HAL Id: hal-02134729
https://hal.archives-ouvertes.fr/hal-02134729
Submitted on 20 May 2019
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of
sci-entific research documents, whether they are
pub-lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destinée au dépôt et à la diffusion de documents
scientifiques de niveau recherche, publiés ou non,
émanant des établissements d’enseignement et de
recherche français ou étrangers, des laboratoires
publics ou privés.
Health monitoring of bearing and gear faults by using a
new health indicator extracted from current signals
Moncef Soualhi, Thi Phuong Khanh Nguyen, Abdenour Soualhi, Kamal
Medjaher, Kamel Eddine Hemsas
To cite this version:
Open Archive Toulouse Archive Ouverte (OATAO)
OATAO is an open access repository that collects the work of some Toulouse
researchers and makes it freely available over the web where possible.
This is
an author's
version published in: https://oatao.univ-toulouse.fr/23769
Official URL :
https://doi.org/10.1016/j.measurement.2019.03.065
To cite this version :
Any correspondence concerning this service should be sent to the repository administrator:
tech-oatao@listes-diff.inp-toulouse.fr
Soualhi, Moncef and Nguyen, Thi Phuong Khanh and Soualhi, Abdenour and Medjaher,
Kamal and Hemsas, Kamel Eddine Health monitoring of bearing and gear faults by using a
new health indicator extracted from current signals. (2019) Measurement, 141. 37-51. ISSN
0263-2241
OATAO
Health monitoring of bearing and gear faults by using a new health
indicator extracted from current signals
Moncef Soualhi
a,⇑, Khanh T.P. Nguyen
a, Abdenour Soualhi
b, Kamal Medjaher
a, Kamel Eddine Hemsas
c aLaboratoire Génie de Production, Université de Toulouse, INPT-ENIT, 47 Av. d’Azereix, 65000 Tarbes, France b
LASPI, CUR, University of Saint-Etienne, France c
LAS, UFAS1, Sétif, Algeria
a r t i c l e i n f o
Keywords:
Bearing and gear faults Health monitoring Signal processing Feature extraction Health indicator Machine learning Artificial intelligence Fault detection and diagnostics Motor current signal analysis
a b s t r a c t
Gear reducer motors play an important role in industry due to their robustness and simplicity of con struction. However, the appearance of faults in these systems can affect the quality of the product and lead to significant financial losses. Therefore, it is necessary to perform Prognostics and Health Management (PHM) for these systems. This paper aims to develop a practical and effective method allow ing an early fault detection and diagnostic for critical components of the gear reducer, in particular gear and bearing defects. This method is based on a new indicator extracted from electrical signals. It allows characterizing different states of the gear reducer, such as healthy state, bearing faults, gear faults, and combined faults. The diagnostic of these states is done by the Adaptive Neuro Fuzzy Inference System (ANFIS). The efficiency and the robustness of the proposed method are highlighted through numerous experimental tests with different levels of loads and speeds.
1. Introduction
Gear reducer motors are widely used in industrial applications due to their robustness and low cost. However, during their life cycle, different degradation types can occur in these systems lead ing to undesirable situations such as: system degradation, down time, high maintenance costs, product quality damages, etc. Therefore, maintaining such systems in a good condition requires the implementation of an adequate maintenance strategy. The pre dictive maintenance, using Prognostics and Health Management (PHM), can be a good candidate. It ensures, on one side, the relia bility, availability, maintainability and safety of industrial systems
[1,2]. And on the other side, it allows the detection and diagnostics of machine faults[1,3]. According to experts statistics, bearing and gear faults represent a significant part of the defects of gear redu cer motors[1,3 5]. Hence, it is essential to adopt efficient monitor ing methods to diagnose their faults.
In literature, fault detection and diagnostics (FDD) approaches can be generally classified into two groups [1,2,6]: model based and data driven based approaches. The model based approaches
use mathematical equations to represent the system behavior. It is more accurate than the data driven approaches. However, con sidering the complexity of systems, it is often difficult to imple ment the model based methods. The second group is based on the analysis of signals extracted from different types of sensors. It is suitable for complex systems where no a priori knowledge is needed to monitor the system. However, its performance strictly depends on the availability of sufficient and representative data
[7]. The choice of an appropriate approach depends on the system knowledge we have on it, and the availability of historical degrada tion data. In reality, it is difficult to model the degradation pro cesses of bearings and gears because of the complexity of their mechanism, which is nonlinear, non stationary and stochastic. Therefore, the data driven approach is chosen because of abundant data acquired by different sensor types[1]. Among them, vibration and current signals are promising and non invasive parameters for monitoring.
The FDD of bearings and gears can be obtained by using the time domain analysis. In this case, statistical features, such as Root Mean Square (RMS), Standard Deviation (StD), Kurtosis (KUR), Skewness (SKE), etc., are extracted from vibration signals to per form condition monitoring[8 10]. On the other hand, the authors in[11 13]propose the use of the frequency analysis to identify the ⇑ Corresponding author.
characteristic frequencies of the localized roll bearing defects. Gear fault diagnostic is addressed in[5,13 15]. The authors in[5,13,14]
extract fault characteristic frequencies to localize gear defects while the work in[15]uses frequency domain features to detect gear abnormalities. In the time frequency domain, the studies
[16 21]propose to use the wavelet transformation methods for the detection of bearing localized faults. In addition, the work in
[22]proposes to combine the fast dynamic time warping method and the kurtosis technique for fault detection of gears. The study
[23]uses angular measurements to diagnose different gear faults. In addition to vibration signals, acoustic emission data can be used as an alternative for condition monitoring, as in[24 26]. In these works, the authors show the effectiveness of acoustic signals in motor fault detection and diagnostics, including bearing, gear and electrical faults such as shorted coils. However, fault detection and diagnostic based on acoustic emission signals may be unfeasi ble when the motor runs too quietly[24].
