≥ (n−1) , there exists a constant cµ such that The algorithm is summarized as follows.
 
2 n K n Algorithm 1 Condition gradient with predefined atoms
m ≥ cµ max log , K log log (24)
δ δ δ (1) Input: Measurements y, predefined A, estimated
is sufficient to guarantee that we can recover x and localize bound c , frequency tolerance ;
the frequencies with probability at least 1 − δ with respect (2) Initialize: a0 ∈ A, τ ← 0, A0 ← [a0 ], c0 ← [c],
to the random samples and signs. Further numerical simu- x0 ← [A0 c0 ];
lations suggest that, although without theoretical support, (3) REPEAT
the separable frequency distance can be as small as n1 aτ +1 ← arg mina∈A h5g(xτ ), ai; {FORWARD}
Tang et al. (2013). The signal recovery process is realized Ãτ +1 ← [Ac aτ +1 ];
by solving an atomic-norm minimization problem γτ +1 ← arg minγ∈[0,1] g[xτ + γ(caτ +1 − xτ )];
1 2 {LINE SEARCH}
min g(x) := ky − Φxk2 s.t. kxkA ≤ c, (25)
x 2 c̃τ +1 ← [(1 − γτ +1 )cτ ; cγτ +1 ];
where y corresponds to observed noisy measurements, c is Merge adjacent frequencies within 
a constant bound and the atomic norm kxkA under the (4) UNTIL convergence
set of atoms A = {ei2πf t+φ } is defined as (5) Output: cτ , Aτ ;
( )
X X 4. SIMULATION
kxkA := inf ca : x = ca a, ca ≥ 0 . (26)
a∈A a∈A 4.1 Algorithm Validation
In our BRB fault detection problem, suppose the fre- To examine our algorithm, we consider a noise free signal
quencies wk lie in [−W, W ], and stator current x(t) is a represented as follows:
continuous signal of the form
K
X x(t) = 1.0 cos(2π50t + φ1 ) + 0.1 cos(2π150t + φ3 )
x(t) := ia (t) = ck ei2πwk t . (27)
+ 0.01 cos(2π48.62t + φ2 ). (29)
k=1
As we can see, the signal is composed of three frequency
By taking regularly spaced Nyquist samples at t ∈ components: a fundamental power supply frequency com-
{0, 1/2W, ..., (n − 1)/2W }, we observe ponent at 50 Hz with a unit magnitude, a third harmonic
frequency component with 1/10 unit magnitude and a real applications, we add noise of 30dB SNR into the sim-
characteristic fault frequency component at 48.62 Hz with ulated current signal. We then employ both the traditional
1/100 unit magnitude. The initial phases are randomly Fourier transform and our off-the-grid compressive sensing
chosen from [0, 2π). The Fourier frequency spectrum is method to reconstruct the sparse frequency components
shown in the upper part of Fig. 3 where we measure 0.7 of stator current signals under both healthy and fault
second current signal with sampling rate of 1kHz. Because conditions, with the results plotted in Fig. 5 for com-
of the leakage of fundamental frequency energy, it is not parison. The top-row figures, from left to right, show the
feasible to detect the fault characteristic frequency compo- state currents of healthy, two-broken bar, one-broken bar,
nent using the traditional Fourier transform method since and one early-stage broken bar conditions, respectively.
the frequency component is very close to the fundamental The middle-row figures show the corresponding Fourier
frequency and its magnitude is too low. When we use off- spectra and the bottom-row figures show the off-the-grid
the-grid compressive sensing technique on the same 0.7 CS reconstruction. As we can see, the characteristic fre-
second signal, we reconstruct all three components exactly, quency components of all the three fault conditions are
in both magnitude and frequency, as shown in the lower extracted successfully, even when the magnitude is about
part of Fig. 3. This result agrees with the aforementioned 40dB lower than the fundamental frequency component
off-the-grid theory in Section 3. and the traditional Fourier spectrum fails. As regarding
to the computational time, our it takes about 2.5 seconds
0 for our off-the-grid CS algorithm to reconstruct the signal
Magnitude(dB)

-20 on a 3.6GHz Intel Xeon CPU using Matlab.

