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

net/publication/337408556

Fault Detection for Railway Traction Motor Bearing through Leakage Current

Article in Quarterly Report of RTRI · November 2019

DOI: 10.2219/rtriqr.60.4_256

CITATIONS READS

3 115

3 authors, including:

Minoru Kondo
Railway Technical Research Institute
47 PUBLICATIONS 369 CITATIONS

SEE PROFILE

All content following this page was uploaded by Minoru Kondo on 28 October 2020.

The user has requested enhancement of the downloaded file.

PAPER

Fault Detection for Railway Traction Motor Bearing through Leakage Current

Yo SAKAIDANI Minoru KONDO

Drive Systems Laboratory, Vehicle Control Technology Division

Ken TAKAHASHI
Lubricating Materials Laboratory, Materials Technology Division

Much research on detecting machinery faults in the early stage is being conducted for the
purpose of preventing failure and reducing maintenance effort simultaneously. In this paper,
a bearing fault detection method for a railway traction motor through leakage currents is pro-
posed. The proposed detection method combines octave band analysis and machine learning.
Experiments simulating abnormalities due to defective bearings were conducted and the ef-
fectiveness of the proposed method was verified. These experiments showed that the proposed
method can successfully detect failures in railway traction systems in relation to specific con-
ditions and that leakage current could potentially be used to detect bearing faults.

Keywords: fault detection, bearing fault, leakage current, octave band analysis, machine
learning, one-class classification, nearest neighbor

1. Introduction cal corrosion: bearing insulation and bearing conduction [8].

Failures in railway traction systems occasionally lead
to disastrous accidents and therefore periodic mainte-
nance is performed to keep railway vehicles safe. In order
to enhance the safety of railway traction systems, much
research is being carried out into condition monitoring for
vehicle equipment. With respect to general motor failures,
almost half originate from bearings [1] and therefore early
stage bearing fault detection is important.
A lot of condition monitoring methods relying on physi-
cal properties such as vibration, acoustic emission, tem-
perature and winding current have been proposed so far
[2]-[6]. They often use signal processing techniques and
some of them combine those techniques and machine learn-
ing. Among them, vibration are widely used in condition
monitoring system because of its reliability as a bearing
fault defect indicator. When applying vibration monitor-
ing to a railway traction system, a vibration sensor has to
be attached to the railway traction motor directly, which
exposes it to severe conditions. Moreover, a long signal line
that extends from an underfloor traction motor is needed.
Therefore, it is preferable to refer to physical properties
other than vibrations for monitoring the condition of rail-
way traction systems. This paper focuses on leakage cur-
rents, which can be acquired directly from the inverter that
provides power to the motor. No sensor has to be attached
to the motor.
Leakage current is the current that occurs in an un-
wanted current path under normal conditions [7]. One of
the causes for unwanted current is common mode voltage.
Inverters generate a common mode voltage, which causes
the common mode current (CMC) flowing from the mo-
tor windings to take an unintended path to the ground.
Current, which flows through the shaft and the bearings,
causes a discharge between the bearing parts and causes
deterioration in the contact surface of the bearing. The
damage from discharge to the bearing is known as electri-
cal corrosion. There are two approaches to prevent electri-
In Japan, electric corrosion of railway traction motor bear-
ings used to be a serious problem, but insulated bearings
were developed to prevent it [9].
Nowadays electrical corrosion is not a major issue
for railway traction motors in Japan because of insulated
bearings. However, insulated bearings may work as ca-
pacitors due to the insulation layer of the outer race. Due
to the capacitance and the number of voltage pulses from
the inverter, a small leakage current may flow through
the bearing. This means that there is a possibility that a
bearing fault could be detected from leakage current vari-
ance due to the bearing fault. In addition, research is be-
ing conducted into bearing fault detection using common
mode currents [10]. In this research, two bearing diagnosis
methodologies were proposed. The first method counts
the number of discharge pulses that exceed a prefixed
threshold value. According to the research, the amplitude
of CMC discharges decreases with faulty bearings so this
means the number of large pulses decreases as the bearing
deteriorates. Second method is to analyze the CMC in fre-
quency domains. The frequency of the highest peak shifts
to the low frequency side with bearing degradation so they
can be detected by monitoring the CMC peak shifts. In this
research, experiments were conducted and demonstrated
that a fault in a bearing affected the CMC somehow.
In parallel, a method for monitoring the condition of
equipment on diesel railway cars is also being developed.
The method combines octave band analysis of vibrations
and machine learning algorithms. Octave band analysis is
a frequency analysis method, which is often used for acous-
tic emissions and noise evaluations. In addition, a clas-
sification technique using machine learning was adopted.
In one-class classification, the detection of abnormalities
is conducted only by using normal condition data. The ef-
fectiveness of the method was verified through bench tests
and experiments using actual cars [11]-[15]. The method
has a general applicability so it was applied to detect bear-
ing faults through leakage current.