The main drawback of the above studies is that the sensitivity of the vibrations and the acoustic emissions can be reduced due to the industrial environment noises. Therefore, it is difficult to diag nose electrical faults in motors via these signals. Hence, electrical signals (current, voltage, etc.) can be used as a non invasive alter native for bearings and gears health monitoring.
Considering the previous studies, the bearing and gear faults are often separately addressed in the literature. Moreover, to our knowledge, no existing research takes into account numerous operating conditions of motors when considering its different fail ure types, such as bearing faults, gear faults and both component faults in the same time. Therefore, this paper aims to fill this liter ature gap. In detail, a new indicator extracted from the three phase current signals is presented to characterize different system states, as healthy state, bearing faults, gear faults and combined faults of an asynchronous motor driving a geared box, with different levels of loads and speeds. The diagnostic of these classes is done by arti ficial intelligence using the Adaptive Neuro Fuzzy Inference Sys tem (ANFIS). The remainder of the paper is structured as follows. Section2presents the proposed methodology to extract and build health indicators. The performance and robustness of our health indicators are highlighted in Section 3 through experimental results carried out on a test bench provided by the LASPI labora tory. Finally, the conclusion and perspective of this work will be presented in Section4.
2. Proposed methodology for fault detection and diagnostics This section presents the main steps of the proposed methodol ogy for bearing and gear fault detection and diagnostics (Fig. 1).
The system analysis allows identifying the critical components and the corresponding failure mechanisms leading to sensor place ment and data acquisition. The recorded data are then processed to extract relevant features. For this purpose, both of the frequency and the time domain are investigated. In detail, the frequency anal ysis is used for each current signal to extract a characteristic value (the MAX value of the FFT) corresponding to different load varia tion states. On the other hand, the time domain analysis is applied to extract values (such as the peak to peak value of the signal amplitude) that allow tracking the evolution of the bearing and the gear degradations. After that, a new feature is evaluated and used to build health indicators. These indicators are exploited in the third step to identify and classify the different health states of the motor’s critical components (healthy state, bearing faults, gear faults, and combined faults). The classification is performed by using pattern recognition methods, which are part of Machine Learning (ML). The following subsections describe in details the above mentioned steps.
2.1. From system to data acquisition
One of the main tasks of health monitoring is to identify the appropriate physical parameters to be observed in order to track the system degradation process. To achieve this task, a methodol ogy is proposed and shown inFig. 2.
At the beginning of the methodology, it is necessary to analyze the architecture, the structure and the functionalities of the system in order to determine the critical components leading to system failure. For this purpose, numerous approaches can be used such as experience feedback, fault tree, event tree, cause and effect tree, etc.[1]. In the framework of this paper, we focus on asynchronous motors. According to the studies in[1,3,27], bearing failures repre sent 41 45% of faults in the induction motors. Moreover, according to the studies presented in[4,5], gear defects are also the main cause that leads to wind turbines or gear reducer motor failures. Therefore, bearings and gears can be considered as the motor crit ical components whose health states should to be monitored over time to detect and diagnose their faults. To do this, it is important to determine the most suitable physical parameters for health monitoring. According to the studies presented in[3 5,8,9], the vibration signal is the most used data to track the bearing and gear degradations. However, the sensitivity of this type of signal can be reduced due to noises in operational environments. Furthermore, one of the major disadvantage of vibration monitoring is its high cost of the accelerometers and the difficulties when accessing to the machine to install the sensors. Besides, the electrical sensors are inexpensive and easy to implement. Therefore, current sensors, which are considered as a non invasive way of monitoring the motors, are used in this paper. This is also known as Motor Current Signal Analysis (MCSA)[28].
2.2. From data to features extraction and health indicators construction
This subsection deals with the extraction of a new health indi cator from the three phase current signals. The extracted health indicator will be used to detect and diagnose bearing and gear defects in gear reducer motors. The extraction process is shown inFig. 3.
Compared to what is reported in the literature[1,8,15,16,24,26], the health indicator proposed in this paper is robust when taking into account the impact of different load levels variation on the machine. Also, this indicator allows detecting both bearing and gear faults simultaneously. The main steps of the features extrac tion and health indicators construction are presented hereafter. 1. Data acquisition. Depending on the operating conditions of the
motor, the three phase raw current signals (iað Þ; it bð Þ, and it cð Þ)t are used to extract the relevant features. These raw signals are shown inFig. 4.
2. Splitting every current signal into N segments of length L. The obtained signals are split into several segments of length L. In this contribution, each recorded signal equals to 10 s and is split into 100 segments of 0:1 second (L 0:1 s, N 100). Each signal segment is denoted by yjh(Fig. 5) where j character izes the phase current (j2 a; b; c½ ) and h represents the segment number (h2 1; . . . ; N½ ). This step aims to reduce the data size for signal processing, and takes only the relevant features, such as the peak to peak and the amplitude maximum values from the observations.
0
System analysis...
•
Three phase currents..
--·r---...
.
---
.
-..
-Signal processing : MAX (FFTJ(y0hl ) , ·• ,. , ,. , ., , ., • ., , +-+----<>--'(
·
~~~
;
~
·
~
·
~~~
·
;
~
;~
~
~~~
·
~;~~;i
·-~
H
p
-
~
:
...
~~
.
~
i
.
~
.
'.l1n~r
... )
•~
[
~
~Jt
_
~
-
-
-
-
.
:
-
·
-
.
-
.
