See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/310159381

Modeling Approach and Fault Index Analysis of a Voltage-Source Brushless

DC Motor

Conference Paper · September 2016

DOI: 10.1109/SYSTOL.2016.7739807

CITATIONS READS

6 527

4 authors:

Kawthar Alameh Riham Ginzarly

ALTRAN France - Usine Renault Cléon Beirut Arab University
12 PUBLICATIONS   115 CITATIONS    12 PUBLICATIONS   46 CITATIONS   

SEE PROFILE SEE PROFILE

Ghaleb Hoblos Georges Barakat

ESIGELEC Université du Havre
105 PUBLICATIONS   881 CITATIONS    161 PUBLICATIONS   2,660 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Developing prognostics and health management approaches for remaining useful life forecasting based on stochastic and zonotopic estimation and prediction
techniques with applications to degraded power electronic systems in an LPV framework View project

Fault detection and diagnosis in Brushless DC motors based on vibration signal analysis View project

All content following this page was uploaded by Kawthar Alameh on 23 October 2017.

The user has requested enhancement of the downloaded file.

 Modeling Approach and Fault Index Analysis of a Voltage-Source
Brushless DC Motor
Kawthar Alameh1, Student Member, IEEE, Riham Ginzarly1, Ghaleb Hoblos1, Georges Barakat2

Abstract— This paper presents a modeling approach and a interest [5][6]. The acquired signals can be thermal, magnetic
fault index analysis of a voltage-source brushless DC motor. (magnetic field, flux density...), electrical (currents,
First, an analytical multi-physical model of the inverter-fed voltages,...) or mechanical (torque, vibration…) [7]. Among
Permanent Magnet DC motor is developed and simulated using these signals, vibration analysis has been one of the most
Matlab/Simulink. This model enables the generation of attractive and successful techniques for health monitoring of
electrical, magnetic and vibration signals under healthy and
rotating machines, as stated in [8][9].
faulty motor behaviors, with several fault categories and
severities. Before simulating faulty conditions in the different The crucial step in a signal-based FDI approach is the
parts of the analytical model, they are compared to Finite analysis of the process data in one or more informative
Element models, developed using Matlab for air-gap flux density domains (time, frequency...) to extract discriminative features
and ANSYS software for stator natural frequency calculations. sensitive to the fault occurrence, type and severity [5]. For this
Simulation results of the motor during normal functioning and purpose, the presented work focuses on the extraction of time,
under faulty conditions are presented. In particular, rotor space, frequency and harmonic indicators from EM vibration
eccentricity and single pole demagnetization faults are studied in signals generated by the motor, and to analyze them with
this paper. Then, different features, including time-, space-, respect to the operating condition and fault severity.
frequency- and harmonic- domain characteristics, are extracted
The main contribution of this paper is:
from vibration signals for different cases. Finally, these
indicators are analyzed with respect to the fault severity to select  The design of an analytical model of an inverter-fed
the most discriminative one(s) allowing an efficient fault BLDC motor combining the electrical, magnetic and
detection and isolation. vibratory behaviors of the PM motor
 The use of the stator vibration signal as an alternative
I. INTRODUCTION
solution for FDI purpose, when the electrical and