256 QR of RTRI, Vol. 60, No. 4, Nov. 2019

 Fig. 1 Schematic diagram of the fault detection method
Fig. 1 Schematic diagram of the fault detection method
applied to detect bearing faults through leakage current. In the case of fault detection, the diagnosis program learns
The following section introduces a bearing fault detec- components, the two must be plotted closely in this space.
The following section introduces a bearing fault detection what a normal condition is from the training data (normal data).
tion method combining octave band analysis and machine Therefore, normal data items are plotted closely together
method combining octave band analysis and machine learning After the training, test data is plotted in the multi-dimensional
learning [16]. Experiments were conducted by simulating as shown in Fig. 1 when normal condition signals are simi-
[16]. Experiments were conducted by simulating an inner race space and the distance between the test data (point X) and the
an inner race fault with a defective bearing. Section II de- lar and stable.
fault with a defective bearing. Section II describes the nearest training data (point A) is calculated. The test data is
scribes the diagnostic method. Section III describes the ex- In the case of fault detection, the diagnosis program
diagnostic method. Section III describes the experimental diagnosed as abnormal when the distance is longer than the
perimental procedure and results of the measurement and learns what a normal condition is from the training data
procedure and results of the measurement and diagnosis. reference distance. The reference distance is determined so that
diagnosis. Section IV discusses the results and the conclu- (normal data). After the training, test data is plotted in
Section IV discusses the results and the conclusion is given in 99% of the data are diagnosed as normal when the training data
sion is given in Section V. the multi-dimensional space and the distance between the
Section V. are diagnosed as the test data. The degree of abnormality is
test data (point X) and the nearest training data (point A)
defined below.
is calculated. The test data is diagnosed as abnormal when
2. Diagnosis method x  NN x the reference distance. The ref-
2. Diagnosis method the distance f is
x  longer
 than 1 (1)
erence distance is determined d so that 99% of the data are
For the For the bearing
bearing fault fault detection
detection fromfrom a leakage
a leakage current, diagnosed
current, a whereas x normal
and NN when are thethe coordinates
training data of thearetest data and the
diagnosed
method
a method combining
combining octave
octave bandband analysis
analysis and and one-class
one-class as thenearest point The
test data. among the training
degree data respectively,
of abnormality is definedand dbe-
is the
classification
classification was was
used.used.
TheThe method
method was wasoriginally
originallydevel-
developed low. reference distance. The degree of abnormality becomes positive
as aavibration monitoring system for railway diesel cars. when the data
oped as vibration
Faults in
monitoring
railway
system
traction
for railway
equipment
diesel
often produce x −isNN
diagnosed
( x ) as abnormal.
cars. ( x ) above
fThe = 1
value is−affected by false-positive(1) data in the
abnormal
Faults noise and
in railway are equipment
traction thus detected oftenbyproduce
train crews
ab- or training data. Therefore, d the average of the distances between
normal passengers
noise and unfamiliar
are thuswith train installations.
detected by train crews This or
implies
pas- that where
the testx point
and NN and aare certain
the number
coordinatesof nearest
of the points
test should
data be
special knowledge about traction gear is
sengers unfamiliar with train installations. This implies not required to detect used
and the as the point
nearest value among
representing the degree
the training dataof respectively,
abnormality. The
a problem
that special if one is about
knowledge aware traction
of how ageartrain is
normally sounds. In
not required and ddetail
is the ofreference
the refineddistance.
degree of The abnormality
degree of is described
abnormalityin [15]-
order
to detect to diagnose
a problem railway
if one traction
is aware equipment
of how a trainproblems
normally in the [16].positive
becomes The programwhen the wasdata created using a scikit-learn
is diagnosed as abnormal. machine
sounds.sameIn way humans
order can, octave
to diagnose band traction
railway analysis, equipment
which is similar The learning
above library
value [18].
is affected by false-positive data in
to human hearing, is used to detect vibrations.
problems in the same way humans can, octave band analy- the training Thisdata.method needs neither
Therefore, abnormalof data
the average the nor parameter
distances
In order to conduct a diagnosis amid the complex vibrating
sis, which is similar to human hearing, is used to detect between adjustments
the testwith point respect
and ato certain
the equipment.
numberThe abnormality
of nearest
environment of a railway traction system, classification
vibrations. pointssimulation
should beexperiments
used as thewere value conducted with the
representing the degree
engine test
technique in machine learning was used. For classification, a bench and itThe was shown
In order to conduct a diagnosis amid the complex vi- of abnormality. detail thatof thethe refined
diagnosisdegree
program of detected
abnor- the
classifier is trained with a sufficient amount of data for each fault [13]. This diagnostic method
brating environment of a railway traction system, classifi- mality is described in [15]-[16]. Thehas general was
program applicability,
created and
predefined group, so that it can classify the data correctly. When
cation technique in machine learning was used. For clas- usingtherefore a scikit-learn the same
machine method was applied
learning library for
[18].fault detection
using machine learning, sufficient data for respective classes are through leakage currentneither
in railway traction motor
sification, a classifier is trained with a sufficient amount of This method needs abnormal data bearings.
nor pa-
needed for correct diagnosis. However, data relating to failures
data for each predefined group, so that it can classify the rameter adjustments with respect to the equipment. The
in railway traction systems is hard to obtain. Thus in this study,
data correctly. When using machine learning, sufficient abnormality
a fault detection system was developed using a one-class 3. Experiment simulation simulating
experiments abnormality
were conducted with
data for respective classes are needed for correct diagno- the engine test bench and it was shown that the diagnosis
classification method [17], which only uses normal data.
sis. However, data relating to failures in railway traction program detected theoffault
3.1 Conditions the [13].
experimentThis diagnostic method
For the detection of abnormalities in diesel railway cars,
systems is hard to obtain. Thus in this study, a fault detec- has general applicability, and therefore the same method
the one-class classification method was designed, which is
tion system was developed using a one-class classification was applied Thefor fault detection
experiment simulating through leakage current
an abnormality in
was conducted
based on the nearest neighbor principle. The method works as
method [17], which only uses normal data. railway
to traction
verify the motor
proposedbearings.
diagnosis method. Currents leakages in
follows:
For the detection of abnormalities in diesel railway the motor were measured on normal and abnormal bearings
Signals are measured and 1/12 octave band analysis is
cars, the one-class classification method was designed, under various conditions. Both bearings were insulated. The test
applied to those signals every second. Next, the dimensions of
which is based on the nearest neighbor principle. The 3. Experiment equipment issimulating
shown in Fig.abnormality
2. Two induction motors that had
the analyzed data are reduced by principle component analysis
method works as follows: been mounted in a train were used. One motor was used as a test
(PCA). The number of principle components is set to 6. Then,
Signals are measured and 1/12 octave band analysis is 3.1 Conditions motor and sensors of thewere experiment
attached to it. The other one was used
preprocessed data is plotted in a multi-dimensional space. If two
applied to those signals every second. Next, the dimensions as a load, which generated torque against the traction motor.
pieces of data have similar frequency components, the two must
of the analyzed data are reduced by principle component The experiment
Their output simulating an abnormality
shafts were connected was through
to each other con- a
be plotted closely in this space. Therefore, normal data items
analysis (PCA). The number of principle components is set ducted to verify
torque gauge theandproposed
the diagnosis
torque was method.
measured Currents
during the
are plotted closely together as shown in Fig. 1 when normal
to 6. Then, preprocessed data is plotted in a multi-dimen- leakages in the The
experiment. motor test were
motor measured
was cooled on normal
with a forcedand ab-
convection
condition signals are similar and stable.
sional space. If two pieces of data have similar frequency normal bearings under various conditions. Both bearings