-
-
~
-
--
-
~~
~
~
·
1
( Ratio= Peak-to-peak (y,.,, >
···1
H
1
~
~
~W
!
ffiM
~
MW
~
IMYl~
W
M~
ii
1
•~
wi1
@
w~1
m
m1
~
}
lndl le
n
05 : MAX (FFT) (Yo•J) •• • "' • ., • ., • .. • !, indi= RMS (ratio) J , .. ~ "''~
·
1
•
::;~ 03 02•
• •...
.Sfo.(Y)_
...
J • " •• •• ,.., - - •• lndiia @. 06 lndiib f;i!i
07 ~i~ '. Ratio= Peak-to-peak(v,11',)···:
H
J::
I
'.
~
'. '.
:
~
: MAX (FFTJ(v0hl J , , .. • .. • .. • !, ind1= R~i~~tit,i
!
{
__
)e
-
·"
..
i
'···I... • --•
Fault detection and diagnostics-
0
Health Indicator construction - • Time and frequency domain analysis fig. 1. The proposed methodology for FOO of gear reducer motors.••
•
•••
•
•
•
• System•
•••••
••
Sub-sv,tem Physical parameter0
fig. 2. From system to data acquisition.
is
n
ecessary
to calculate the
rat
io
between the peak to peak
values of
eac
h
segmented signa
l
in
time domain
(yih)
and
its
spectra
l
amplitude
in fr
eque
ncy
domain
(MAX(
FFT
(yih))
. This
ratio allows grouping a
ll
the observations of the different
load
leve
ls
( e
.g.
0%, 25%, 50%, 75%)
in
one
class
.
Figs. 6 and 7 illustrat
e
the dispersion problem and the
effect
iv
eness
of the ratio,
respectively.
In
detail,
Fig. 6
shows an
examp
l
e
of the
distribut
io
n
of the
health
ind
icators
observations. These health
ind
icators
repre
sent three different health states
(e
.g.
healthy, faulty
1
, fau
lty
2) without the normalization to the MAX(FFT(yih)) value.
Each
health state contains severa
l
groups of observations, which cor
respond to
d
iffe
r
ent
load l
eve
ls
( e
.g.
0%,
25
%,
50%, 75
%)
. Fig. 7
shows the
e
ffect
of the normalization on the
reduct
io
n
of the
observations dispersion caused by the
load
leve
l
variations.
Th
e
Fig. 8
shows the
frequ
e
ncy
and the time domain
plots
of
the recorded signal.
Th
ese plots
allow separating the classes
by a decision making rule.
This
rule is based on the
va
lu
es
of
the ratio between
Yih
amplitudes and MAX(yih(f)), which are
then used to represent
eac
h h
ea
lth
state
(regroup
ing
different
load leve
l
observations,
in
Fig. 6
)
by a
class
as
illustrat
ed
in
Fig. 7.
Yih
·
where Y
ih(f)
MAX(y
ih(f))'
(1)rr,nprr,~n.,,, 1 '11'1 111·1"' "''" "''"" '""'"' I ~ 1 ti ) U I
..
u s~ ~~-: -: -: -:
l
.
,.
TI • It 10 U IM 1114 Mlf')RIIUID•••
....
•.
Data
•
•
acquisition
•••••
••
>
Data processing
>
Feature extraction
Fig. 3. From data to features extraction and health indicators construction.
Number of samples
Fig. 4. Three-phase raw current signals.
RM
S
(
zih
)
S
tD
(
Y
i)
(2)where
zihis the h
thsignal segment of the/' phase,
zihand
Yi consist of
ne
and
Ne
sampling points, respectively. The RMS and the
StD are respectively the root mean square and the standard devi
ation. The RMS value measures the average energy of the signal,
each degradation will vary the R
MS values
[30]
.
On
the other
hand, the StD value allows to limit the d
i
spersion between the
indicators of different health states.
RM
S
(
z
ih)
✓
-1
X2
~
)
jh
(
n
)2
ne
n ,(
3
)
S
t
D
(
Y
i)
{
~e
x
t
(v
j(
n
)
y
ir
(4) 500 1000 1500 2000 2500 3000 Number of samples4. Construction of h
ea
l
th indicators
.
The obtained indicator is
then
exp
l
oited to build
health indicators which are used to
classify, detect and diagnose the bearing and
gear faults.
Note that the utilizat
i
on
of only one
indicator
is not suffi
cient to detect severa
l
faults. In this case, the combinat
i
on
of
different
ind
i
cators is
necessary
to
move
from
one
dimens
i
ona
l
space to three dimens
i
ona
l
space by using a
pattern
recognit
i
on techn
i
que as shown in
Fig. 9
.
This
t
echnique is based on
class
i
fying N observat
i
ons
denoted
(
indiiah
,
indiilil, indi;,h)into
classes
where
h
E[
1
,
...
,
N]
.
Each
observat
i
on
is characterized
by a vector
including the three health ind
ic
ators corresponding to the
three phase current signa
l
s of the motor. This vector is then
used to build
a
matrix of health indicators as illustrated
hereafter.
indiia1
indi
ib
1 indi
;
,1
indiia2 indi
ib2
indi
;
,
2
indi
indiia3 indi
ib
J indi
;
,3
indiiaN indiibN indiicN
2.3.
From health indicators to fault detection and
diagnosticsTh
i
s step aims to map each value of the above
matrix to a cor
responding
class
(
e.g. healthy, degraded, faulty, etc.
)
by using a
pattern recognit
i
on technique. This technique is based on a dassi
tier
mode
l
using a training database for the recognition
of
the
membership
class of
each
observat
i
on.
Note that t
his
techn
i
que
requires
a prioriknowledge of the
classes for
the construct
i
on
of
g:
Q> 0.S-
yal
J
Q. -0 •~
E -1 < 0i
~.