I n recent years, Brushless DC Motors (BLDCMs), have

been attractive candidates for several drive applications
such as hybrid electric vehicles (HEV), due to their high
magnetic signals present some limitations in fault
detection issues
This paper is structured as follow. A Matlab/Simulink
efficiency, important torque-to-inertia and torque–to-volume analytical model of the BLDCM is presented in section II. The
ratios and high dynamic response performance. EM and mechanical parts of the model are then compared to
In some applications, BLDCMs operates in severe finite element (FE) models using Matlab and ANSYS. In
environment conditions like high temperature levels, high section III, the two studied faults REF and DMF are simulated
stator currents and strong vibrations [1]. Consequently, in the analytical model, by changing some of its parameters.
several kinds of faults are unavoidable in such motors and can The simulation results, obtained for different cases, are
occur in the rotor, stator, voltage/current inverter, or in the presented in section IV. In section V, different indicators,
connected mechanical components [2]. As mentioned in [1], extracted from vibration signals generated with different
the main faults in permanent magnet (PM) motors can be faults, are analyzed with respect to fault type and severity, for
classified, into: electromagnetic (EM), like stator-winding fault detection and diagnosis purposes. Finally, conclusion
short-circuits and PM demagnetization (DMF), and and future perspectives are formulated in the last section.
mechanical faults, like rotor eccentricity (REF) and bearing
II. ANALYTICAL MODEL OF BLDCM
damages [3]. This paper discusses machine faults, mainly
REF and DMF, without focusing on the inverter and bearings. The inverter drive system for a BLDCM is shown in Fig.1.
In order to detect these failures at early stages, many It consists of a DC voltage source, a three-phase inverter and
detecting and monitoring approaches have been developed so a surface-mounted PM synchronous motor (SM-PMSM).
far [2][4]. A fault detection and isolation (FDI) system is a To rotate the BLDC motor continuously, the stator
monitoring approach used to detect and then characterize windings should be energized in a proper sequence, based on
faults in a process. FDI approaches are, mainly, classified into the current rotor position [10]. Three-phase Hall sensors are
three categories, namely qualitative model-based, quantitative usually used to detect the rotor position information. A
model-based and signal-based classes [5]. In fact, for complex Matlab/Simulink block is modeled to generate the three-
and large-scale process plants as in the case of electrical sensor signals from the back-emf waveforms [11]. The
machines, signal-based condition monitoring is of particular switching logic of the inverter transistors is set, in the Gate-
1 2
K. Alameh, R. Ginzarly and G. Hoblos are with IRSEEM/ESIGELEC, G. Barakat is with the Electrical Engineering Department, GREAH,
76801, Saint Etienne du Rouvray, France (e-mail: University of Le Havre, 76600, Le Havre, France (e-mail:
k.kawthar.Alameh@ieee.org, rmginzarly@gmail.com and Georges.Barakat@univ-lehavre.fr).
Ghaleb.Hoblos@esigelec.fr)
signal generation block, to follow the 180°-conduction mode. Maxwell force is assumed to be the only source of EM
This section discusses the analytical multi-physical model of vibration and acoustic noise radiation.
the SM-PMSM.
𝑃𝑟𝑎𝑑 (𝜃𝑠 , 𝑡) = 𝐵𝑟𝑎𝑑 (𝜃𝑠 , 𝑡)2 ⁄2µ0 (5)
Gate-signal Hall-signal
generation generation FE calculation of 𝐵𝑟𝑎𝑑 using Matlab
FE analysis (FEA) is a very popular numerical method used
stator currents in the field of electrical machine analysis. It uses Laplace and
Electrical model EM torque
Poisson’s equations to describe the machine and can perfectly
Three-phase
back-emf analyze PMs of any shape and material [15]. To compare the
six-step
voltage Magnetic model air-gap flux density analytical calculation, a FEA has been performed to calculate
inverter
the vector potential 𝐴𝑧,𝑖 at each point 𝑋𝑖 in the laminated sheet
vibration signals
Vibratory model of the machine, as detailed in [16]. Fig. 2 gives a comparison
between 𝐵𝑟𝑎𝑑 calculated by the FEA and analytical modeling.
SM-PMSM multi-physical model As shown in this figure, the results obtained by the two
models are almost similar and this could validate the proposed
Figure 1: Inverter-drive system of the Brushless DC motor
analytical approach for the 𝐵𝑟𝑎𝑑 calculation.
Fault-free electrical model of the stator 1

Air-gap Flux Density due to rotor PMs, (T)

0.8
For a three-phase star-connected stator system without a
0.6
neutral return, the fault-free dynamic model is expressed by
(1), (2) and (3), where 𝑉𝐴𝐵,𝐵𝐶 are phase-to-phase voltages (V)
0.4
0.2
at the inverter output, 𝑖𝐴,𝐵,𝐶 are motor phase currents (A), 0
𝑒𝐴,𝐵,𝐶 are back-emf voltages induced in stator phases (V), 𝑅𝑆 -0.2
and 𝐿𝑆 are motor winding resistance (Ω) and inductance (H). -0.4 FEA model
Analytical model

𝑉𝐴𝐵 = 𝑅𝑆 (𝑖𝐴 − 𝑖𝐵 ) + 𝐿𝑆 𝑑(𝑖𝐴 − 𝑖𝐵 )⁄𝑑𝑡 + (𝑒𝐴 − 𝑒𝐵 )

