Electric Power Systems Research 121 (2015) 1–13

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Analysis of fault signatures for the diagnosis of induction motors fed

by voltage source inverters using ANOVA and additive models
Oscar Duque-Perez a,∗ , Luis-Angel Garcia-Escudero b , Daniel Morinigo-Sotelo a ,
Pedro-Esteban Gardel c , Marcelo Perez-Alonso d,e
a
Department of Electrical Engineering, University of Valladolid – UVA, 47011 Valladolid, Spain
b
Department of Statistics and Operational Research, University of Valladolid, 47011 Valladolid, Spain
c
National University of Asuncion, UNA, Paraguay
d
Renfe Fabricación y Mantenimiento SA, 28036 Madrid, Spain
e
Antonio de Nebrija University, 28040 Madrid, Spain

a r t i c l e i n f o a b s t r a c t

Article history: Condition monitoring of voltage source inverters-fed induction motors is challenging since the influence
Received 12 June 2014 of the supply complicates the use of methods valid for utility fed motors. Analysis of variance is performed
Received in revised form over data obtained in a controlled laboratory experiment where a hole was progressively drilled into a
20 November 2014
rotor bar. Additive models are obtained to stand out the influence of the operating conditions and the
Accepted 21 November 2014
supply on the amplitude of the fault signatures. Useful conclusions for the design of diagnosis systems
for the detection of broken bars in induction motors are drawn from the analysis.
Keywords:
© 2014 Elsevier B.V. All rights reserved.
Additive models
Condition monitoring
Fault diagnosis
Induction motors
Spectral analysis
Statistical analysis

1. Introduction Motor current signature analysis (MCSA) is currently consid-

Induction motors are essential in many industries. Their sim-
plicity and ruggedness are outstanding advantages and make them
by far the most commonly used type of motor in sizes ranging
from fractional horsepower to grades of industrial applications [1].
However, owing to the thermal, electrical and mechanical stresses,
industry is confronted with unexpected faults and lifetime reduc-
tion of these motors [2].
The need to increase reliability against possible faults has
attracted considerable interest in fault diagnosis of induction
motors in recent years [3–7]. The main aim of a condition monitor-
ing system is to detect incipient faults before a potentially harmful
machine failure occurs with low missed and false alarms rates, in
order to discriminate among various machine conditions, classify-
ing faulty modes from normal modes.
∗ Corresponding author at: Paseo del Cauce 59, 47011 Valladolid, Spain.
Tel.: +34 983184533.
E-mail addresses: oscar.duque@eii.uva.es (O. Duque-Perez), lagarcia@eio.uva.es
(L.-A. Garcia-Escudero), daniel.morinigo@eii.uva.es (D. Morinigo-Sotelo),
pgardel@pol.una.py (P.-E. Gardel), marcelop@renfe.es (M. Perez-Alonso).
ered as a standard in preventive maintenance basically due to
the advantages of using non-invasive sensors [8,9]. Nevertheless,
the introduction of VSI-fed (voltage source inverters) motors in
applications where variable speed and torque are needed has pro-
duced significant changes in the field of diagnostics needing further
research in order to overcome various challenges such as higher
noise level (inherent floor noise reduces the possibility of true
fault signature recognition using line current spectrum) [10,11],
dynamically changing excitation frequency and the fact that fault
signatures can significantly change from open-loop to closed-loop
VSI operation. All these effects complicate the utilization of fre-
quency analysis methods [12]. VSI-fed motor faults have been
analyzed and initial results are given in literature [13–19] but fur-
ther investigation is still required [20].
The amplitude of fault harmonics is also affected by the sup-
ply voltage, making any of these anomalies more noticeable as the
supply voltage increases, provided that the machine is not working
under saturation condition. This is very important, and it must be
taken into account, particularly, with controlled-speed induction
motors, when this control is based on voltage regulation, keep-
ing constant the voltage/frequency ratio. Therefore, to develop a
generic condition monitoring method valid for any supply and