:
:
-~
10 -- ycJ; 0
'is. •0-
.~
5c
E -1 < 0+
•-····
ne•
ya!i
yas --- yah---
- - -- - -
' -0.5 2 2.5 yb!i ybs ---· ybh---
~ - - -- - -
--'----0. 5 2 2.5 yc!i yes --- ycb 0.5 2 2.5tn
di
(c)
First
he
a
l
th stat
e
/
Second
he
al
t
h
s
t
ate
/
Third
hea
l
th
s
ta
t
e
/
0
%
75
%
25
%
0
%
lnd
i
(
a)
Fig. 6. Dispersion of features caused by the variation of load levels.
t
h
e t
r
a
inin
g database
. In
our case, the
n
u
m
be
r
o
f
s
t
ud
i
ed classes
i
s
equa
l
t
o 7
. E
ach class is compose
d
o
f
100
observat
i
ons
. T
hus, t
h
e
n
umber o
f
a
ll
observat
i
ons N is equa
l
to 700
. T
o perfor
m
t
h
is task,
t
h
e co
n
structed mat
ri
x indi is
d
iv
i
ded
in
to a tra
inin
g database a
nd
a test database
. Th
e tra
inin
g database is der
i
ved
f
ro
m
classes
n
a
m
e
d
05,1
,-;;
s ,-;;
7, w
h
ere s
r
eprese
n
ts t
h
e
n
umber o
f h
ea
l
t
h
states o
f
t
h
e gearbox compo
n
e
nt
s
.
Each t
r
a
inin
g class is
r
ep
r
e
se
n
ted by a
m
at
ri
x composed o
f
50% o
f
observat
i
o
n
s taken
r
a
n
domly
f
rom eac
h
class
a
n
d
defi
n
ed
by
t
h
e
vecto
r
(indi;ah, indiibh, indi;,h)-
Th
e t
r
a
inin
g database correspo
n
ds fina
ll
y
to a mat
ri
x o
f
n
lin
es
(
n 350), co
n
ta
inin
g t
h
e healt
h in
dicato
r
s
observat
i
ons, a
n
d th
r
ee colu
m
ns co
rr
espo
n
di
n
g to t
h
e t
h
ree
p
h
ase cu
rr
e
n
t
ind
icato
rs. Th
is mat
ri
x is used to tra
in
t
h
e
cl
ass
i
fie
r
0
%
25
%
50
%
75
%
First
hea
l
th
s
t
ate
•
•
•
•••••
••••
••
•
---
• •
•••••
••
•
•
••
ln
d
i
(
c)
m
ode
l.
Th
e test database is a
l
so co
n
structed random
l
y
f
rom
india
nd
composed o
f
50
%
o
f
each class to test the accu
r
acy o
f
t
h
e das
s
i
fter
m
ode
l
as
ill
ustrated
in
Fig
.
1
0
. Th
e test database correspo
n
ds
fi
n
a
ll
y to a mat
ri
x o
f
n
350
lin
es
.
In
t
h
e
li
teratu
r
e,
n
u
m
erous
m
ach
i
ne lea
rnin
g tech
n
iques are
used to detect a
n
d d
i
ag
n
ose bear
in
g a
n
d gea
r
de
f
ects
.
Amo
n
g these
tec
hn
i
ques, we ca
n
cite the most effect
i
ve met
h
ods suc
h
as Neura
l
n
etworks
(
NN)
[31,32)
, K Nearest Neig
h
bors
(
K NN)
[33)
, Support
vecto
r
mac
hin
es
(
SVM)
[31,34,35)
, Na
i
ve
B
ayes class
i
fie
r
(NB)[36)
a
n
d Adapt
i
ve Neu
r F
uzzy
I
nf
erence System
(
AN
FI
S)
[37,38,3
0
)
.
Eac
h
o
f
these tech
n
i
q
ues p
r
ese
n
ts
i
ts ow
n
adva
n
tages
a
n
d drawbacks
. H
oweve
r
, as t
h
e AN
FI
S comb
i
n
es bot
h
art
i
ficia
l
n
eu
r
a
l
n
etwo
r
ks a
nd f
uzzy
in
fe
r
e
n
ce systems,
i
t a
ll
ows exp
l
o
i
t
in
g
Second
he
a
l
t
h
s
t
at
e
~
•
• •
• •
•
•
••
•••
•
.
•••••
.
..
---• ---•---•---•
0
%
25
%
50
%
75
%
ln
d
i (b)
Third
hea
l
t
h
s
t
a
te
lndi (
a)
(.
••••
....
••
.
\
\·
..
::·
---"'-••• 1/
0
%
25
%
50
%
75
%
0.6 ~
A
:;
~
o.4 Ii 0.2 -0.4 -0.6 la -0.8 ' - - - - " - - - - ' - - - ' - - - - " - - - - ' - - - ' - - - - " - - - - ' 1000 2000 3000 4000 samples 5000a) Time domain
analysis
FFT 150 200 f(Hz) 6000 7000 8000
:
l
250 300b
)
Frequency domain analysis
Fig. 8. Combination of time and frequency analysis for vectors dispersion limitation.
1.8
4
- - - - -
r-91.82-!!!
Ill:lE
~ 1.8g
,,
,E1.78
.
.
~
o oto'bo o,~
•' •
>.
O O,t
~I
/Oct.0<\)
,.,00
\
•$.s
N•
(16•!,
t
" (!:, 0~ o O ofi.f9,..1J• 0, 0 O'\g .... lo 0 8"~ •o•'::
o E2: GNr surftct damage
o El: GNr 1/2 tooth break
o E4: S.arlng ~ntr " " fautl
ES: Bearing oultr race fault
ES: E2+£5 (Combloed ""lbl
o E7: E4+H (Combloed ""lb)
Fig. 9. Health indicators construction characterizing different health states.
•••
...
.
•
.