-0.6
(1)
-0.8

𝑉𝐵𝐶 = 𝑅𝑆 (𝑖𝐵 − 𝑖𝐶 ) + 𝐿𝑆 𝑑(𝑖𝐵 − 𝑖𝐶 )⁄𝑑𝑡 + (𝑒𝐵 − 𝑒𝐶 ) (2) -1

0 5 10 15 20 25 30 35 40 45 50 55 60
Angle, (mech. degrees)
𝑖𝐶 (𝑡) = −(𝑖𝐴 (𝑡) + 𝑖𝐵 (𝑡)) (3)
Figure 2: Comparison of the air-Gap flux density due to rotor PMs for a
The three-phase back-emf 𝑒𝐴,𝐵,𝐶 are calculated, using the healthy case, by analytical and FE approaches
Faraday’s induction law, as the time derivative of the open-
circuit rotor flux induced in the stator phase windings. Vibratory Model of the stator structure
The motor internal EM torque 𝑇𝐸𝑀 (N.m) is expressed by The vibratory model computes the stator mechanical
(4), with 𝛺𝑚 is the rotor speed (rad/sec). parameters and vibration signal, due to its excitation by the
𝑇𝐸𝑀 = (𝑒𝐴 × 𝑖𝐴 + 𝑒𝐵 × 𝑖𝐵 + 𝑒𝐶 × 𝑖𝐶 )⁄𝛺𝑚 (4) EM pressure 𝑃𝑟𝑎𝑑 . The analytical stator natural frequencies
are compared to those obtained by a FE model developed
In this model, an inter-turn short-circuit fault can be using ANSYS software.
represented by an additional resistor 𝑅𝑓 in parallel with the Analytical vibratory model
shorted turns of the faulty phase [1] and consequently, an
extra electrical equation is added to the above healthy ones. The crucial step in the prediction of the vibro-acoustic
emissions of any structure is the determination of its natural
EM Model of the motor frequencies 𝑓𝑚 using: analytical models, FEA or experimental
This section presents the analytical and FE models of the measurements. Due to their fast computation speed, the
EM part of the motor. analytical approaches are generally preferred in multi-physics
modeling issues. In [17], the most encountered 2D and 3D
Analytical EM model
analytical models used for 𝑓𝑚 calculation and their limitations
This model calculates the radial flux density in the air-gap have been detailed. In this paper, an equivalent 2D ring model
(𝐵𝑟𝑎𝑑 ) and the radial pressure applied to the inner surface of is used to calculate 𝑓𝑚 of the stator structure, taking into
the stator (𝑃𝑟𝑎𝑑 ). 𝐵𝑟𝑎𝑑 is defined as the product of the air-gap account the teeth and windings as additional masses to the
permeance [12] and the total magnetomotive force, created by core, as proposed in [12].
the stator (𝑀𝑀𝐹𝑠 ) and the rotor (𝑀𝑀𝐹𝑟 ), in both space (𝜃) and The radial dynamic deflections 𝑋𝑑,𝑚𝑓 of the stator are
time (t) domains. The 𝑀𝑀𝐹𝑠 is expressed in the stator-space calculated by (6), based on Hooke’s law and by means of a
(𝜃𝑠 ) domain in function of the instantaneous stator currents complex magnification factor 𝜂𝑚 (𝑓) given in [12], where 𝑅𝑠 ,
[13]. The 𝑀𝑀𝐹𝑟 is represented by a rectangular function [14] 𝐿𝑠𝑐 , 𝐾𝑚 and 𝑃𝑟𝑎𝑑,𝑚𝑓 are the stator inner radius and active
in the rotor-space (𝜃𝑟 ) depending on the PM characteristics.
length, lumped stiffness and complex magnitudes of the 2D
According to [9], the tangential component of the flux density
FFT of 𝑃𝑟𝑎𝑑 respectively. Finally, the 2D vibration
can be neglected and 𝑃𝑟𝑎𝑑 is calculated as given in (5), where
displacement 𝑋(𝑡, 𝜃𝑠 ) is obtained by a 2D inverse fast Fourier
µ0 is the air magnetic permeability. The calculated radial
transform (FFT) of the calculated 𝑋𝑑,𝑚𝑓 .
 rotor and the stator and it is generally classified into three
𝑋𝑑,𝑚𝑓 = (2𝜋𝑅𝑠 𝐿𝑠𝑐 𝑃𝑟𝑎𝑑,𝑚𝑓 ⁄𝐾𝑚 ) × 𝜂𝑚 (𝑓) (6)
types: static (SE), dynamic (DE) and mixed eccentricity
Natural frequencies calculation using ANSYS (ME). The distinction between SE and DE depends on
whether the air-gap length is varying with time or not [1]; the
A 2D FE model is achieved using ANSYS to validate the ME is a combination between SE and DE.
analytical formulas used to calculate the natural frequencies This fault is modeled by changing the smoothed air-gap
of the stator. Only the stator core, represented by an length “g” in the calculation of the air-gap permeance as in
equivalent ring, and stator teeth are considered in this model. [19]. Moreover, the transformation from 𝜃𝑟 to 𝜃𝑠 is modified
The stator windings are modeled as additional masses to the as expressed in (9), (10) and (11), respectively, for healthy,
teeth and the rotor is assumed as a negligible noise source SE and DE cases, where 𝜀𝑠 , 𝜀𝑑 , 𝜃𝑠0 , 𝜃𝑑0 are, respectively, the
[18]. Teeth and core are laminated iron with an equivalent static and dynamic degrees and initial positions.
modulus of elasticity 𝐸𝑐 = 210 Pa, copper and insulation give
an equivalent modulus of elasticity 𝐸𝑤 = 9.4 Pa. The modeling 𝜃𝑟 = 𝜃𝑠 − Ω𝑚 𝑡 (9)
error (𝛿𝑟 ), between the analytical and numerical approaches
𝜀𝑠 𝑔
and calculated by (7), is given in Table I. Based on this 𝜃𝑟 = cos−1 (𝐾 (cos 𝜃𝑠 − cos 𝜃𝑠0 )) − Ω𝑚 𝑡
𝑅𝑠
comparison, it is shown that the analytical 2D ring approach (10)
can be effectively used for stator natural frequencies 𝜀 𝑔 𝜀𝑠 𝑔
with 𝐾 = 1⁄√1 + ( 𝑅𝑠 )2 − 2 𝑅𝑠
cos(𝜃𝑠 − 𝜃𝑠0 )
calculation. 𝑠