http://dx.doi.org/10.1016/j.epsr.2014.11.021
0378-7796/© 2014 Elsevier B.V. All rights reserved.
2 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13
Table 1
Specifications of the tested motors.
Motor Rated power (kW) Rated voltage (V)
M1 0.75 3 × 230/400
M2 1.1 3 × 230/400
at different operating conditions, further considerations must be
taken into account.
Considering this goal, in this paper it is analyzed the suitability
of a number of fault signatures obtained from the stator current
to diagnose rotor faults in induction motors fed by different sup-
plies. First, statistical tools such as boxplots and analysis of variance
(ANOVA) are used to analyze the data obtained from a case study
in which a cage fault condition has been produced by drilling
a hole in one of the bars of an induction motor. Different fault
conditions have been obtained by progressively making the hole
deeper.
Secondly, additive models are obtained, which are easy to inter-
pret but are more powerful analysis tools than linear models, as
it is shown in Section 5. The aim of this analysis is to evaluate
the potential and suitability of the data, composed by the afore-
mentioned fault signatures, to be used in a condition monitoring
task and to study the influence on the diagnosis of the type of
motor (two motors of different rated powers have been tested),
the motor supply (five supplies have been used, utility supply and
four different frequency inverters with different types of control)
and the motor load level (the motors have been tested from low
load to full load). It is also evaluated the capability to distinguish
among four different rotor bar states: healthy, incipient fault (cor-
responding to a 6 mm depth hole), half-broken bar and full broken
bar.
The results show that the same fault severity produces differ-
ent fault signature amplitudes depending on the type of the motor,
supply and load level. This is critical in the construction of a uni-
versal diagnosis procedure. Another significant contribution of this
paper is the use of the additive models to quantify the influence of
each of these influential variables in the amplitude of the fault sig-
natures, allowing the subtraction of the contribution of a variable in
the fault signature value. This way, when constructing a universal
diagnosis procedure, the bias produced by any of these variables
can be avoided.
2. Detection of broken rotor bars in VSI-fed induction
motors
Induction motor failure through broken rotor bars (where crack-
ing is experienced in the rotor conductors) is common in many
industrial applications. One of the reasons for this type of failures
is that large starting currents occur when cooling is at minimum,
resulting in thermal and mechanical stresses being at maximum.
The incidence of this failure mode is greatest when the start-up
time is relatively long and when frequent starts are required as
part of a heavy-duty cycle [21]. Initially, they start as high resis-
tance, causing high temperature, and then progress as cracking or
small holes in the rotor bars [22]; healthy bars in imminent vicin-
ity must conduct extra current, thus increasing the possibility of
cracking due to increased stresses [23]. To prevent such cumula-
tive destructive process, the problem should be detected as early as
possible [24]. Recently, improvements in machine design and man-
ufacturing have been very important in diminishing the occurrence
of other types of faults in induction motors but squirrel cage design
has only changed slightly [25].
When the cage winding is symmetrical and assuming a purely
and balanced sinusoidal voltage supply, there is only a forward
Rated current (A) Rated speed (rpm) Pole pairs
1.9 1395 2
2.6 1415 2
rotating field at slip frequency with respect to the rotor. If a rotor
asymmetry occurs, there will be a resultant backward rotating field
at slip frequency with respect to the forward rotating rotor. This
backward rotating field induces a voltage and a current in the
stator winding at (1 − 2s)f1 frequency (known as lower-sideband
harmonic, LSH), where s is the motor slip and f1 is the funda-
mental frequency. This induced current is the cause of torque and
speed pulsations, which at the same time, induce new electromo-
tive forces in the stator, and as a result, new counter currents are
produced at frequency (1 + 2s)f1 (known as upper-sideband har-
monic, USH). This process goes on indefinitely, until it is damped
and a pair of new sidebands appears around the main frequency f1
[26].
In VSI-driven induction motors, the line current will contain
additional time harmonics whose number, frequency and ampli-
tude will depend on the switching strategy of the semiconductors.
These time harmonics will generate new air gap spatial harmonics
or will modify the amplitude of the existing ones. That is, in addi-
tion to the characteristic motor spectrum, new harmonics will be
introduced, related to the fault condition, to the driven load or to the
system performance. Therefore, the current spectrum is affected
by many factors, including the motor characteristics (stator con-
nection, design and application), the power supply and the fault
condition.
When a motor with a cage asymmetry is fed by a non-sinusoidal
voltage supply, the process is very similar to a motor fed by a bal-
anced and sinusoidal one [27]. A pair of sidebands around the time
harmonics of the line current will appear at frequencies (k ± 2 ns)f1 ,
where k is the order of the line current time harmonic and n is
any positive integer. The number of sidebands will increase pro-
gressively, but their amplitude will decrease as they move away
from the main frequency, although the magnitude of this amplitude
attenuation will depend on the machine and driven load inertia and
the building characteristics of the motor, which, in turn, depend on
its purpose or application.
3. Case study
In this work, stator current is the signal measured and processed
to determine the condition of the induction motor. Two different
motors, star connected, were tested. Each motor was fed from a
set of three different sources. The first motor, named M1, was fed
from the utility supply (UT), an inverter by Telemecanique (TE)
and a laboratory inverter by Lucas Nülle (LN). The second motor,
named M2, was also fed from three sources: utility supply (UT)
and two frequency inverters by Allen Bradley (AB) and by Siemens
(SM). Table 1 shows the specifications of both motors. The control
of the TE and SM inverters was of V/f linear type, the AB inverter
had sensorless vector control, while the control of the LN inverter
permits an independent adjustment of voltage and frequency. All
four inverters operated with the slip compensation option off and
with an assigned frequency of 50 Hz. In each test, the stator cur-
rent was measured using a Hall effect current probe by Fluke (i30s)
and a DAQ board by National Instruments (PCI-6250M) as data
acquisition equipment. The software used was Labview DLLs for
data acquisition, Matlab for signal processing and R as statistical
software for data analysis.
 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13
Table 2
Description of the tests performed.
Condition state Description Hole depth M1 Hole depth M2
(mm) (mm)
1 Healthy motor 0 0
2 Incipient fault 6.4 5.7
3 Semi-broken bar 11.7 12.1
4 Full-broken bar 17 18
Four rotor conditions, corresponding to different fault severities,
were studied. The four fault severities were obtained by drilling a
hole of different depths (Table 2) into one of the rotor bars. As the
goal was to perform a statistical analysis, 200 tests were performed
for each motor and voltage supply, covering a broad range of load
conditions.
Power spectral density (PSD) using the fast Fourier transform
(FFT) was used to extract fault features. This choice determines
the parameters of the data acquisition process. To avoid frequency
aliasing, a high sampling frequency of 80 kHz was used. One of
the drawbacks of this signal processing technique is the spectral
leakage due to the use of a finite dataset to perform the spectrum
calculation. This is solved by the use of an appropriate window
function. In this work, the Hann window was applied to every full
measure since it decreases the spectral leakage with a good com-
promise between resolution and amplitude accuracy. In order to
have a reasonable frequency resolution (0.1 Hz), the time acquisi-
tion period was 10 s, and this gives a measurement with 800,000
samples.
The amplitude of the LSH and USH harmonics are normally used
as fault features for condition monitoring purposes. Instead of using
directly the FFT of each stator current, the PSD of the signal spec-
trum is calculated because it allows a much better identification
of the fault harmonics compared to the main component. Further-
more, the PSDs were normalized so that the fundamental harmonic
frequency was always 0 dB.
As mentioned in the introduction, condition monitoring of VSI-
driven induction motors is more challenging than for the utility
supply case. For example, the stator current spectrum is highly
affected in a different way by the specific VSI used as motor sup-
ply. To illustrate this, the PSD of the line current consumed by
motor M1 with a full broken bar in the rotor under two differ-
ent supplies is shown in Fig. 1. In Fig. 1a (UT supply), the fault
sidebands around the main harmonic are clearly seen and the
noise level is very low. The current PSD of the same motor fed
by TE supply is represented in Fig. 1b. Although operating con-
ditions in both tests are similar (slip: 4.6%), it is very remarkable
that the noise level is very high when the motor is fed with
the TE inverter. Additionally, an interharmonic appears between
the lower sideband and the main harmonic and this makes more
difficult to identify this fault signature in the current spectrum
[28,29].
For condition monitoring purposes, instead of the peak absolute
value in dBs, as fault signature is considered the difference between
the peak amplitude and the noise floor in the PSD of the stator cur-
rent calculated by means of a wavelet function (the Daubechies
orthonormal wavelet of length equal to 8, which is the least asym-
metric family, has been used [30]). The reason of choosing this
measurement, instead of the peak absolute value in dBs, is because
it is considered essential to take into account the different level
of noise introduced by the power supply. For example, in Fig. 1,
the first band of fault signatures has similar amplitudes for both
motor supplies. However, while they are clearly observed for UT
supply, in the inverter-fed motor case they are almost hidden by
the noise.
3
Table 3
Analysis of variance. Motors M1 and M2. LSH.
Estimate Std. error t p-value
Intercept 11.9338 0.3336 35.768 <2e−16***
Motor M2 −3.8749 0.3446 −11.244 <2e−16***
Condition 2 1.8392 0.4683 3.928 9.06e−05***
Condition 3 9.8605 0.4687 21.038 <2e−16***
Condition 4 14.7252 0.4742 31.050 <2e−16***
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.
Residual standard error: 5.984 on 1214 degrees of freedom.
Multiple R-squared: 0.5314, adjusted R-squared: 0.5298.
F-statistic: 344.1 on 4 and 1214 DF, p-value: <2.2e−16.
4. Statistical analysis of the results
In this section, the results of statistical analysis are shown with
the aim of drawing relevant conclusions for diagnosing rotor faults
induction motors with different types of power supply.
4.1. General analysis
First, boxplots are used to analyze and compare the collected
data. Boxplots are a useful visual tool to compare means and vari-
ability between different datasets. In these boxplots, the legend in
the x-axis is as follows: the first numeric value stands for the fault
severity or condition state (1 for a healthy motor; 2 for an incipient
fault; 3 for a half-broken bar; 4 for a full broken bar); next, it is
denoted the motor by M1 or M2; the additional letters are to indi-
cate the type of supply using the acronyms previously mentioned
in Section 3.
Fig. 2 shows the boxplots of the amplitude of the fault signatures
around the main harmonic (Fig. 2a and b) for both tested motors, the
four condition states including the three different supplies for each
motor and all motor operating conditions. Motor M1 data are more
spread out indicating a larger variance and causing an overlapping
of the boxes.
The graphic information given by the boxplots is completed with
some ANOVA models. In this analysis, the LSH and USH features are
taken as continuous response variables and “the type motor” and
“fault condition” as categorical explanatory variables. This analysis
demonstrated that the two categorical variables, “motor” and “con-
dition”, had a clear statistical significance in each ANOVA model
adjusted.
Table 3 shows the associated estimated coefficients resulting
from the ANOVA where the LSH feature was considered as response
variable and motor M1 and Condition 1 (healthy motor) as the ref-
erence levels. Since the associated p-values are much smaller than
0.01 and 0.001 (the standard cut-off values normally accepted), we
conclude that the following differences detected are statistically
significant: means of the LSH measures between motors M1 and
M2, and means corresponding to the conditions 2, 3 and 4 (repre-
senting different fault severity degrees) with respect to condition
1 (healthy motor).
Tukey’s honest significant differences method was also applied
as post hoc testing procedure to make a statistical comparison of the
means of the 4 condition states for the LSH and USH response vari-
ables. Tukey’s method provides a multiple testing correction when
comparing all these mean values (see [31] as reference for ANOVA
testing and post hoc multiple comparisons). Although the results of
applying this multiple comparison technique are not reported here,
all differences observed between pairs of means were statistically
significant in the analysis of the fault signatures. The p-values again
were very close to 0 and thus statistical significant differences were
detected among these 4 mean values.
The boxplots, ANOVA models and post hoc comparisons suggest
that these fault signatures may be adequate to discriminate among
4 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13