QR of RTRI, Vol. 60, No. 4, Nov. 2019 257

 Torque
Torquegauge
gauge Test
Torque Testmotor
motor
Load
Load Torquegauge
gauge Test
Testmotor
motor
Load
motor
Load Torque gauge Test motor
motor
Load
motor
motor
motor

Fig.
Fig.22 Experimental
Experimentalsetup
setup
Fig.
Fig.2 22Experimental
Fig. Experimental
Experimental setupsetup
setup
Fig. 2 Experimental setup

1.5
1.5 No-load
No-load
1.5
1.5 No-load
No-load
Load
1.5 Load
No-load
Load
Load
Load

Fig. Rotation
Rotation(rpm)(rpm)
Fig.3.3. Leakage
Leakagecurrent
currentsensor.
sensor. Rotation (rpm)
Fig.
Fig.3.3. Leakage
Leakagecurrent
currentsensor.
sensor. Fig. 66 Load torque Rotation
versus (rpm)
rotational speed.
Fig. 3Fig.Leakage
3. Leakage current sensor
current sensor. Fig. Fig. 6 Load torqueRotation
Fig.
Fig.6 6 Load
Load torque
Loadtorque
versus (rpm)
rotational speed. speed
versus rotational
torqueversus
versusrotational
rotationalspeed.
speed.
Fig. 6 Load torque versus rotational speed.
Table
Table II
Table
Test
1Table
TestI Icondition
condition
Test Table
condition
Load-side
Load-side Load
Test
Test
Load Table
conditionIRotational
condition Rotational Load
Load
Load-side Load Rotational Load
bearing
Load-side
bearing
Load-side Test
Loadcondition
condition
Load
condition speed[rpm]
Rotational
speed[rpm]
Rotational [Nm]
Load
[Nm]Load
bearing condition speed[rpm] [Nm]
Load-side
bearing
bearing Load
condition
condition Rotational
speed[rpm]
900
speed[rpm]
900 Load
[Nm]
[Nm]
bearingst condition 900
speed[rpm] [Nm]
1800
900
1800
900
Normal(1
Normal(1 ) ))
stst 1800
Normal(1 Full-load
Full-load 900
2700
1800
2700
1800 Shown
Shown
Full-load 2700 Shown
nd
Normal(2
Normal(1
Normal(2
Normal(1 ndst)st)
nd))
Normal(2 No-load
Full-load
No-load
Full-load 1800
3600
2700
3600
2700
inFig.
Shown
in Fig. 6
Abnormal
Normal(1
Normal(2
Abnormal
Normal(2 ndndst)
) ) No-load 3600
4500 inShown
Fig.66
Abnormalnd Full-load
No-load
No-load 2700
3600
4500
3600 in Shown
Fig.
in Fig.6 6
Normal(2
Abnormal )
Abnormal 4500
5400
No-load 3600
4500
5400
4500 in Fig. 6
Abnormal 5400
4500
5400
5400
justed to 20 m3/min. Thermocouples were installed in the
5400
bearing
bearingto
motor (showncheckinin
(shown Fig.
the Fig. 4)4)was
temperaturewasinstalled
installedandtoto thetheload-side.
prevent load-side.
overheating.After
After
Fig.
Fig.4.4. Normal
Normalbearing
bearing(insulated).
(insulated). bearing
the bearing
the (shown
installation, (shown
installation, inthe
theinFig.Fig.
traction4)4)was
traction was installed
motor
motor installed
was
was to tothe
driven
driven theaload-side.
load-side.
for
for 33 hours
hoursAfter
After
toto
Fig.
Fig.4.4. Normal
Normalbearing
bearing(insulated).
(insulated). Thebearingleakage (shownthe currentin Fig. was measured
4) was installed usingto the for current
load-side. sensor
After
thethe
stabilizeinstallation,
stabilize installation,
the
the grease
grease the
inintraction
traction
the bearing.
the bearing. motor
motor was
was driven
driven for 3 3 hours
hours toto
Fig.Fig.
4 4.Normal
Normalbearing (insulated)
bearing (insulated). (Pearson
the 301X)
installation, asthe shown
traction in motor
Fig. 3.was driven for 3 hours to
stabilize
stabilize
To the grease
the grease
To simulate in
a in the bearing.
the bearing.
bearing abnormality,
The
stabilize simulate greaseaaain
experimental
the bearing
theconditionsabnormality,
bearing. are ashown a bearing
bearing in with
with
Table an
an1.
artificial
artificialToTofault simulate
simulate
fault was
was made. bearing
made. bearing abnormality,
Simulating
Simulatingabnormality,
aalocal
local adefect,
abearing
bearing
defect, the
the with
withanan
surface
surface
First, an
Tofault experiment
simulate with
a bearing a normal
abnormality, bearing was
adefect,
bearing conducted.
with an
Flat artificial was made. Simulating a ait
local defect, the assurface
Flatplane
plane of artificial
the
Beforetheinner
ofartificial inner fault
race
starting race
fault
was
was
was the made.
machined
machined
made.
Simulating
experiments, totomake
Simulating make the
a
local
partially
itlocal
partially
load-side
defect,
flat
flat the
as
and
the
surface
shown
shown
anti-
surface
Flat
Flatplane
plane of
ininofthe
ofFig.
the 5.
Fig.
load-side
inner
inner
5. Then race
Then
bearings
was
racethe was
the machined
machined
bearing
bearing
of the was
to
wasto
test
make
tocleaned
cleaned
motor
it
make itwere partially
partially
and
and new new flat
flat
grease
removed.
as
grease shown
as shown
was
was
Then
Flat plane in in
packed
the
Fig.
Fig. inner
5. 5.
into Then racethe
Then
it.
was
Afterthe machined
bearing
bearing
completion was
was make
cleaned
cleaned
of the
it and
partially
and
experimentnewnew flat
grease as shown
grease
with waswas
the
packed
thein cleaned
Fig. into
5. Thenit. After