𝛿𝑟 (%) = |𝑓𝑚 𝑎𝑛𝑎𝑙𝑦𝑡𝑖𝑞𝑢𝑒 − 𝑓𝑚 𝐹𝐸𝐴 |⁄𝑓𝑚 𝐹𝐸𝐴 × 100 (7) 𝜀𝑑 𝑔

𝜃𝑟 = cos−1 (𝐾′ (cos 𝜃𝑠 − cos(Ω𝑚 𝑡 + 𝜃𝑑0 ))) − Ω𝑚 𝑡
𝑅𝑠
(11)
𝜀 𝑔 𝜀𝑑 𝑔
Table I: Errors of the analytical method compared to FEA results for with 𝐾′ = 1⁄√1 + ( 𝑅𝑑 )2 − 2 𝑅𝑠
cos(𝜃𝑠 − Ω𝑚 𝑡 − 𝜃𝑑0 )
natural frequency calculation 𝑠

Mode FEA (Hz) Analy (Hz) Error (%) IV. SIMULATION RESULTS
0 3410,06 3346 1,87 The 36-slot 12-pole BLDCM, proposed in [20], is used as
2 165,93 155,77 6,12 a reference motor. It is driven by a 180°-commutated voltage
3 420,13 403,6 3,94 inverter with a voltage DC bus 𝑉𝐷𝐶 = 490 V at the nominal
4 715,98 697,3 2,61 speed 𝑁𝑛 = 668 tr/min. SE and DE faults with different
degrees (𝜖𝑠 and 𝜖𝑑 = 10, 25, 37, 45, 55, 67, 80, 93 %), partial
III. FAULTS MODELING IN THE ANALYTICAL MODEL
PM cracks in south and north poles with various severities (𝜖𝑝
The first step toward a data-based fault diagnosis is to = 10, 25, 45, 55, 67, 80, 93, 100 %) are simulated in the
simulate the faults in the model and then, to analyze their analytical model. Then, their effects are studied on different
effects on the analyzed signals. In this section, the REF and signals (current, vibration…).
DMF are modeled. Due to the machine symmetry, the effects of REF and DMF
DMF in the rotor PM on the three stator phases are almost the same. Fig. 3 gives
current waveforms in phase A, in the time (a) and the
The rotor PMs of a SM-PMSM could be demagnetized due frequency (b) domains, for a healthy, SE, DE, south and north
to high temperature levels, stator MMF or even by cracks pole cracks, cases. Based on this comparison, it is shown that
produced in rotor PMs during manufacturing [1]. the studied faults have no significant effects on the current
In the magnetic model, a crack is created in a rotor pole to signals, due to the winding configuration; so they cannot be
imitate the DMF. It is introduced by removing certain PM effectively detected by the conventional motor current
elements using an analytical signal (β(𝜃𝑟 )) in the 𝜃𝑟 -domain signature analysis (MCSA). As well, the EM torque has been
added to the 𝑀𝑀𝐹𝑟 . The mathematical equation of β(𝜃𝑟 ), studied under different cases. Fig.4 gives the 𝑇𝐸𝑀 waveforms
given in (8), depends on the angular position [𝜃j0 : 𝜃j1 ] in (rad) for a healthy, 45% SE, 45% DE and 45% crack, cases.
and the depth 𝜀𝑝 in (%) of the crack in the faulty pole “j”. 50
Healthy
40 10
1 45% SE

β(𝜃𝑟 ) = 𝐴0 + ∑ 𝐴𝑐,𝑚 cos(𝑚𝜃𝑟 ) + 𝐴𝑠,𝑚 sin(𝑚𝜃𝑟 )

45% DE
30 45% PDMF
1
0 45% PDMF
Phase A Current (A)

20 10 2
𝑚
Magnitude (A)