•
• Health Indicators
•••
••
}
Test matrix (350 X 3) Train matrix (350 X 3) ANFIS dassifier Classifier models m•
.,.
~~ lncll'I.)}
lndi(cJ OS -;1\ \,$ n, 02 @•
04 07•
f/:\ Fault detection and diaanostlcsFig. 10. From health indicators to fault detection and diagnostics.
MF2 MF 3 MF4 lndi••- --MF 1 MF2 MF 3 MF4 Inputs Layer l Fuzzification
Layer 2 Layer 3 Layer4
Defezzification
Layers
Output
Weighting Normalization
Fig. 11. ANAS structure.
t
h
e ab
ili
ty o
f
a
n
eura
l
n
etwork to class
i
fy a
n
d
ide
n
t
if
y patte
rn
s
w
i
t
h
a ru
l
e based fuzzy logic mode
l
lea
din
g to a
n in
crease o
f
t
h
e
lea
rnin
g capacity
[39)
.
Th
e AN
FI
S was
ini
tia
ll
y developed by
[37)
in
1993
. I
t is made o
f
five
l
aye
r n
eu
r
a
l
n
etwo
r
k, w
h
e
r
e each
l
aye
r
pe
rf
o
r
ms a step o
f
a
fuzzy
in
fere
n
ce system o
f
type
T
akag
i
Sugeno, a
l
so ca
ll
ed
(p
r
e
n
euro
n
a
l
a
r
c
hi
tectu
r
e)
.
AN
FI
S uses t
h
e hybrid
lea
rn
ing a
l
go
ri
t
h
m
between the descent g
r
adie
n
t met
h
od a
n
d the
least s
q
ua
r
e method
to m
i
n
i
mize t
h
e e
r
ro
r
betwee
n
t
h
e AN
FI
S output (p
r
edictio
n
va
t
ues) a
n
d t
h
e target va
l
ues (true va
l
ues)
. Th
e genera
l
struc
t
ure o
f
AN
FI
S is s
h
ow
n in
Fig
.
11
.
In
deta
il
s, AN
FI
S has a set o
f in
puts va
l
ues
fr
om the hea
l
th
in
d
i
cato
r
s observations de
n
oted
(
indiah
,
indibh
,
indi,h)
-
Th
ese
in
dicato
r
s
a
r
e ext
r
acted from t
h
e tra
inin
g database t
h
at character
i
zes t
h
e d
if
fere
n
t hea
l
t
h
states o
f
t
h
e gea
r
box
. E
ach observatio
n
o
f
t
h
e t
r
a
inin
g
mat
ri
x is associated to a members
hi
p fu
n
ction, w
h
ich ca
n
be t
r
ia
n
gula
r
, gaussian, t
r
apezoida
l
, etc
. In
ou
r
case whe
r
e fou
r
fu
n
ctions
a
r
e considered, t
h
e AN
FI
S has
j3
4fuzzy
r
ules
µ,
with 3 is t
h
e
n
umber o
f in
puts (hea
l
t
h in
dicators
=
3), and 4 is the
n
umber o
f
members
hi
p fu
n
ctio
n
s (M
F:
gauss, gauss2mf
...
)
. Th
ese
r
u
l
es ca
n
be g
i
ve
n
as fo
ll
ows
:
ifindia11 is
µ{
,
indibh is
µ
t
and indich
is
µ~
34Y
' f
"
)
.v
i
fi
(india11
,
indibh
,
indi,h)
j 1
with
:
fi
(
indi
(
a
,b
,
c
J
h)
W
j
-
(
e{
.indi,,h
+
e{indibh
+
e{indich
+
eL,)
(5) (6) (7)
and where
erepresents the coefficients of the
ru
lej,
w
is the weight
of the rule j, and
Y
corresponds to the outp
u
t of the A
NF
IS model
(the predicted values of the d
i
fferent health states of the motor)
.
First layer (Fuzzification
)
Th
e fi
r
st
l
ayer co
n
ta
in
s as ma
n
y
n
eu
rons as poss
i
ble o
f
t
h
e fuzzy subset
in
t
h
e
i
nfe
r
e
n
ce system
. I
t pe
r
forms the fuzz
i
ficatio
n
o
f
the tra
inin
g
in
put set
f
rom
(
indiah
,
indibh
,
indich),
by calcu
l
at
in
g t
h
e membership degree o
f
each
in
put t
h
roug
h
a membe
r
s
hi
p fu
n
ction
.
O 1
µ
ij
(
indi
(
a,b
,
c
)
h) with i
1
,
2
(
8
)
where
µij
re
p
resents the membershi
p
functions used for the fuzzi
fication
. I
n this paper, the Ga
u
ssian function
[37)
is considered
.
Second layer (Weighting of the fuzzy rules
)
Th
e seco
n
d
l
ayer ca
t
culates t
h
e act
i
vatio
n
degree o
f
t
h
e p
r
emises
(output o
f
the first
l
ayer), where each
n
euro
n in
th
i
s
l
ayer ma
r
ked
1tcorrespo
n
ds to
a fuzzy
r
u
l
e o
f
t
h
e type Suge
n
o
. Th
e output o
f
t
h
is
l
aye
r
(weig
h
ts
W
j
)
correspo
n
ds to t
h
e p
r
oduct o
f
the fuzzy
in
puts
.
The act
i
vatio
n
fu
n
ctions used o
n
t
h
ese
n
eurons depe
nd
o
n
t
h
e ope
r
ato
r
s AN
D
/
OR cited
in E
q
.
(5)
.