Fig. 1. Normalized PSD around main harmonic. Motor M1 with a full-broken bar. Supplies UT and TE.

the four condition states. Fig. 2 shows overlapping among the boxes,
which indicates that in some operating conditions it could be dif-
ficult to distinguish between adjacent fault severity conditions:
a healthy state and an incipient fault, between an incipient fault
and a semi broken bar, and between the semi broken and the full
broken bar. The use of additive models will allow us to overcome
some of these difficulties as it will be shown in Section 5.
Fig. 2 shows too, the boxplots corresponding to the sidebands
around the fifth (Fig. 2c and d) and seventh harmonic (Fig. 2e and f),
respectively. It can be inferred from these plots that the LSH around
the fifth harmonic is a useful fault signature to complement the
information provided by the signatures around the main harmonic.
However, this is not the case for the rest of fault signatures, at least,
when the data obtained in all the different operating conditions are
analyzed together.
It must be pointed out that this analysis used mixed data from
two different motors, six supplies and different load conditions. To
allow us to be more conclusive, the analysis has to consider the
influence of the different operating conditions. In Section 4.2, it
will be analyzed the influence of the supply and in Section 4.3 the
influence of the load.
4.2. Influence of the supply on the signature faults
In this section, data are segregated according to motor and sup-
ply with the aim to point out the diagnostic differences due to the
supply. The boxplots of the amplitude of the LSH and USH fault
signatures are shown in Fig. 3. Tables 4 and 5 present the results
of the ANOVA models adjusted for both motors. Now, the type of
power supply is included in the analysis as a variable. Several inter-
esting conclusions can be drawn from the analysis of these figure
and tables. The most obvious one is that results are not similar for
all combinations of motors and supplies. Each case studied will be
analyzed separately next.
 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13 5

Fig. 2. Boxplots of the amplitude of the broken bar signature faults LSH (left) and USH (right) around the main harmonic (top figure), fifth harmonic (middle figure) and
seventh harmonic, for motors M1 and M2, all types of supply and four condition states (denoted as 1 to 4, from healthy to full broken bar).