10
𝑗
𝐴0 = ((−1) 𝜀𝑝 𝑀𝑀𝐹𝑟,𝑚𝑎𝑥 (𝜃j1 − 𝜃j0 ))⁄(2𝜋) 0
10
-1

𝐴𝑐,𝑚 = (8) -10

10
-2

-20
(−1)𝑗 (𝜀𝑝 𝑀𝑀𝐹𝑟,𝑚𝑎𝑥 (sin(𝑚𝜃j1 ) − sin(𝑚𝜃j0 )))⁄(𝑚𝜋) -30 -3
10

𝐴𝑠,𝑚 = -40
-4

(−1)𝑗+1 (𝜀𝑝 𝑀𝑀𝐹𝑟,𝑚𝑎𝑥 (cos(𝑚𝜃j1 ) − cos(𝑚𝜃j0 )))⁄(𝑚𝜋) -50 10

0.14 0.15 0.16 0.17 0.18 0.19 0 66.8 200.4 334 400
Time (sec) Frequency (Hz)
(a) (b)

Rotor eccentricity fault Figure 3: Comparison between current waveforms in phase A, in

The REF is a condition of non-uniform air-gap between the the time (a) and the frequency (b) domains, for different cases.
 Similarly to the current signal, it is noticed that the studied frequency-, space,- and harmonic- domains.
faults don’t affect significantly the torque spectrum and it can
Scalar time- and space domain indicators
neither be used for accurate fault detection. Figs 5 and 6 give,
respectively, the time- and the space- domains vibration In the time-domain approach, characteristic features are
signals, 𝑋(𝑡) and 𝑋(𝜃𝑠 ), with their spectral and harmonic extracted from the discrete time signal ((𝑥𝑖 )𝑖=1,..,𝑁 )𝑗 at the
components, for different cases. As shown in these figures, position 𝜃𝑗 . The commonly used scalar time indicators are
SE, DE and partial crack have different effects on the listed in [21]. The considered indicators, in this study, are:
vibration signal in the time-, space-, frequency- and peak-to-peak (P-2-P); root-mean-squared (RMS), Kurtosis
harmonic-domains. In the next section, several indicators will (Ku), skewness, crest (CF) and impulse (IF) factors.
be extracted, from vibrations in these domains, to select the Due to the asymmetrical type of REF and DMF in the
most discriminative feature(s) for each type of fault. space-domain (𝜃𝑠 ), the vibration distribution is significantly
2
affected in this domain, as shown in fig.6 (a). For this reason,
10
200
Healthy
45% SE
the same scalar indicators, extracted from the time-domain,
10
1
45% DE
45% PDMF
1
are calculated from the discrete space-domain signal
(((𝑥𝑗 )𝑗=1,..,𝑀 )𝑖 at time 𝑡𝑖 .
Magnitude (N.m)

45% PDMF2
EM torque (N.m)

150
0
10

100
Frequency-domain indicators
-1
10
The frequency-domain indicators are extracted from the
50 10
-2 Fourier transform (X (f)) of the time signal ((𝑥𝑖 )𝑖=1,..,𝑁 )𝑗 . As
shown in fig.5 (b), the studied faults give rise to additional
frequencies 𝑓𝑘′ multiples of the mechanical frequency (𝑓𝑒 ⁄𝑝),
-3
0 10
0.18 0.182 0.184 0.186 0.188 0.19 0.192 0.194 0 100 200 300 400.8
Frequency (Hz)
where 𝑓𝑒 the electrical frequency and p is the number of pole
Time (sec)
(a) (b)
pairs. In this study, the frequency-domain indicators are the
Figure 4: Comparison between EM torque waveforms, in the time
(a) and the frequency (b) domains, for different cases. magnitudes “𝐴𝑓𝑘′ ” in (µm) of frequencies “𝑓𝑘′ ”.
The harmonic-domain indicators are extracted from the
6
Fourier transform (X (h)) of the space signal ((𝑥𝑗 )𝑗=1,..,𝑀 )𝑖 .
Healthy
10
0 45% PDMF
45% SE
2 As shown in fig.6 (b), the studied faults give rise to additional
4

-1
45% DE
45% PDMF1
space-harmonics “k”. The considered harmonic-domain
10
indicators are magnitudes “𝐴ℎ𝑘 ” in (µm) of harmonics “k”.
Vibration X(t) (µm)

Magnitude (µm)