(9)
Third layer (Normalization
)
Th
e t
hi
rd
l
aye
r n
o
rm
a
li
zes t
h
e act
i
va
tio
n
deg
r
ee o
f
each ru
l
e,
i.
e
.
each
n
ode
in
this
l
ayer, marked
N,rece
i
ves at
i
ts
in
put t
h
e output o
f
t
h
e p
r
evious
l
aye
r
o
f
the
fhn
euro
n
,
a
nd
t
h
e
n
calcu
l
ates t
h
e
r
atio betwee
n
t
h
e
i'
hrule weig
h
t a
nd
t
h
e sum
ofa
ll
ru
l
e weig
h
ts
. Th
e output o
f
th
i
s
l
ayer is the
n
ormalized weig
h
ts
.
3 Wj
o Wj - - (10)
IJ
'
,w
i
Fourth layer (Defuzzification
)
Th
e fou
r
t
h l
aye
r
e
n
su
r
es t
h
e
defuzz
i
ficatio
n
o
f
the previous
l
ayer to determi
n
e t
h
e pa
r
amete
r
s
o
f
t
h
e act
i
vatio
n
fu
n
ctio
n
e,
i.
e
.
e
is t
h
e co
n
seque
n
t pa
r
ameters
.
o
4wj/i
<'-0
-
(
c{
.indi,,h
+
~
.indi
b
h
+
e{indi,h
+
eL,)
(11)
Fifth layer(Output
)
Th
e
fi
f
t
h l
aye
r
co
n
ta
i
ns a s
in
g
l
e
n
euro
n in
a ci
r
cle ma
r
ked
L
·
I
ts role is to calcu
l
ate the sum o
f
the previous output.
J'
o
5Y
I:<'-0
-
(
c{
.indiah
+
ei
.indibh
+
e~ .indich
+
eL
1 ) j 1(12)
Th
e
?
r
ep
r
ese
n
ts t
h
e predicted va
l
ues o
f
t
h
e d
i
ffere
n
t hea
l
t
h
states o
f
the moto
r. Th
ese va
l
ues w
ill
be used to eva
l
uate and clas
s
i
fy t
h
e data
fr
om the test
in
g set.
3.
Appli
ca
tion
a
nd
re
sult
s
Th
is sectio
n
p
r
ese
n
ts t
h
e applicatio
n
use
d
to test a
n
d ve
rif
y t
h
e
pe
r
forma
n
ce a
n
d t
h
e robust
n
ess o
f
the proposed methodology for
bea
rin
g a
n
d gea
r
fau
l
t detectio
n
a
n
d diag
n
ostics
. Th
e applicatio
n
consists o
f
a test bench
in
stalled at
l
aboratory
leve
l
(
Fig
.
12
)
.
Th
ree phase cu
r
re
n
t signa
l
s a
r
e co
n
t
in
uously
r
eco
r
ded
f
rom t
h
e
output o
f
t
h
e pu
l
se generator p
l
aced before t
h
e motor
. T
hese sig
n
a
l
s
a
r
e
r
ecorded at d
iff
e
r
e
n
t operat
in
g co
n
ditions by vary
i
ng the speed
a
nd
t
h
e load
. Th
ey a
r
e t
h
en processed sepa
r
ate
l
y,
in fr
equency a
nd
t
i
me doma
in
s, to s
h
ow the
li
mits o
f
t
h
ese traditiona
l
app
r
oaches
a
nd
to emphasize t
h
e adde
d
va
l
ue o
f
the proposed hea
l
t
h i
ndicato
r.
Fin
a
ll
y, t
h
e obta
in
ed hea
l
th indicators a
r
e
f
ed
in
to class
i
fier mode
l
s
Fig. 12. Test bench installed at LASPI laboratory.
3.1.
Description
of
the test bench
Th
e
Fig. 12
shows the
test be
n
ch
inst
a
ll
ed at
the
lASP
I
I
abora
tory
in Fran
ce
. I
ts overa
ll
scheme is presented
in
Fig. 14
.
I
n detail,
the
asy
n
chronous cage
motor
drives a
three
axis gearbox
.
This
I
at
te
r
compone
nt
is
composed o
f
three rotat
i
ng shafts.
The
first shaft
AE,
also
named
as
the
in
put
shaft,
is
directly driven
by
the rotor
shaft. At the
output
shaft,
a
n
e
l
ectromag
n
etic brake is
placed
in
o
rd
e
r
to apply a
l
oad
l
eve
l
to
the motor.
The
second shaft
Alco
n
tains the
gea
r
a
nd
the
bea
rin
g compo
n
ents
used
du
rin
g
the
exper
im
e
nts (
the
g
r
ee
n
zo
n
e
in
Fig. 14
)
. Th
is
shaft
is geare
d
by
the
in
put
shaft.
Fin
a
ll
y,
the third shaft AS
is
n
amed as
the
output
shaft
a
nd
is
gea
r
ed by
the
Alshaft.
Th
e
motor
is
powered
by a
pulse
ge
n
e
r
ato
r
that
ad
j
usts
i
ts
speed
by vary
in
g
the rotating
freq
ue
n
cies
(
25
H
z,
35
H
z a
nd
45
H
zi
Th
e ac
hi
eved expe
rim
e
n
ts co
rr
espo
nd
to
fou
r
load levels
(
0
%
,
25
%
,
50
%
and 75
%)
at
different speeds. Regarding
the data
ac
q
uis
iti
o
n
part,
the current sensors
a
r
e
inst
a
ll
ed at
the
stator leve
l
a
nd
con
n
ected
to
a
n
acquis
iti
o
n
card
(
reference
9234
from
Nat
i
ona
l I
nstrument
)
to
r
eco
rd
the three phase currents.
The
r
ecorded
data
a
r
e stored
in
csv
files
by us
in
g
Matlab software.
Each file
conta
ins
10 s
o
f
the
cu
r
re
nt
s
i
g
n
a
l
sampled
at a
frequency
equa
l
to 25.6 kHz.