anti-load-side
the completion
bearing bearing
was of the experiment
was reinstalled
cleaned and new with
carefully
grease the
was
packed
packed
normal
normal into
intoit.new
bearing,
bearing, it.After
After
the
the completion
completion
load-side
load-side bearing ofofthe
bearing the
was
wasexperiment
experiment
replaced
replaced with
withthe
with
with the
the
the
andpacked a brand into it. the bearing
After (shown
completion in
of the Fig. 4) was
experiment installed
with theto
normal
normal
defective
defective bearing,
bearing,
bearing
bearing the
and
and load-side
load-side
the
the bearing
experiment bearing
experiment was
was was
was replaced
replaced
conducted
conducted with
with
after
afterthethe
aa
the load-side.
normal bearing, After the the installation,
load-side bearingwas the traction
wasconducted
replaced with motor was
defective
defective
running
running
driven in.bearing
in.
for
bearing
3 hours and
and and thetheexperiment
to stabilize
experiment
the grease
was conducted afterthe
after aa
defective
running
running in. in. bearing the experiment wasinconductedthe bearing. after a
For
To normal
Forsimulate
normal and
and abnormal
abnormal
a bearing bearings,
bearings,the
abnormality, theaexperiments
experiments
bearing with were
were an
runningFor in.
Fornormal normal and and abnormal
abnormal bearings,
bearings, thetheexperiments
experiments were
were
conducted
conducted
artificial atat
fault six
six different
wasdifferentmade. speeds
speeds with
Simulatingwith full-load
full-load
a local and
and no-load.
no-load.
defect, the
conducted
conducted For normal
at six
at six the and
different abnormal
different speeds bearings,
with
speedsgenerated the
full-load
with full-load experiments
andand no-load.were
no-load.
With
With
surface thetheof full-load,
full-load,
the inner load
therace loadwas motor
motor generated
machined totorque
torque
makeand against
against the
it partially the
Fig. conducted
With
With thethefull-load,at six
full-load, different
the
theload load speeds
motor
motor with full-load
generated
generated torque
torqueagainst no-load.
against the
the
Fig.5.5. Artificial
Artificialinner
innerrace
racefault.
fault. test
test
flat motor.
motor.
as shown The
The in torque
torque
Fig. 5.ofof the
Then the load
load
the motor
motor
bearing depended
depended
was cleaned onon the
the
and
Fig.
Fig.5.5. Artificial
Artificialinner
innerrace
racefault.
fault. With
test
testmotor. the
motor. full-load,
TheThe torquethe
torque load
ofof motor
the
theload loadgenerated
motor
motor torque
depended
depended against
ononthe the
the
cooling rotational
rotational speed
speed of
of the
the test
test motor
motor as
as shown
shown inin Fig.
Fig. 6.6. Without
Without aex-
a
cooling system with5 5.the air
air volume adjusted toto 20
20 mm/min.
33
Fig.
system Fig.
with Artificial
Artificial
the inner
volume race
inner race
adjustedfault
fault. /min. newtest
rotational grease
motor. speed was
Theofof packed
torque
the test into
of
motortheit.as Aftermotor
load completion depended of theon the
cooling
coolingsystemsystemwith withthe air
airvolume adjusted rotational
load, the load speedmotor the
did test
not motor
generate asshown
shown
torque inand
inFig.
Fig.
the 6.6. Without
testWithout
motor aa
motor to to20 mm/min.
3
Thermocouples
Thermocouples were
were the
installed
installed volume
inin thethe adjusted
motor toto 20
check
check
3
/min.
3 the
the load,
periment
rotationalthe load
with
speed motorthe
of the did
normal not
test generate
motorbearing, torque
as shown the and the
load-side
in Fig. test
6.test motor
bearing
Without a
cooling
Thermocouples system with
were the air
installed volume
in the adjusted
motor to
to 20
check m /min.
the load,
load,the
generated
generated theload load
small
small motor
motor
torque
torque did totonot
did not generate
maintain generate
maintain the torque
torque
rotational
thebearing
rotational and
and thethe
speed.
speed. test motor
InIn motor
each
each
Thermocouples
temperature
were insulated.
temperature and
and were
prevent
The
preventtestinstalled
overheating.
equipment
overheating. in the
The
is
The motor
leakage
shown
leakage in to check
current
Fig.
current 2.was
wasthe
Two wasload, replaced
the load with
motor the
did defective
not generate torque and
and thethe test experi-
motor
Thermocouples
temperature and were installed in The
the motor tocurrent
checkwas the generated
generated
case, the
the motorsmall
smallwas torque
torque to maintain
to maintain
accelerated the
from rotational
the arotational
a standstill, speed.
speed. In
continued each
In each
temperature
measured
induction
measured using
usingand
motors aprevent
a prevent
current
that had
current overheating.
overheating.
sensor been
sensor (Pearson Theleakage