-2
10
0 Results and discussions
-2 10
-3
Different time- and space-domains scalar indicators are
-4 -4
listed in Table II. In order to study the relevance of these
10
indicators, their relative values (𝐼𝑛𝑑𝑟 ) in (%) with respect to
-6
0.18 0.2 0.22 0.24 0.26 0 100 200 300 400
the healthy case (𝐼𝑛𝑑𝑎,ℎ ), calculated by (12), are considered
Time (sec) Frequency (Hz)
(a) (b)
instead of their absolute ones (𝐼𝑛𝑑𝑎 ).
Figure 5: Comparison between time-domain vibration signals (a) 𝐼𝑛𝑑𝑟 (%) = (𝐼𝑛𝑑𝑎 − 𝐼𝑛𝑑𝑎,ℎ )⁄𝐼𝑛𝑑 𝑎,ℎ × 100 (12)
and their frequency components (b) for different cases.
The frequency- and harmonic- domains indicators
6
(𝐴𝑓𝑘 ; 𝐴ℎ𝑘 ), in (µm) for k = {0, 1, 2, 3 and 4} are given in Table
Healthy

10
0
45% SE
45% DE
III. For the healthy case, dominant frequency and harmonic
4 45% PDMF1
components are {𝑚𝑝, 𝑛𝑓𝑒 } with m = {0, 2} and n = {0, 2, 4,
Vibration X(theta) (µm)

45% PDMF
2
2
6}. Indeed, “𝐴ℎ𝑘 ” for k = {0, 2p (12)} and “𝐴𝑓𝑘′ ” for k’ = {0,
Magnitude(µm)

-1
10

0
2p (12), 4p (24), 6p (36)} exist even for a healthy case while
10
-2
others are only produced by faults.
-2
The P-2-P and RMS values and “𝐴ℎ2 ” can effectively detect
-4
10
-3
the occurrence of a fault. They have almost increasing
monotonic evolutions in function of the fault gravity, with
-6 -4
0 60 120 180 240 300 360
10
0 1 2 3 4 5 6 7 8 9 10 11 12 13 high values compared to the healthy case. However, they are
Harmonic
Space (degree)
(a) (b) overlapping for different fault types and they can’t identify
the fault type. In addition, “𝐴ℎ1 ” gives the same result for a
Figure 6: Comparison between space-domain vibration signals (a)
and their harmonics (b), for different cases SE or DE with more important values compared to the crack
case. Therefore, it is an effective indicator for the eccentricity
V. FEATURE EXTRACTION AND ANALYSIS FROM case (SE or DE). The “𝐴𝑓2 ” gives an almost zero value for the
VIBRATION SIGNALS SE case. However, it’s significantly affected by both DE and
crack and it can be used to distinguish them from SE.
In this section, the set of vibration signals obtained for As fault degrees increase, the kurtosis, time domain “IF”
different fault types and severities, are analyzed in the time-, manifest almost the same behaviors for different faults with
more or less sensitivity. They have weak relative values with Table III: Evolution of frequency and harmonic indicators (µm) in function
of fault degrees (%)
respect to the healthy case, compared to other indicators.
Furthermore, their evolutions with fault severity are not FREQUENCY HARMONIC
monotonic and they are overlapping for different fault types. 1200
PDMF
1
25
PDMF
1

Similarly, “𝐴ℎ0 ” gives weak relative values compared to the SE SE

1000 20 DE
DE

𝐴𝑓0 ; 𝐴ℎ0 (%)

PDMF PDMF
2 2

healthy case (around 25% for 100% of fault). Therefore, they 800 15

can be used neither for detection nor isolation purposes. 600 10

The magnitudes “𝐴𝑓0 ” and “𝐴ℎ3 ” (for 𝜎𝑠 > 0.5), the “CF” 400 5

(for 𝜎𝑠 > 0.4) and “IF” (for 𝜎𝑠 > 0.4) calculated in the space 200
0

domain are only sensitive for the SE case. 0

-5
0 20 40 60 80 100 0 20 40 60 80 100
1
0.8 PDMF PDMF
Table II: Evolution of scalar indicators (%) in function of fault severity (%) SE
1
0.9 SE
1

0.7 DE DE

𝐴𝑓1 ; 𝐴ℎ1 (µm)

0.8
PDMF PDMF2
2
TIME SPACE 0.6 0.7

100 120 0.5 0.6

PDMF1 PDMF
1
90 0.4 0.5
PDMF PDMF
P-2-P (*100 %)

2 100 2
80 SE SE 0.4
DE 0.3
DE
70 0.3
80
0.2
60 0.2

50 60 0.1
0.1

40 0 0
0 20 40 60 80 100 0 20 40 60 80 100
40 18
30 25
PDMF
1
20 16 SE PDMF
20 1
DE

𝐴𝑓2 ; 𝐴ℎ2 (µm)

10 20 PDMF
14 PDMF2 2
SE
0 0 DE
0 20 40 60 80 100 0 20 40 60 80 100 12
15 20 15
14 PDMF PDMF 10
1 1
18
PDMF PDMF
2 2
RMS (*100%)