First, we focus on motor M1 and UT supply analysis results. It
is observed that the evolution of the fault severity can be clearly
traced by extracting the LSH and UHS fault signatures from the cur-
rent spectrum. An incipient fault is even easily observable because
there is no overlapping between the boxes of healthy motor and
incipient fault conditions. Nevertheless, this is not the case when
it comes to distinguish between half-broken bar and full broken
bar conditions. The peak amplitudes characterizing a half-broken
bar condition are so high (and therefore, so easily observable) that
they produce an overlapping with the box corresponding with a full
broken bar, making very difficult to distinguish these two condition
states.
When the motor is fed by the LN frequency inverter, the ampli-
tude of the fault peaks shows a behavior somehow similar to the
UT supply. It is remarkable that, when only the LSH is considered,
the ANOVA models of Table 6 show that the differences between
both supplies are not significant. But, this is not the case when the
USH is studied. In the case of the LSH study, there is an overlapping
between boxes corresponding to the condition states 1 and 2. But,
on the other hand, half-broken and complete broken bar condition
states can be clearly observed and distinguished. On the contrary,
if only the USH is considered, there is no overlapping between the
boxes of conditions states 1 and 2.
The condition monitoring is more challenging when the motor
M1 is fed by the TE inverter. This supply is characterized by a
high noise level in the current spectrum. While an incipient fault
is hardly distinguishable from a healthy motor condition, a half-
broken bar condition is observable but not as easily as for UT or
6 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13

a)

40

30
dB

20

10

1 .M 1 _ U T

2 .M 1 _ U T

3 .M 1 _ U T

4 .M 1 _ U T

1 .M 2 _ U T

2 .M 2 _ U T

3 .M 2 _ U T

4 .M 2 _ U T
1 .M 1 _ T E

2 .M 1 _ T E

3 .M 1 _ T E

4 .M 1 _ T E

1 .M 2 _ A B

2 .M 2 _ A B

3 .M 2 _ A B

4 .M 2 _ A B

1 .M 2 _ S M

2 .M 2 _ S M

3 .M 2 _ S M

4 .M 2 _ S M
1 .M 1 _ L N

2 .M 1 _ L N

3 .M 1 _ L N

4 .M 1 _ L N

b)
40

30
dB

20

10
1 .M 1 _ U T

2 .M 1 _ U T

3 .M 1 _ U T

4 .M 1 _ U T

1 .M 2 _ U T

2 .M 2 _ U T

3 .M 2 _ U T

4 .M 2 _ U T
1 .M 2 _ A B

2 .M 2 _ A B

3 .M 2 _ A B

4 .M 2 _ A B
1 .M 1 _ T E

2 .M 1 _ T E

3 .M 1 _ T E

4 .M 1 _ T E

1 .M 2 _ S M

2 .M 2 _ S M

3 .M 2 _ S M

4 .M 2 _ S M
1 .M 1 _ L N

2 .M 1 _ L N

3 .M 1 _ L N

4 .M 1 _ L N

Fig. 3. Boxplots of the amplitude of the broken bar signature fault (LSH: top figure, USH, bottom figure) around the main harmonic, for motors M1 and M2, six different
supplies and four condition states (denoted as 1 to 4, from healthy to full broken bar).

LN supplies. Furthermore, all condition states show lower fault sig- state 1 as reference, there is not significant evidence of interaction
nature amplitudes, and consequently, the fault detection is not as between the four condition states. Therefore, the two fault signa-
straightforward as it was with the other supplies. Therefore, fault tures are useful for condition monitoring purposes. In Table 5, the
detection is more difficulty when the motor is fed by this inverter. p-value for the condition state 2 is much bigger than for the other
Next, the results corresponding to the motor M2 analysis are condition states. This result means that it could be difficult to dis-
shown. The ANOVA models of Table 5 indicate that, taking condition tinguish between condition states 1 and 2. This is also observed
 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13 7

Table 4
Analysis of variance. Motor M1. LSH and USH.

LSH USH

Estimate Std. error t p-value Estimate Std. error t p-value

Intercept 10.2955 0.7514 13.701 <2e−16*** 12.5562 0.7720 16.265 <2e−16***

typeM1 LN 0.3798 0.5503 0.690 0.49 −2.6006 0.5654 −4.600 5.16e−06***
typeM1 TE −10.0824 0.4108 −24.541 <2e−16*** −10.3997 −0.4221 24.639 <2e−16***
Slip 114.9593 13.7517 8.360 4.33e−16*** 89.6634 14.1280 6.346 4.32e−10***
Condition 2 3.2075 0.4953 6.477 1.94e−10*** 4.5680 0.5088 8.978 <2e−16***
Condition 3 10.2642 0.4855 21.142 <2e−16*** 9.8418 0.4988 19.732 <2e−16***
Condition 4 15.8637 0.5193 30.547 <2e−16*** 15.0091 0.5335 28.132 <2e−16***

Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.
Residual standard error: 4.461 on 605 degrees of freedom.
Multiple R-squared: 0.7985, adjusted R-squared: 0.7965.
F-statistic: 399.5 on 6 and 605 DF, p-value: <2.2e−16.
Residual standard error: 4.584 on 605 degrees of freedom.
Multiple R-squared: 0.7579, adjusted R-squared: 0.7555.
F-statistic: 315.7 on 6 and 605 DF, p-value: <2.2e−16.

Table 5
Analysis of variance. Motor M2. LSH and USH.