mounted
(Pearson leakage
301X)
301X) in a asascurrent
shown
train
shown was
inin
were case,
ment
generated
case, was motor
conducted
small was was
torque accelerated
after a running
to maintain from in.
the rotationalstandstill, speed. continued
In each
temperature
measured anda aprevent overheating. The301X)
leakage current was case,the
rotating the
at motor
themotor
specified wasaccelerated
accelerated
rotational from
from
speed afor
for astandstill,
standstill,
180 seconds continued
continued
at least
used.
Fig.3.3.Oneusing
measured
Fig. using
motor current
wascurrent
used sensor
sensor
as a test(Pearson
(Pearson
motor 301X) asas
and sensors shown
shown wereinin rotating
case,
rotating Forthe at the
normal
motor
atstopped.
the
specified
and rotational
abnormal
was accelerated speed
bearings,
fromfor a 180 180 the seconds
standstill, at
experiments least
continued
measured
Fig.
Fig.3.3. using a current sensor (Pearson 301X) as shown in and rotating
and then
then at thespecified
stopped. specified
First,
First,
rotational
rotational
the
the “1st
“1st
speed
speed
Normal
Normal for
run”
run” 180 seconds
seconds
experiment
experiment
atatleast
wasleast
was
attachedThe experimental
to it. The conditions
other one are
was shown
used
The experimental conditions are shown in Table I. First, in
as Table
a load, I. First,
which were
androtating
thenconducted
atstopped.
the specified
stopped. at
First, six different
rotational
the “1st speeds
speed for
Normal run” with
180 full-load
seconds atwas
experiment and
least
Fig. The
3. experimental
The torque
experimental conditions
conditions are shown
arewas
shown in Table I.
in TableBefore First,
I. First, and
conducted
conducted then using
using athea brandFirst,
brand newthe
new “1st
bearing.
bearing. Normal
Then,
Then, run”
a “2nd
amotor experiment
“2nd Normal
Normal was
run”
run”
anan experiment
generated
experiment with
with aa normal
againstnormal thebearing
traction
bearing was conducted.
motor.
conducted. Before no-load.
and
conducted then With
stopped.
using a abrand full-load,
First, the
new “1st the
Normal load run” experimentgenerated was
an The experimental
anexperiment with conditions
a anormal bearing arewas
shown in Table Before
conducted. I. First, was conducted
was performed
performed usingusing
using brand
the
thesame newbearing.
same bearing.
bearing
bearing Then,
forThen,
for theasame
the a“2nd
“2nd Normal
Normalrun”
experiment. run”
experiment
starting
Their
starting
an
the
the
experiment
withshafts
experiments,
output
experiments, normal
with a normal
the
were
the bearing
load-side
load-side
bearing
wasand
connected conducted.
and
wasand
to each Before
anti-load-side
anti-load-side
conducted.
other
Before
torque
wasconducted
was
against
performed
performed using the
using a brand
using
test
thethe samemotor.
new
same bearing Thefor
bearing.
bearing
torque
Then,
for thethe asame
sameof the
“2nd
same
experiment.
load mo-
Normal
experiment.
experiment. run”
starting
starting
bearings
through the
ofathe
the experiments,
experiments,
test
torque motor
gauge were the
the
and load-side
load-side
removed.
the Then
torque and
the
wasanti-load-side
anti-load-side
cleaned
measuredanti- After
After
tor
this,
this,
depended
anan “abnormal
“abnormal
on thethe
run”
run”
rotational
was
was performed
performed
speed
with
with the
the defective
defective
bearings
starting of thethe test motor
experiments, were theremoved.
load-side Then andthe cleaned anti-
anti-load-side was
Afterperformed
After this,
this,anan“abnormal using
“abnormal same
run”
run” was bearing
was performed forof
performed
the
thewith same
with
test
thethe
motor as
experiment.
defective
defective
bearings
bearings
load-side
during of
the of the
the
bearing test
testmotor
was
experiment. motor were
were
reinstalled
The removed.
removed.
test carefully
motor Then
Thenand
was the
theacleaned
cleaned
brand
cooled anti-
new
withanti-a
bearing.
bearing.
shown in Fig. 6. Without awasload, the load motor did not
load-side
bearings bearing
ofbearing was
the testwas motorreinstalled
were removed. carefully
Then and a brand
the cleaned new
anti- After
bearing.
bearing. this, an “abnormal run” performed with the defective
load-side bearing reinstalled The leakage current was measured using the developed
load-side
forced convection was
cooling systemcarefully
reinstalled carefully
with the and aira avolume
and brand
brandnew new
ad- generate
bearing. The
The
leakage
torque andcurrent the was
test measured
motor using
generated the developed
small torque
load-side bearing was reinstalled carefully and a brand new Theleakage leakagecurrent currentwas wasmeasured
measuredusing usingthe thedeveloped
developed
The leakage current was measured using the developed