8
12 SE 16 SE 10
DE DE
14 6
10
12 4 5
8
10 2
6 8 0
0 0 20 40 60 80 100
0 20 40 60 80 100
6 1.2 9
4
PDMF PDMF
1 1
4
SE 8 PDMF
2 2
2 1 DE
𝐴𝑓3 ; 𝐴ℎ3 (µm)

SE
PDMF 7
0 2 DE
0 20 40 60 80 100 0
0 20 40 60 80 100 6
0.8
50 45
PDMF PDMF
1 1 5
PDMF 40 PDMF2
2 0.6
40 SE 4
SE
35
DE DE
0.4 3
30 30
Ku (%)

2
25
20 0.2
20
1

10 15 0 0
0 20 40 60 80 100 0 20 40 60 80 100
0.4 0.4
10 PDMF PDMF
0 1 1
0.35 PDMF 0.35 PDMF
5 2 2
𝐴𝑓4 ; 𝐴ℎ4 (µm)

SE SE
-10 0 0.3 DE 0.3 DE
0 20 40 60 80 100 0 20 40 60 80 100
14
150 x 10 0.25 0.25
6
100
SKEWNESS (%)

PDMF 0.2 0.2

1
50 PDMF 4 PDMF
2 1
0.15 0.15
SE PDMF2
0 2
DE SE
0.1 0.1
DE
-50 0
0.05 0.05
-100 -2
0 0
0 20 40 60 80 100 0 20 40 60 80 100
-150 -4

-200 -6 The DE case can be distinguish from other fault types by

-250
0 20 40 60 80 100
-8
0 20 40 60 80 100 monitoring the magnitude “𝐴𝑓1 ” or the “CF” (for 𝜎𝑑 > 0.3)
70 80

60 70
calculated in the time domain. Moreover, the magnitudes
50 60 “𝐴𝑓3 , 𝐴𝑓4 , 𝐴ℎ4 ” can be used to identify the partial crack in
PDMF
1
different poles.
CF (%)

40 50
PDMF
2
SE
30 40
DE The skewness of the vibration signals calculated in the time
20 30
PDMF
1
and space domains are given in row (4) of Table II. In the time
10 20

0
PDMF
SE
2
10
domain, the skewness under SE and DE are non-monotonic
-10
0 20 40 60
DE

80 100 0
for fault severities less the 60%; but significantly different for
0 20 40 60 80 100
120
PDMF
120 the crack case (around 120%). Moreover, the skewness in the
1
100 PDMF
SE
2 100 space domain for SE and DE are only overlapping for small
80 DE
80
faults (with a severity less than 15%). Over this value, this
IF (%)