LSH USH

Estimate Std. error t p-value Estimate Std. error t p-value

Intercept 6.6327 0.5401 12.280 <2e−16*** 5.8980 0.5724 10.304 <2e−16***

typeM2 AB −2.2867 0.3808 −6.005 3.32e−09*** −1.6869 0.4036 −4.180 3.35e−05***
typeM2 SM −0.1410 0.3882 −0.363 0.7166 0.3993 0.4114 0.971 0.332
Slip 101.6038 15.2920 6.644 6.85e−11*** 112.7421 16.2057 6.957 9.13e−12***
Condition 2 1.1622 0.4287 2.711 0.0069** 2.9176 0.4543 6.422 2.73e−10***
Condition 3 9.5110 0.4410 21.566 <2e−16*** 10.5088 0.4674 22.485 <2e−16***
Condition 4 11.9837 0.4300 27.871 <2e−16*** 13.1294 0.4557 28.814 <2e−16***

Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.
Residual standard error: 3.856 on 600 degrees of freedom.
Multiple R-squared: 0.663, adjusted R-squared: 0.6597.
F-statistic: 196.8 on 6 and 600 DF, p-value: <2.2e−16.
Residual standard error: 4.086 on 600 degrees of freedom.
Multiple R-squared: 0.6647, adjusted R-squared: 0.6614.
F-statistic: 198.3 on 6 and 600 DF, p-value: <2.2e−16.

in the boxplots of Fig. 3, where there is an overlapping between Table 6

Analysis of variance. Comparison of LSH versus the sum of the amplitudes of the
condition states 1 and 2.
peaks.
This M2 motor was also fed by three different voltage supplies
(UT, AB and SM inverters), which were also compared, consider- Res. Df RSS Df sum of Sq F p-value
ing the UT supply as the reference. The ANOVA results show that Motors LSH 1263 61,872
the LSH and USH fault signatures are significantly different when Condition LSH 1260 23,448 688.26 <2.2e−16
the motor is fed by the AB inverter compared to the UT voltage Motors LSH + USH 1263 230,558
Condition LSH + USH 1260 80,765 778.97 <2.2e−16
supply. This important result should be considered in the design
of a general-purpose condition monitoring strategy. Fault signa- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.
tures when the motor is fed by the SM inverter show a different
behavior. The ANOVA results provide high p-values for both fault 4.3. Influence of the slip
signatures, which proves that they have relatively similar distribu-
tions for both motor supplies (UT and SM). These different results As aforementioned, the motor was tested in a broad range of
could be explained by the different control strategies used by each load conditions in each group of tests. Hence, so far analyzed
inverter. results mix up all motor load conditions. The performed ANOVA
This subsection has shown the results obtained when different (Tables 4 and 5) shows the significant impact of slip in the dis-
motor supplies are considered. The type of voltage supply has a sub- tinction between condition states. The analysis is now performed
stantial effect on the amplitude of the fault signatures. It has been independently for different load conditions with the aim to deter-
observed that when the inverter has an independent adjustment of mine how the fault diagnosis could be affected by the slip or motor
voltage and frequency, the results are similar to the UT case. In the load. Fig. 4 shows the boxplots of the LSH (Fig. 4a) and the USH
case of sensorless vector control, the results are very different than (Fig. 4b) distributions for the six supplies and four condition states
for the rest of supplies. Two different situations have been observed considered in this study. But here, the datasets are divided in two
for the V/f linear control. In one of them (SM supply) the results are groups depending on the value of the slip. One group is labeled as
quite similar to the UT supply. Meanwhile, the condition monitor- “High” slip if its value is greater than 0.04. If this value is lower than
ing task would be more difficult if the motor is fed by the TE supply. 0.04, the other group is labeled as “Low” slip.
The reason is that this inverter produces a high noise level in the Again, the results are different depending on supply. Addition-
current spectrum. ally and not surprisingly, the boxes are less overlapped at high slips
8 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13

Fig. 4. Boxplot of the amplitude of the broken bar signature fault (LSH: top figure, USH, bottom figure) around the main harmonic, for motors M1 and M2, six different
supplies and four condition states (denoted as 1 to 4, from healthy to full broken bar). Two sets of tests are considered according to slip being higher or lower than 0.04
(denoted as “high” and “low”).
 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13
80
60
dB
40
20
1.M1
2.M1
3.M1
4.M1
Fig. 5. Boxplot of the sum of the amplitudes of the broken bar signature faults LSH and USH around the main harmonic, for both motors, six different supplies and four
condition states.
than at low ones. This means that fault detection should be easier
at high slips. More precisely, when motor M1 is fed by the UT sup-
ply and at low slips, it is difficult to distinguish between condition
states 3 and 4. However, boxes hardly overlap at high slips. Con-
clusions are analogous for conditions states 1 and 2 when the M2
motor is fed by the same supply.
Condition states 3 and 4 are more easily differentiated at high
slips when the motor is fed by the LN and TE inverters than when the
AB inverter is used to feed the motor. The incipient fault detection
is also easier at high slips than at low ones. When the SM inverter
is used as voltage supply, no significant differences are observed
between high and low slips.
4.4. LSH versus USH
The goal of this section is to study and compare the suitability
for condition monitoring of the LSH and USH fault signatures. In an
overall analysis and from a statistical point of view, the LSH and USH
distributions are quite similar not showing important differences.
Their variances are similar or, in some cases, larger for the USH.
A deeper analysis about the distinction between condition states
1 and 2 shows that the boxes of the USH fault signature have less
overlapping than the boxes of the LSH. Let us not forget that incipi-
ent fault detection is the most challenging situation for a condition
monitoring system.
All in all, it could be concluded that the USH is best suited as
a fault signature for condition monitoring purposes. Although the
influence of the motor and drive inertia should not be dismissed
since the LSH and USH have a strong dependence on the motor
plus load inertia. Rotor speed oscillations that give rise to the USH
are limited by the drive inertia. On the other hand, these fluctua-
tions reduce the amplitude of the LSH. Therefore, the LSH is higher
mainly in high inertia drives, but the USH may be predominant for
very low inertia ones [20,32,33] and this is the case of our labo-
ratory tests. Nevertheless, amplitude changes are similar for both
fault signatures, but in opposite directions and, consequently, the
sum of the amplitudes of LSH and USH remains quite constant
with the inertia [34]. The boxplots of the sum of the amplitudes
of the LSH and USH of both motors are shown in Fig. 5. Table 6
shows a comparison of the ANOVA of the LSH and the sum of
both fault harmonics. The results demonstrate that these fault sig-
natures are equally suited for detection purposes, but the sum
9
1.M2
2.M2
3.M2
4.M2
should be preferred if the influence of the drive inertia has to be
considered.
5. Additive models
The results analyzed in the previous sections show that there is
usually some overlapping between the boxplots of the fault sig-
natures. These box overlappings make the fault detection more
difficult. It is also harder to estimate the fault severity by a sim-
ple observation of the amplitude of the fault related harmonics.
Besides, the different operational conditions considered, character-
ized by the voltage supply, type of motor and variable motor load,
have been proved to have a significant impact on the amplitude of
the fault signatures. In this paper, it is proposed the use of additive
models to separate the effect of these factors on the amplitude of
the fault signatures with the aim of improving fault detectability
and severity estimation.
Additive models are useful data analytic tool to estimate
the effect of different variables in the response of a system by
using nonparametric regression techniques. They are based on the
approximation of a multivariate function f(x1 , . . ., xn ) as the sum of
univariate functions fj (xj ) [35].
The estimation of the individual terms explains how the
dependent variable changes with the corresponding independent
variables and which variables do not have a significant impact in
the response. Therefore, additive models can be seen as a general-
ization of multiple regression models, which maintain the additive
nature of the models, but it is allowed to replace simple linear terms
as predictors by unspecified (non-parametric) functions that can be
estimated through “backfitting” algorithms. These models are flex-
ible and easy to implement, and once the model has been identified,
it is easy to assess the influence of each primary variable xj on f(x1 ,
. . ., xn ) just by plotting fj against xj [36]. They have been previously
used to solve problems such the design of heat transfers [37], for the
planning of distributed generation [38], to obtain optimum work-
ing conditions of heat-pumps [39] and to forecast the prices in the
electric market [40,41].
In this work, additive models have been fitted using the R
free software environment for statistical computing and the “gam”
package available at the CRAN repository [42]. Each response vari-
able is modeled as a sum of a common overall mean value , plus
the effects of the type of motor and voltage supply Ti , the effect of
the motor load as a continuous function of the value s, plus the fault
10 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13