258 QR of RTRI, Vol. 60, No. 4, Nov. 2019

 Table 2 Specifications of the condition monitor
Item Specification
Input signals Leakage current
Sampling 12.5 kHz
Octave frequency
band Analysis 1 Hz
analysis frequency
Frequency range 10 Hz to 5 kHz
Band width 1/12 octave
RMS values
Recorded data
Analysis results
Recording frequency 1 Hz
to maintain the rotational speed. In each case, the motor
was accelerated from a standstill, continued rotating at the
specified rotational speed for 180 seconds at least and then
stopped. First, the“1st Normal run”experiment was con-
ducted using a brand new bearing. Then, a“2nd Normal
run”was performed using the same bearing for the same
experiment. After this, an“abnormal run”was performed
with the defective bearing.
The leakage current was measured using the devel-
oped condition monitoring system. The sampling frequency
for monitoring was 12.5 kHz and octave band analysis was
applied to the measured signals every second. The specifi-
cations of the condition monitoring are shown in Table 2.
The recorded and processed data were saved automatically
in an external USB device when the specified input value
exceeded the preset threshold values.
3.2 Measurement results
The condition monitor saves the root mean square
(RMS) values and octave band analysis data from the mea-
sured values. This section presents the octave band analy-
sis data from each experiment.
The octave band analysis results of the leakage cur-
rent are shown in Fig. 7. Each colored domain in the graph
represents the range in fluctuation of magnitude during
the 180 seconds for each running condition. The blue and
orange lines represent no-load and the green and red lines
represent the full-load condition. The blue and orange
bands are results of the normal bearing and the green and
red bands are that of the abnormal bearing. The X and Y
axes show the frequency of the leakage current and the
root-mean-square value of the leakage current, respectively.
Under a full-load, some peaks were found. The highest
peak came from the inverter current frequencies, which
corresponds to the product of rotational speed and the pole
pairs of the induction motor. The effect of the bearing fault
on leakage current was small, and two band graphs over-
lapped at almost all the frequencies.
With no-load and during the PWM mode (900 and 1800
rpm), the two band graphs overlap and the effect of the
bearing fault was small. However, when the inverter shifted
to the rectangular wave mode, the magnitudes of frequency
components (especially more than 1000 Hz) changed obvi-
ously. The bearing fault increased the high frequency com-
ponents of the leakage current by more than 10 dB.
From the octave band analysis results, the effect of
the bearing fault on the leakage current frequencies were
shown especially for the no-load condition and during the
rectangular wave mode.
QR of RTRI, Vol. 60, No. 4, Nov. 2019
3.3 Diagnosis results
This section evaluates the performance of the diagnosis
based on leakage currents, using measured data. Training
and normal test data were created from“1st Normal”and
“2nd Normal”experimental data. Data corresponding to
the same load and speed conditions were mixed and then
separated randomly into training data and normal test
data. The diagnostic program studied“normal conditions”
based on training data and diagnoses were conducted us-
ing normal and abnormal test data.
From the octave band analysis results, it was shown
that impact of a fault differed with frequency. As shown in
Fig. 7, the effect of a fault on a low frequency component
(less than 100 Hz) is lower than on a high frequency com-
ponent. In this paper, only the high-frequency band (more
than 1000 Hz) are used for the diagnosis.
The diagnosis results obtained using the high-frequen-
cy band (1000-5000 Hz) are shown in Fig. 8. Abnormalities
in the normal test data had negative values under all con-
ditions and correct diagnoses were obtained for the normal
bearing. In the case of abnormality detection with a full
load, the diagnosis program only detected a fault at 3600
rpm and the calculated abnormalities were small. In the
case of no-load, the diagnosis program did not detect any
fault when the inverter was operated in PWM mode. How-
ever, when the inverter shifted to the rectangular wave
mode (more than 3600 rpm), all the abnormalities became
positive values and the diagnosis program successfully de-
tected the fault in the bearing. Abnormalities increased in
accordance with the speed and it was confirmed that the
bearing fault was detected clearly when there was no load
during the rectangular wave mode.
4. Discussion
A comparison of measurements and results of the diag-
noses with previous work [10], led to the following findings:
Different fault types could affect the leakage cur-
rent differently and the impact of each defect on the
leakage current was unclear.
According to previous work, the capacitance of the bear-
ing decreases as the bearings and their lubricant deteriorate.
As shown in Fig. 5, the fault type in this study was a single
point defect and the high frequency components of the leak-
age current increased when there was no load during the
rectangular wave mode. This suggests the likelihood that ca-
pacitance increased due to the bearing fault. Although strict
comparisons are not possible because the experimental con-
ditions were not identical, different bearing fault types could
have a different impact on the leakage current.
The operating conditions of the motor such as
load and inverter operation mode, could also affect
the leakage current.
The degree of abnormality calculated from the high fre-
quency components of the leakage current grew when the
motor was operated with no load during the rectangular
wave mode. Moreover, the high frequency components did
not change with a full load or in PWM mode. The operating
conditions of the motor could influence the leakage current.
In previous work [10], variations were found in the leakage
259
 Increase