-40

-60 While we take a closer look at the sparse characteristic

-80
0 50 100 150 200
frequency, we observe that under the same load condition,
Frequency(Hz) the characteristic frequencies are slightly different, which
0 True are 47.575Hz, 47.675Hz, and 48.125Hz for two-broken bar,
Magnitude(dB)

Recovered
-20 one-broken bar, and one early-stage broken bar conditions
-40
respectively. This interesting observation indicates that
there is a possibility to diagnose the severity of the
-60
0 50 100
Frequency(Hz)
150 200 fault conditions using the high resolution off-the-grid CS
analysis. The closer to the power supply frequency, the less
severe the fault. This is however under investigation in our
Fig. 3. Fourier spectrum (upper) and CS reconstruction future work.
(lower)
Phase-A stator current(A)

20

10

4.2 Off-the-Grid CS Based Fault Detection 0

-10

-20
To further test our method, we simulate the stator current 0 0.2 0.4
Time(s)
0.6 0.8 1

under both healthy condition and fault condition using our 3000
multi-loop circuit model. The model parameters are ex-
Speed(RPM)

2000
tracted from a 1.5kW three-phase squirrel-cage induction
motor with parameters listed in Table 1. 1000

0
Table 1. Motor parameters 0 0.2 0.4
Time(s)
0.6 0.8 1

n fs (Hz) Rs (Ω) Re (µΩ) Rb (µΩ)

29 50 3.43 37 133.4 Fig. 4. Stator current (upper) and rotor speed (lower)
Ls (H) Lr (mH) Lb (µH) Le (µH) J(kg · m2 ) under healthy condition
0.51563 5.443 1.08 0.01 0.0014
5. CONCLUSION
Under healthy condition, the simulated stator current of
phase-A and the rotor speed are shown in Fig. 4. We notice In this paper, we propose to use off-the-grid compressive
that after about 0.5 second transient operation, the motor sensing technique to detect broken bar fault in squirrel-
runs in a steady state with a constant speed. cage induction motor. Simulation results show that due
For fault condition, we consider three different situations: to its high frequency resolution, off-the-grid compressive
two consecutive broken bars, one broken bar and one sensing can effectively detect broken-bar faults, even early
early-stage broken bar. For fully broken bar situations, we stage broken-bar faults with very short time measure-
employ the equivalent circuit model as shown in Fig.2(a). ments. In our future work, tests on experimental data
While for the early-stage broken bar, which is generally not will be considered for detecting and diagnosing broken-bar
trivial to simulate, we consider resistance Rb0 = 1000Rb to fault in squirrel cage induction motors.
simulation this early fault degradation for simplicity.
REFERENCES
We sample stator current in a period of 0.7 second when
the motor is operating in a steady state, assuming in Baraniuk, R.G. (2007). Compressive sensing [lecture
practice the fluctuated load can be treated as unchanged notes]. IEEE Signal Processing Magazine, 24(4), 118–
in such a short time. Considering measurement noise in 121.
 4 4 4 4

3 3 3 3

2 2 2 2

1 1 1 1
Magnitude(A)

Magnitude(A)

Magnitude(A)

Magnitude(A)
0 0 0 0

-1 -1 -1 -1

-2 -2 -2 -2

-3 -3 -3 -3

-4 -4 -4 -4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Time(s) Time(s) Time(s) Time(s)

(a) (b) (c) (d)

0 0 0 0

-10 -10 -10 -10

-20 -20 -20 -20

Magnitude(dB)

Magnitude(dB)

Magnitude(dB)

Magnitude(dB)
-30 -30 -30 -30

-40 -40 -40 -40

-50 -50 -50 -50

-60 -60 -60 -60

-70 -70 -70 -70

-80 -80 -80 -80

0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Frequency(Hz) Frequency(Hz) Frequency(Hz) Frequency(Hz)