Fig. 6. Additive model for the amplitude of LSH (top figure) and USH (bottom figure) around the main harmonic, for both motors, six different supplies and four condition
states.

condition state Cj where j = 1, 2, 3 and 4, plus the sum of a random
error term ε. That is:
Dependent variable = m + Ti + s(slip) + Cj + ε
where the dependent variable can be the LSH, the USH or LSH + USH.
Twice-standard-error confident limits and bands have been also
included for each considered effect in the additive models. Notice
that this fitted model allows understanding whether it is possible to
detect differences among the four considered motor fault condition
states after controlling the differences in type of motor/supply or in
the slip. Moreover, in future developments, these estimated effects
can be taken into account and subtracted in order to achieve more
accurate classifications in case of an unknown condition state.
Fig. 6 shows the additive models for the LSH (Fig. 6a) and the
USH (Fig. 6b). Comparing both Figs., no significant differences exist
apparently between both peaks. The additive models calculated
confirm the ANOVA and Boxplots results analyzed in the previous
sections. The models show that the three variables are relevant and
should be taken into account to develop a condition monitoring
expert system. The voltage supply used to feed the motor has an
important influence in determining the amplitude of the fault sig-
natures. As it was observed in the previous sections, some supplies
lead to similar results: (i) LN and UT supplies in the case of motor
M1; (ii) SM and UT supplies in the case of motor M2.
 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13 11

M1_LN M1_TE M1_UT 1 2 3 4

10

10

10
5

5
partial for condition
partial for type

s(slip)
0

0
-5

-5

-5
-10

-10

-10
0.02 0.04 0.06 0.08
type condition
slip

a)
M2_AB M2_SM M2_UT 1 2 3 4
10

10

10
5

5
partial for condition
partial for type

s(slip)
0

0
-5

-5

-5
-10

-10

-10

0.01 0.02 0.03 0.04 0.05 0.06

type condition
slip

b)
Fig. 7. Additive model for motor M1 (top figure) and M2 (bottom figure). Independent variable: amplitude of the USH around the main harmonic. Dependent variables: type
of supply, slip and condition state.