Increase

Fig. 7 Octave band analysis results (leakage current)

Fig. 7 Octave band analysis results (leakage current)

Fig. 7 Octave band analysis results (leakage current)

Fig. 8Fig.Diagnosis
8 Diagnosisresults
results byby leakage
leakage currents
currents

current but the operating conditions of the motor were not experiment. From the average values, it was deduced that
the effectiveness
described in detail.of the proposed method. Fig. 8 Diagnosis results by frequency
the bearing
leakage
components
fault decreased
currents
of the the
leakage current especially
impedance when
of the bearing
Using the newly developed condition monitoring system, and there was no load in the rectangular wave mode.
increased the leakage current values. The octave band
the average values and results of octave band analyses analysis Abnormalities were then
results revealed that diagnosed
the bearing through octave band
fault affected the
the effectiveness
conducted
5. Conclustion on of the
leakage proposed method.
currents, were recorded during the frequency components of the leakage current
high frequency components of the leakage currentUsing
analysis, followed by one-class especially
classification. when the
especial-
Using theFrom
experiment. newly thedeveloped condition
average values, monitoring
it was deducedsystem, there
that the ly was
diagnosis no load
program,in the
the rectangular
degree of wave mode.
abnormality
when there was no load in the rectangular wave mode. in normal and
the average
bearing
In thisfaultvalues
decreased
paper, and detection
a fault results
the ofmethod
impedanceoctave bandbearing
ofbased
the onanalysesand
leakage Abnormalities
abnormal test data were
Abnormalities was then diagnosed
werecalculated.
then Using
diagnosedthrough octave
thethrough
high band
frequency
octave
conducted
increased
current was
onproposed
the leakagecurrent
leakage currents, were
values.
and experiments Therecorded
octave during
band
simulating
the
analysis analysis, followed
bandsanalysis,
abnor- band of the leakage by one-class
current,
followed by the
classification.
diagnosis
one-class program wasUsing
classification.
the
capable
Using
experiment.
results From
revealed the
thataverage
the values,
bearing it was
fault deduced
affected that
the the
high diagnosis
of program,
detecting the the
bearing degree
fault of abnormality
correctly
malities were carried out using a bearing with an artificial the diagnosis program, the degree of abnormality in nor- when in
there normal
was no and
load,
bearing fault decreased the impedance of the bearing and
fault, to verify the effectiveness of the proposed method.
abnormal test data was calculated. Using the high frequency
mal and abnormal test data was calculated. Using the high
increased the leakage current values. The octave band analysis bands of the leakage current, the diagnosis program was capable
Using the newly developed condition monitoring sys- frequency bands of the leakage current, the diagnosis pro-
results revealed that the bearing fault affected the high of detecting the bearing fault correctly when there was no load,
tem, the RMS values and results of octave band analyses gram was capable of detecting the bearing fault correctly
conducted on leakage currents, were recorded during the when there was no load, during the rectangular wave