Thi
s
test bench is
dedicated
fo
r
bea
rin
g and gea
r
fault d
i
ag
n
os
ties.
It
is equ
ipped
with specified
compone
n
ts
r
ep
r
ese
nting
differ
Gear
Exp 29 Teeth 100 Teeth 36 Teeth (Al : AS) 90Teeth El Healthy Healthy Healthy Healthy
E2 Healthy Healthy Surface damage Healthy
E3 Healthy Healthy ½ Tooth break Healthy
E4 Healthy Healthy Healthy Healthy
ES Healthy Healthy Healthy Healthy
E6 Healthy Healthy Surface damage Healthy E7 Healthy Healthy 1/, Tooth break Healthy
ent states
such
as outer and
inn
e
r r
ace fau
l
ts
in
bea
rin
gs,
surface
damage
and
half tooth
b
r
eak
faults
in
gea
rs. T
he
Fig. 15
shows
the
components used
in
the
exper
im
enta
l
tests
with different
fa
il
u
r
e
types
cons
i
dered
in
this work.
I
n th
i
s
applicat
i
on,
seven
types o
f
exper
im
ents
that
cha
r
acte
ri
ze
seven health states
o
f
the motor were performed
o
n
the test
bench
.
These
expe
rim
e
n
ts are
summarized
i
n
Fig. 13
.
E
ach exper
im
e
n
t,
from
£1
to
£7,
required the
cha
n
ge o
f
the
corn
ponents mounted
between
the
Aland
ASshafts,
as
ill
ust
r
ated
in
Fig. 14
. From
the
Fig. 13
, one ca
n
see that
in
the first
exper
im
ent
£1, a
ll
the
compo
n
ents are
in
a
healthy state whereas
in
the second
expe
ri
ment £2 a
surface damage
is
present
in
the
gea
r
at
the
A/
shaft.
3.2.
Investigation
on
the
fault signatures using spectral analysis
Th
is
subsect
i
o
n
dea
l
s with the
ext
r
act
i
o
n
o
f
the characteristic
f
requencies co
rr
espond
in
g to
the
component
defects.
In
the
case
o
f
bea
rin
gs,
the
faults can be
distributed
in diff
e
r
e
nt
components
as
shown
in
Fig. 16
.
Th
e v
i
brat
i
on caused by
these
component
defects
(inn
e
r
race,
oute
r r
ace, cage and
rolling
ba
lls
)
a
ff
ect
the
cu
rr
e
nt
s
i
gna
l
s
by
pro
d
ucing
harmonic
freq
ue
nci
es
d
ue
to the rad
i
a
l
mot
i
on betwee
n
the
motor rotor
and
the stator. Each defect
ca
n
be
l
oca
li
zed
through
its
cha
r
acte
ri
st
ic
frequency
by
the following equat
i
on:
...
LegendAl :AS Al:AE
Healthy Healthy Healthy state
++
Healthy Healthy Gear surface damage
++
Healthy Healthy Gear ½ tooth break
++
Inner race Healthy Bearing inner race fault++
Ouwmce Healthy Bearing outer race fault
..
hmermce Healthy Combined fault (E2 + E4)++
Ouwrace Healthy Combined fault (E3+ ES)++
Supply source
TT
Pulsegenerator Squirrel cage
motor
Used for experiences Electromagneticj
Brake supply
,__
Three current .._ _ _ _ ----' Acquisition card (National Instrument 9234)
senSO<$
Fig. 14. Overall scheme of the test bench.
I : Healthy gear 2 : Gear surface damage 3 : Gear half tooth (
½
)
broken.. : Healthy bearing 6 : Bearing inner race fault 6 : Bea.ring outc.r race fault Fig. 15. Illustration of different component experiences.
Ball
---1-.,__,__,.,._
_
d
Inner race
--J1.
1":
::t::...=~_J
PDCa
ge
Fig. 16. Bearing components.(
13
)
where
f
,
represents the electrica
l
supp
l
y frequency, k
(
1
,
2
,
3 ...
)
is the harmonic n
u
mber generated by
t
he c
u
rrent signals and/
bcor
responds to the characteristic frequencies of the bearing e
l
ements
.
Th
e
spectra
l
a
n
a
l
ys
i
s co
n
ducted
h
e
r
ea
f
te
r
a
im
s to s
h
ow t
h
e
im
poss
i
b
ili
ty o
f
detect
in
g fau
lt
s by us
in
g t
h
e c
h
a
r
acte
r
ist
ic
fr
eque
n
ci
es ext
r
acted
fr
o
m
t
h
e reco
rd
e
d
cur
r
e
n
t s
i
g
n
als at d
if
ferent ope
r
at
in
g co
ndi
t
i
o
n
s
.
Howeve
r
, for
ill
ustrat
i
o
n
a
n
d
cl
arity
Table 1
Characteristic parameters of the bearing used for the experimental tests. Number of roUing elements, Nr
Contact angle, fi (0 )
Rotating frequency.[, (Hz)
Diameter of the rolling elements. d (inch)
Pitch diameter, PD (inch)
9 0
43.75
0.2762 0.7342
speed o
f
45 Hza
n
d a load leve
l
o
f
75%is co
n
sidered
in
t
h
is
applicatio
n
T
he c
h
aracteristic
f
reque
n
cies o
f
the bea
rin
g
inn
er
a
n
d oute
r
r
ace de
f
ects a
r
e
expressed by
E
q
.
(14)w
h
ereas
t
h
e
characteristic
pa
r
ameters o
f
t
h
e
bea
rin
g a
r
e
g
i
ve
n in
Table 1
.