As it is observed in the middle graphic of Fig. 6, the model
shows a quasi-linear dependence with the slip, especially for slips
between 2 and 6%.
An important result can be observed in the right subplots of
Fig. 6, where it is plotted the contribution of the fault severity to
the amplitude of the fault signatures. It can be considered that it is
represented the fault signature after subtracting the contribution
of the type of motor and supply (left subplot) and the contribu-
tion of the slip (central subplot). Now, the four conditions states
are clearly distinguished. There is no overlapping between the four
boxes. The additive models allow us to point out the contribution
of the fault severity and to separate and eliminate the influence of
the other factors. This way, the performance of a diagnostic or con-
dition monitoring system (some classifier, for example) would be
significantly improved.
Fig. 7 shows the results of the additive model of the USH consid-
ering separately the data of each motor (the models for LSH and for
LSH + USH are very similar). One of the most important results is
that the influence of TE inverter is clearly differentiated from the
others voltage supplies. This inverter had a high noise level in the
current spectrum.
Fig. 8 is also very interesting. It shows the additive model results
of the case when the UT supply was used to feed both motors. Again,
in the right subplot, the contribution of the four fault severities to
12 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13
Fig. 8. Additive model for the amplitude of the USH around the main harmonic, for motors M1 and M2 fed by the utility supply, UT.
the fault signature is completely distinguishable. The left subplot
represents the contribution of the type of motor (labeled as type
in the figure). It is important to notice that each motor contributes
differently to the amplitude of the fault signature.
Thus, the additive models demonstrate the influence of each
factor in the value of the fault signature. The most important result
is that the additive model allows us to highlight the fault severity
contribution to the dependent variable (the amplitude of the fault
signature). This is easily observed in the right subplots in all the
figures. This variable, Cj , can be used as input data to a classifier
trained for the diagnosing of the motor. The performance of the
classifier would be greatly improved comparing to the case where
it is trained with the raw data obtained from the spectra.
6. Conclusions
The main goal of this paper was the use of statistical tools to
study if some fault signatures extracted from the stator current
were suitable for diagnosing rotor faults in induction motors in
steady state operation. A case study is presented where a rotor bro-
ken bar fault evolution has been tested and analyzed. The fault is
simulated by a progressive drilling of a hole in one of the rotor bars
in two different induction motors. These motors were fed by differ-
ent supplies: utility supply and four different frequency inverters
with different control strategies. The motors were also tested at
different loads, covering from low load to full load.
This fault produces a pair of bands around the main harmonic.
The amplitudes of these two harmonics are considered as fault sig-
natures. The results demonstrate statistically that these data are
suitable for performing a condition monitoring task. It was also
demonstrated the importance and significance on the fault diagno-
sis of the following variables: type of motor, voltage supply and the
motor load.
One of the most important results of this analysis is that the type
of voltage supply, used to feed the motor, has a substantial impact
on the detectability of the fault and the subsequent diagnosis. This
makes the detection of the fault at an early stage especially difficult
when the motor is fed by some inverters.
All the statistical analyses have also proven that the USH around
the main harmonic is the best suited as a fault signature. In fact, in
most of the cases both fault signatures are equally valid, but, in the
most challenging situations, the USH fault signature has proved
to be the most adequate. In our laboratory setup the motor-load
system was characterized by a low inertia. Nevertheless, the sum
of both sidebands is also valid as fault signature and it is mostly
independent of the inertia.
The obtained results also show that a universal diagnosis
method for an induction motor must consider the information
about the type of motor, the type of supply, the load level or the
slip. On the other hand and for the presented cases, these results
offer guidance about which fault signatures are more suitable for
diagnosing and if the slip or load level must be taken into account.
The most important result of this study is that the same fault
severity produces different fault signature amplitudes depending
on the type of the motor, supply and load level. This is crucial in the
construction of a diagnosis procedure for different motors, supplies
or motors operating under a wide range of load levels. It has been
proven that the additive models allow us to quantify the influence
of each of these variables in the amplitude of the fault signatures.
These results confirm the conclusions obtained with the ANOVA
analysis. They can be very useful for diagnosis purposes allowing
the subtraction of the contribution of a variable in the fault signa-
ture value. So, the bias produced by any of these variables in the
amplitude of the fault signature can be avoided when constructing
a universal diagnosis procedure or system.
Acknowledgements
García-Escudero’s research was partially supported by the
Spanish Ministerio de Ciencia y Tecnología, grant MTM2011-
28657-C02-01, by Consejería de Educación y Cultura de la JCyL,
grant VA212U130.
References
[1] Y. Oguz, M. Dede, Speed estimation of vector controlled squirrel cage asyn-
chronous motor with artificial neural networks, Energy Convers. Manag. 52
(2011) 675–686.
 O. Duque-Perez et al. / Electric Power Systems Research 121 (2015) 1–13
[2] J. Faiz, B.M. Ebrahimi, Locating rotor broken bars in induction motors using
finite element method, Energy Convers. Manag. 50 (2009) 125–131.
[3] M.R. Mehrjou, N. Mariun, M.H. Marhaban, N. Misron, Rotor fault condition mon-
itoring techniques for squirrel-cage induction machine – a review, Mech. Syst.
Signal Process. 25 (2011) 2827–2848.
[4] H. Keskesa, A. Brahama, Z. Lachiri, Broken rotor bar diagnosis in induction
machines through stationary wavelet packet transform and multiclass wavelet
SVM, Electr. Power Syst. Res. 97 (2013) 151–157.
[5] G.K. Singh, S.A.S. Al Kazzaz, Induction machine drive condition monitoring and
diagnostic research – a survey, Electr. Power Syst. Res. 64 (2003) 145–158.
[6] A.G. Garcia-Ramirez, L.A. Morales-Hernandez, R. Osornio-Rios, J.P. Benitez-
Rangela, A. Garcia-Perez, R.J. Romero-Troncoso, Fault detection in induction
motors and the impact on the kinematic chain through thermographic analysis,