(e) (f) (g) (h)

5 5 5 5

0 0 0 0

-5 -5 -5 -5

-10 -10 -10 -10

Magnitude(dB)

Magnitude(dB)

Magnitude(dB)

Magnitude(dB)
-15 -15 -15 -15

-20 -20 -20 -20

-25 -25 -25 -25

-30 -30 -30 -30

-35 -35 -35 -35

-40 -40 -40 -40

-45 -45 -45 -45

0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
Frequency(Hz) Frequency(Hz) Frequency(Hz) Frequency(Hz)

(i) (j) (k) (l)

Fig. 5. (a)-(d) Stator current under healthy, two-broken-bar fault, one-broken-bar fault, and one partial-broken-bar fault
respectively, (e)-(h) Fourier spectrum of (a)-(d), (i)-(l) off-the-grid CS reconstruction of (a)-(d).

Bellini, A., Filippetti, F., Franceschini, G., Tassoni, C., Rao, N., Shah, P., and Wright, S. (2013). Conditional
and Kliman, G.B. (2001). Quantitative evaluation of gradient with enhancement and truncation for atomic
induction motor broken bars by means of electrical norm regularization. In NIPS workshop on Greedy
signature analysis. IEEE Transactions on Industry Algorithms.
Applications, 37(5), 1248–1255. Santos, P.M., Correa, M.B.R., Jacobina, C.B., da Silva,
Boucherma, M., Kaikaa, M.Y., and Khezzar, A. (2006). E.R.C., Lima, A.M.N., Didiery, G., Raziky, H., and
Park model of squirrel cage induction machine including Lubiny, T. (2006). A simplified induction machine model
space harmonics effects. Journal of Electrical Engineer- to study rotor broken bar effects and for detection.
ing, 57(4), 193 – 199. In Proceedings of the 37th IEEE Power Electronics
Butcher, J.C. (1987). The numerical analysis of ordinary Specialists Conference(PESC), 1–7.
differential equations: Runge-Kutta and general linear Tang, G., Bhaskar, B., Shah, P., and Recht, B. (2013).
methods. Wiley-Interscience. Compressed sensing off the grid. IEEE Transactions on
Candès, E.J. and Wakin, M.B. (2008). An introduction to Information Theory, 59(11), 7465–7490.
compressive sampling. IEEE Signal Processing Maga- Toliyat, H.A. and Lipo, T.A. (1995). Transient analysis of
zine, 25(2), 21–30. cage induction machines under stator, rotor bar and end
Filippetti, F., Franceschini, G., Tassoni, C., and Vas, P. ring faults. IEEE Transactions on Energy Conversion,
(1998). AI techniques in induction machines diagnosis 10(2), 241–247.
including the speed ripple effect. IEEE Transactions on Xu, B., Sun, L., Xu, L., and Xu, G. (2012). An ESPRIT-
Industry Applications, 34(1), 98–108. SAA-based detection method for broken rotor bar fault
Kim, Y., Youn, Y., Hwang, D., Sun, J., and Kang, D. in induction motors. IEEE Transactions on Energy
(2013). High-resolution parameter estimation method Conversion, 27(3), 654–660.
to identify broken rotor bar faults in induction motors. Zhang, P., Du, Y., Habetler, T.G., and Lu, B. (2011). A
IEEE Transactions on Industrial Electronics, 60(9), survey of condition monitoring and protection methods
4103–4117. for medium-voltage induction motors. IEEE Transac-
Luo, X., Liao, Y., Toliyat, H., El-Antably, A., and Lipo, tions on Information Theory, 47(1), 34–46.
T.A. (1993). Multiple coupled circuit modeling of Zhao, Y. and Lipo, T.A. (1996). Modeling and control
induction machines. In Conference Record of the 1993 of a multi-phase induction machine with structural
IEEE Industry Applications Society Annual Meeting, unbalance. IEEE Transactions on Energy Conversion,
volume 1, 203–210. 11(3), 570–577.