60 indicator can be only used to identify the fault type (SE, DE

60
40 or partial crack) because it remains constant as the fault
20
40
PDMF
gravity increases (greater than 40%). However, none of the
1
0 20 PDMF
SE
2 aforementioned indicators could be used to identify the
-20
0 20 40 60 80 100 0
0 20 40 60 80
DE
100
cracked pole in the DMF case.
 CONCLUSION & FUTURE PERSPECTIVES Diagnosis with Kullback-Leibler Divergence : From
An analytical based simulation study was performed to Theory to Applications,” Supélec, 2015.
analyze the effects of REF and DMF on EM vibration signals [6] Y. Lei, Z. He, and Y. Zi, “A new approach to intelligent
in a three-phase BLDC motor. First, an analytical model of an fault diagnosis of rotating machinery,” Expert Syst.
inverter-fed PM DC motor, able to generate electrical, Appl., vol. 35, no. 4, pp. 1593–1600, 2008.
magnetic and vibration signals for different cases, was [7] H. A. Toliyat, S. Nandi, S. Choi, and H. Meshgin-kelk,
developed. The analytical flux density and stator natural Electric Machines - Modeling Condition Monitoring
frequencies calculations were compared to FEA under healthy and Fault Diagnosis, CRC Press . 2013.
conditions. Both REF and DMF with different severities were [8] H. Su, K. T. Chong, and R. Ravi Kumar, “Vibration
simulated in the model by changing some of its parameters, signal analysis for electrical fault detection of induction
according to the fault type. Due to the limitation of electrical machine using neural networks,” Neural Comput. Appl.,
(current) and magnetic (torque) signals to identify any fault, vol. 20, no. 2, pp. 183–194, 2011.
the EM vibrations were selected to extract fault-related [9] Z. Yang, X. Shi, and M. Krishnamurthy, “Vibration
indicators. Then, time-, space-, frequency- and harmonic monitoring of PM synchronous machine with partial
indicators were extracted from these signals under different demagnetization and inter-turn short circuit faults,” in
fault types and degrees. The evolution of each indicator was IEEE ITEC, 2014, pp. 1–6.
studied in function of the fault severity in order to select [10] J. C. Gamazo-Real, E. Vazquez-Sanchez, and J. Gomez-
proper criteria able to identify the operating condition of the Gil, “Position and speed control of brushless dc motors
motor. Based on this analysis, different indicators were using sensorless techniques and application trends,”
classified according to their sensitivity to fault type and their Sensors, vol. 10, no. 7, pp. 6901–6947, 2010.
effectiveness for detection or isolation purposes. [11] M. M. Momenzadeh, A. F. Ahmed, and A. Tolba,
Moreover, other types of faults may occur in PMSMs such “Modelling and Simulation of The BLDC Electric Drive
as the inter-turn short-circuit fault (ISF) and bearing faults. In System Using SIMULINK/MATLAB for a Hybrid
the next step, these faults will be modeled and their effects on Vehicle,” Paderborn, Germany, 2014.
the vibration behavior of the machine will be studied. In this [12] J. F. Gieras, C. Wang, and J. C. Lai, Noise of polyphase
paper, only the case of abrupt faults with constant severity was electric motors. 2006.
studied; the incipient types with variable severities could be [13] J. Le Besnerais, “Réduction du bruit audible d’origine
considered in our future works. In fact, applied conventional magnétique dans les machines asynchrones alimentées
signal processing techniques are only effective under par MLI” École Centrale de Lille, 2008.
stationary conditions, at constant speed. However, under non- [14] A. B. Proca, A. Keyhani, A. El-Antably, W. Lu, and M.
stationary cases as time-varying processes, these techniques Dai, “Analytical model for permanent magnet motors
have some limitations which explain the need of mixed- with surface mounted magnets,” IEEE Trans. Energy
domain signal processing tools (time-frequency, wavelet Convers., vol. 18, no. 3, pp. 386–391, 2003.
decomposition...) for feature extraction. [15] S. Salon, Finite Element Analysis of Electrical
Machines. Troy, New York, 1995.
ACKNOWLEDGMENT [16] R. Ginzarly, K. Alameh, G. Hoblos, and N. Moubayed,
“Numerical Versus Analytical Techniques for Healthy
The authors gratefully thank the “Haute Normandie” Region and Faulty Surface Permanent Magnet Machine,” in 3rd
for financially supporting this work. Inter. Conf. on EECEA, 2016, pp. 83–87.
[17] G. Verez and C. Espanet, “Natural Frequencies
REFERENCES Analytical Modeling of Small Industrial Radial Flux
[1] A. Djerdir, J. A. Farooq, A. Rezig, and A. Miraoui, Permanent Magnet Motors,” in 18th ICEMS, 2015, pp.
“Faults in permanent magnet traction motors: State of 1963–1969.
the art and modelling approaches,” in IEEE PES [18] N. Bracikowski, D. Ilea, F. Gillon, M. Hecquet, and P.
General Meeting, PES 2010, 2010, pp. 1–5. Brochet, “Design of permanent magnet synchronous
[2] Y. Da, X. Shi, and M. Krishnamurthy, “Health machine in order to reduce noise under multi-physic
monitoring, fault diagnosis and failure prognosis contraints,” in IEEE IEMDC, 2011, pp. 29–34.
techniques for brushless permanent magnet machines,” [19] B. M. Ebrahimi and J. Faiz, “Magnetic field and
in VPPC 2011, 2011, pp. 1 – 7. vibration monitoring in permanent magnet synchronous
[3] M. Riera-Guasp, J. A. Antonino-Daviu, and G.-A. motors under eccentricity fault,” IET Electr. Power
Capolino, “Advances in Electrical Machine, Power Appl., vol. 6, no. 1, pp. 35–45, 2012.
Electronic, and Drive Condition Monitoring and Fault [20] Z. Wu, “Conception optimale d’un entrainement
Detection: State of the Art,” IEEE Trans. Ind. Electron., électrique pour la chaîne de traction d'un véhicule
vol. 62, no. 3, pp. 1746–1759, 2015. hybride électrique,” Université de Franche Comté, 2012.
[4] K. Alameh, N. Cité, G. Hoblos, and G. Barakat, [21] K. Alameh, N. Cité, G. Hoblos, and G. Barakat, “Feature
“Vibration-based Fault Diagnosis Approach for extraction for vibration-based fault detection in
Permanent Magnet Synchronous Motors,” IFAC- Permanent Magnet Synchronous Motors,” in 3rd Inter.
PapersOnLine, vol. 48, no. 21, pp. 1444–1450, 2015. Conf. on TAEECE, 2015, pp. 163–168.
[5] J. Harmouche, “Statistical Incipient Fault Detection and

View publication stats