Electr. Power Syst. Res. 114 (2014) 1–9.
[7] J. Cusidó, L. Romeral, J.A. Ortega, A. Garcia, J.R. Riba, Wavelet and PDD as fault
detection techniques, Electr. Power Syst. Res. 80 (2010) 915–924.
[8] M.E.H. Benbouzid, A review of induction motors signature analysis as a medium
for faults detection, IEEE Trans. Ind. Electron. 47 (2000) 984–993.
[9] M. Pineda-Sanchez, M. Riera-Guasp, J. Perez-Cruz, R. Puche-Panadero, Tran-
sient motor current signature analysis via modulus of the continuous complex
wavelet: a pattern approach, Energy Convers. Manag. 73 (2013) 26–36.
[10] F. Briz, M.W. Degner, P. Garcia, A.B. Diez, High-frequency carrier-signal voltage
selection for stator winding fault diagnosis in inverter-fed ac machines, IEEE
Trans. Ind. Electron. 55 (2008) 4181–4190.
[11] R. Wieser, C. Kral, F. Pirker, M. Schagginger, On-line rotor cage monitoring of
inverter-fed induction machines by means of an improved method, IEEE Trans.
Power Electron. 14 (1999) 858–865.
[12] D.F. Pires, V.F. Pires, J.F. Martins, A.J. Pires, Rotor cage fault diagnosis in three-
phase induction motors based on a current and virtual flux approach, Energy
Convers. Manag. 50 (2009) 1026–1032.
[13] B. Akin, U. Orguner, H.A. Toliyat, M. Rayner, Low order PWM inverter harmonics
contributions to the inverter-fed induction machine fault diagnosis, IEEE Trans.
Ind. Electron. 55 (2008) 610–619.
[14] G.R. Bossio, C.H.D. Angelo, G.O. Garcia, J.A. Solsona, M.I. Valla, Effects of rotor
bar and end-ring faults over the signals of a position estimation strategy for
induction motors, IEEE Trans. Ind. Appl. 41 (2005) 1005–1012.
[15] M.F. Cabanas, F. Pedrayes, C.H. Rojas, M.G. Melero, J.G. Norniella, G.A. Orcajo,
J.M. Cano, F. Nuño, D.R. Fuentes, A new portable, self-powered, and wireless
instrument for the early detection of broken rotor bars in induction motors,
IEEE Trans. Ind. Electron. 58 (2011) 4917–4929.
[16] C. Bruzzese, Analysis and application of particular current signatures (symp-
toms) for cage monitoring in nonsinusoidally fed motors with high rejection
to drive load, inertia, and frequency variations, IEEE Trans. Ind. Electron. 55
(2008) 4137–4155.
[17] D. Morinigo-Sotelo, L.A. Garcia-Escudero, O. Duque-Perez, M. Perez-Alonso,
Practical aspects of mixed eccentricity detection in PWM voltage source
inverter fed induction motors, IEEE Trans. Ind. Electron. 57 (2010)
252–262.
[18] I.P. Georgakopoulos, E.D. Mitronikas, A.N. Safacas, Detection of induction motor
faults in inverter drives using inverter input current analysis, IEEE Trans. Ind.
Electron. 58 (2011) 4365–4373.
[19] M. Akar, Detection of a static eccentricity fault in a closed loop driven induction
motor by using the angular domain order tracking analysis method, Mech. Syst.
Signal Process. 34 (2013) 173–182.
[20] A. Bellini, F. Filippetti, C. Tassoni, G.A. Capolino, Advances in diagnostic
techniques for induction machines, IEEE Trans. Ind. Electron. 55 (2008)
4109–4126.
[21] B. Payne, A. Ball, F. Gu, Detection and diagnosis of induction motor faults using
statistical measures, Proc. Int. J. COMADEM 5 (2002) 5–19.
13
[22] H. Çalıs, A. Çakır, Experimental study for sensorless broken bar detection in
induction motors, Energy Convers. Manag. 49 (2008) 854–862.
[23] K. Drobnic, M. Nemec, R. Fiser, V. Ambrozic, Simplified detection of broken rotor
bars in induction motors controlled in field reference frame, Control Eng. Pract.
20 (2012) 761–769.
[24] L.M. Rabelo Baccarini, J.P. Braga Tavares, B. Rodrigues de Menezes, W.M. Camin-
has, Sliding mode observer for on-line broken rotor bar detection, Electr. Power
Syst. Res. 80 (2010) 1089–1095.
[25] I. Aydin, M. Karakose, E. Akin, A new method for early fault detection and
diagnosis of broken rotor bars, Energy Convers. Manag. 52 (2011) 1790–1799.
[26] F. Filippetti, G. Franceschini, C. Tassoni, P. Vas, Impact of speed ripple on rotor
fault diagnosis of induction machines, in: Proc. International Conference of
Electric Machines and Drives, 1994, pp. 452–456.
[27] O. Duque-Perez, D. Morinigo-Sotelo, A. Rodriguez, L.A. Garcia-Escudero, M.
Perez-Alonso, Experimental identification of induction motor broken bar char-
acteristic frequencies using a programmable power source, in: Proc. 2010 19th
International Conference on Electrical Machines, 2010.
[28] S.R.H. Amrei, X. Dian, A new study on harmonics and interharmonics reduction
with frequency selective closed-loop control in VSI inverters, in: Vehicle Power
and Propulsion Conference, 2008, pp. 1–6, VPPC ‘08.
[29] J.H. Jung, J.J. Lee, B.H. Kwon, Online diagnosis of induction motors using MCSA,
IEEE Trans. Ind. Electron. 53 (2006) 1842–1852.
[30] L.A. García-Escudero, O. Duque-Perez, D. Morinigo-Sotelo, M. Perez-Alonso,
Robust condition monitoring for early detection of broken rotor bars in induc-
tion motors, Expert Syst. Appl. 38 (2011) 2653–2660.
[31] R. Christensen, Analysis of Variance, Design, and Regression: Applied Statistical
Methods, Chapman and Hall, London, 1996.
[32] C. Kral, F. Pirker, G. Pascoli, The impact of inertia on rotor fault effects – theo-
retical aspects of the Vienna monitoring method, IEEE Trans. Power Electron.
23 (2008) 2136–2142.
[33] F. Filippetti, G. Franceschini, C. Tassoni, P. Vas, AI techniques in induction
machines diagnosis including the speed ripple effect, IEEE Trans. Ind. Appl.
34 (1998) 98–108.
[34] A. Bellini, C. Concari, G. Franceschini, E. Lorenzani, C. Tassoni, A. Toscani,
Thorough understanding and experimental validation of current sideband
components in induction machines rotor monitoring, IEEE 32nd Annual Con-
ference on Industrial Electronics, IECON (2006) 4957–4962.
[35] T. Hastie, R. Tibshirani, Generalized Additive Models, Chapman and Hall, Lon-
don, 1990.
[36] G. De Nicolao, R. Scattolini, C. Siviero, Modelling the volumetric efficiency of ic
engines: parametric, non-parametric and neural techniques, Control Eng. Pract.
4 (1996) 1405–1415.
[37] B. Sahin, A. Demir, Thermal performance analysis and optimum design param-
eters of heat exchanger having perforated pin fins, the planning of distributed
generation, Energy Convers. Manag. 49 (2008) 1684–1695.
[38] M.R. AlRashidi, M.F. AlHajri, Optimal planning of multiple distributed genera-
tion sources in distribution networks: a new approach, Energy Convers. Manag.
52 (2011) 3301–3308.
[39] O. Çomakli, C. Çelik, S. Erdogan, Determination of optimum working condi-
tions in heat-pumps using nonazeotropic refrigerant mixtures, Energy Convers.
Manag. 40 (1999) 193–203.
[40] R.J. Bessa, M.A. Matos, Global against divided optimization for the participation
of an EV aggregator in the day-ahead electricity market. Part I: theory, Electr.
Power Syst. Res. 95 (2013) 309–318.
[41] R.J. Bessa, M.A. Matos, Optimization models for an EV aggregator selling sec-
ondary reserve in the electricity market, Electr. Power Syst. Res. 106 (2014)
36–50.
[42] R Development Core Team R, A language and environment for statistical com-
puting, R Foundation for Statistical, 2009.