I
nn
e
r rac
e
f
,
r
Out
e
r rac
e
for
/Y!_
[
1
d.cos(Rl
]
f
2 PD • r
t{r
[
1
+dcos(p)
]
f
2 PD • r
(14)
where Nr is the n
u
mber of rolling
elements,
d represents the rolling
elements d
i
ameter, PD is the pitch diameter,
fJ
represents
the con
tact angle
.
T
he above fau
l
t characte
r
istic
fr
eque
n
cies
(in
n
e
r r
ace a
nd
oute
r
r
ace defec
t
s) are calcu
l
ated by t
h
e
in
formatio
n
prese
n
ted
in
t
h
e
Table
1.T
he
r
efo
r
e,
the
inn
e
r r
ace
f
reque
n
cy is/
;,
276.319Hz,a
n
d the outer race
fr
eque
n
cy is f
o
r
125.231 Hz.T
he
Fig.
17s
h
ows t
h
e
spect
r
um o
f
t
h
e
cu
rr
e
n
t signa
l
in
t
h
e
cases
o
f
a hea
l
t
h
y a
n
d a fau
l
ty bea
rin
g
. F
rom t
h
ese
figu
r
es,
one ca
n n
otice
t
h
at t
h
e
mag
ni
tude at t
h
e
ha
r
mo
n
ic
f
reque
n
cies
o
f
t
h
e
defect bea
r
in
g is
diff
e
r
ent
w
h
en
compa
r
ed w
i
t
h
t
h
e
o
n
es
calcu
l
ated above
.
Th
erefore,
o
n
e
ca
n
co
n
clude t
h
at t
h
e
de
f
ects ca
nn
ot be detected
by us
in
g t
h
is tec
hn
i
q
ue
. T
o remedy to t
h
is s
i
tuatio
n
,
t
h
e
proposed
h
ea
l
t
h ind
icator
is applied
in
t
h
e
fo
ll
ow
in
g subsectio
n.
3.3. Health indicators construction
I
n
t
h
is applicatio
n
,
seve
n
expe
rim
enta
l
tests
(see
Fig. 13
)
were
pe
r
fo
rm
e
d
fo
r
d
i
ffere
n
t operat
in
g co
n
d
i
tio
n
s to acquire t
h
e
t
h
ree
p
h
ase cu
rr
ent
signa
l
s and
extract
h
ea
l
t
h
in
d
icators
. T
hese
in
dica
to
r
s a
r
e
t
h
e
n
used to
eva
l
uate t
h
e
pe
rf
o
rm
a
n
ce a
n
d the robust
n
ess
o
f
t
h
e
methodology proposed
in
Sectio
n
2
.
Firs
t, the performa
n
ce o
f
t
h
e
proposed hea
l
t
h i
ndicator R
M
S
/
St
D
is hig
h
lig
h
ted aga
in
st t
h
e
hea
l
t
h ind
icato
r
s us
in
g o
nl
y root mea
n
square
(R
M
S), va
r
ia
n
ce
(VAR)
a
n
d ku
rt
osis
(KU
R
)
va
l
ues, w
h
ic
h
a
r
e
t
h
e
most
employe
d
featu
r
es
in
t
h
e
li
teratu
r
e
. Th
us, seve
n cl
asses
correspo
n
d
in
g to t
h
e
m
o
t
ors bea
rin
g a
n
d gea
r
hea
l
t
h
states
(
cha
r
Fast Fourier Transform of current signal
-10
-20
m-
3o
~ a> -40 -0 ::,=
a. -50 E <( -60 -70 -80 -10 -20 ~ -30 m -0 °;'-40 -0 ::, ~-50E
<( -60 -70 -80100
200
)t4,l't V\4.o, 100 200 300 400500
600 Frequence (Hz)a)
O
ut
e
r rac
e
defect
700Healthy bearing
Outer race fault
800 900 1000
Fast Fourier Transform of current signal
- Healthy bearing
- Inner race fault
300 400 500 600 700 800 900 1000
Frequence (Hz)
b
)
I
nn
e
r ra
ce
d
efect
Health indicators construction at speed corresponding to 45 Hz with 75o/, load level 0.415 0.41
g
t110.405 :I!a:
0.4-• Et: Heahhy state
• El: Gear surfac• damage
• El: Gear 112 tooth bruit
0.395- ----:-:-;--- - - - : ~:---- - - - , ~ - - -
-0.395 0.4 0.405 0.41 0.415 0.39
0.41
RMS(ia) RMS(ib)
Fig. 18. Health indiGltors construction using RMS values.
Health indicators construction at speed corresponding to 45 Hz with 75¾ load level
g
..
.,,
> 0.17 0.165 0.16 -• El:H..ithystalo• E2: Gear twfact damage
• E3: Gear 112 tooth break
• E4: Bearing inMr race fluft
• ES: Staring outer 1'108 fault
H: E2>£4(Comb;nod ,._lb) • E6: El•E Combined faults
0.155 0.155-r---:::7 - ~ - - - r - - - r -- ,~ ----/ 0.156 0.158 0.16 0.15 0.16 0.162 0.164 0.166 0.168 var (lb) var (la)
Fig. 19. Health indiGltors construction using VAR. values.
Health indicators construction at speed corresponding to 45 Hz with 75% load level
1.57 -1.56 ;g:t55 Cll 'iii 1.54 -0 1: ::i .a: 1.53 1.52
-• Et: Healttly atate
• E:z: GNII' turfaet darn.age
• E3: GNII' 112 tooth btNk
+ e.t: S.al'ing iMtr race fault • ES: S.al'ing outer race flutt
E&: EZ+E4(Combinod f>ultl) • El: U+ES(Combinod f>ultl)
1.51
1
-
1-
, ----,---r--r- - . --.--
..-1.58 1.57 1.56 1.55 1.54 1.53
kurtosis (ib) 1.52 1.51
1.55
kurtosis (la)