260 QR of RTRI, Vol. 60, No. 4, Nov. 2019

mode. These results verified that leakage current could po-
tentially serve as an index for abnormalities in the form of
bearing faults, under specific conditions. Leakage current
may however be influenced by other components, which
were not included in the present experiments. This means
that the mechanisms underlying leakage current variance
have to be clarified.
References
[1] Bonnet, A. H., and Yung, C.,“Increased Efficiency
and Increased Reliability,”IEEE Industry Application
Magazine, Vol. 14, pp.29-36, Jan. 2008.
[2] Marquez, F. P. G., Tobias, A. M., Perez, J. M. P., and
Papaelias, M.,“Condition monitoring of wind turbines:
Techniques and methods,”Renewable Energy, Vol. 46,
pp.169-178, Oct. 2012.
[3] Knight, A. M., and Bertani, S. P.,“Mechanical Fault
Detection in a Medium-Sized Induction Motor Using
Stator Current Monitoring,”IEEE Transactions En-
ergy Conversion, Vol. 20, pp. 753-760, Dec. 2005.
[4] Mathew, J., and J. Alfredson, R.,“The Condition Moni-
toring of Rolling Element Bearings Using Vibration
Analysis,”Journal of Vibration, Acoustic, Stress, and
Reliability in Design, Vol. 106, pp. 447-453, July. 1984.
[5] Stack, J. R., Habetler, T. G., and Harley, R. G.,“Fault
Classification and Fault Signature Production for
Rolling Element Bearings in Electric Machine,”IEEE
Transactions Industry Applications, Vol. 4, pp. 735-
739, May 2004.
[6] Blodt, M., Granjon, P., Raison, B., and Rostaing, G.,
“Models for Bearing Damage Detection in Induc-
tion Motors Using Stator Current Monitoring,”IEEE
Transactions on Industrial Electronics, Vol.55, no.4,
pp.1813-1822, 2008.
[7] International Electrotechnical Commission,‘Voltages
and currents, leakage current,’IEC 60050 826-11-20.
[8] Suzumura, J.,“Prevention of Electrical Pitting on Roll-
ing Bearings by Electrically Conductive Grease,”QR of
RTRI, Vol.29，No.4，pp.29-34，2015.
[9] Watanabe, T.,“Electric Corrosion of Traction Motor
Bearings and its Prevention,”RTRI report, Vol.7, No.4,
Authors
Yo SAKAIDANI, Dr. Eng.
Researcher, Drive Systems Laboratory,
Vehicle Control Technology Division
Research Areas: Traction Motor
Minoru KONDO, Dr. Eng.
Laboratory Head, Drive Systems Laboratory,
Vehicle Control Technology Division
Research Areas: Traction Motor, Condition
Monitoring
QR of RTRI, Vol. 60, No. 4, Nov. 2019
View publication stats
pp.27-32, 1993 (in Japanese).
[10] Delgado, M., Garcia, A., Ortega, J. A., Urresty, J.,
and Riba, J. R.,“Bearing Diagnosis Methodologies by
means of Common Mode Current,”presented at the
13th European Conference on Power Electronics and
Applications, Barcelona, Spain, Sept. 8-10, 2009.
[11] Kondo, M., Manabe, S., Takashige, T., and Kanno,
K.,“Traction Diesel Engine Anomaly Detection Us-
ing Vibration Analysis in Octave Bands,”QR of RTRI,
Vol.57, No.2, pp.105-111, 2016.
[12] Kondo, M., Takashige, T., Manabe, S., and Kanno, K.,
“Abnormality Detection in a Contaminated Diesel En-
gine by a Vibration Monitoring Method,”RTRI report,
Vol.30, No.4, pp.47-52, 2016 (in Japanese).
[13] Nishitani, K., Kondo, M., Takashige, T., Kataoka, Y.,
and Noguchi, K.,“Application of the Condition Moni-
toring System by Vibration Analysis to Driving De-
vices on Railway Vehicles in Operation,”RTRI report,
Vol.32，No.8，pp.23-28，2018 (in Japanese).
[14] Nishitani, K., Kondo, M., Takashige, T., Kanno, K.,
“Application of the Condition Monitoring System by
Vibration Analysis to Driving Devices on Railway Vehi-
cles in Operation,”RTRI report, Vol.30，No.11，pp.35-
40，2016 (in Japanese).
[15] Kondo, M., Sakaidani, Y., and Takahashi, K.,“Ab-
normal Bearing Detection for Traction Motors Using
Vibration Octave Band Spectra,”presented at the Dy-
namics and Design Conference 2018, Tokyo, Japan,
August 28-31, 2018, (in Japanese).
[16] Sakaidani, Y., Kondo, M.,“Bearing Fault Detection for
Railway Traction Motors Through Leakage Current,”
presented at the International Conference of Electrical
Machines, Greece, Oct. 3-6, 2018.
[17] Tax, D.,“One-class classification,”Ph.D. thesis, Delft
University of Technology, 2001.
[18] Pedregosa, F., Varoquaux, G., Gramfort, A., Michel,
V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer,
P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
Cournapeau, D., Brucher, M., Perrot, M. and Duch-
esnay, E.,“Scikit-learn: Machine Learning in Python,”
Journal of Machine Learning Research, Vol.12, pp.
2825-2830, 2011.
Ken TAKAHASHI
Senior researcher, Lubricating Materials
Laboratory, Materials Technology Division
Research Areas: Roller Bearing, Mechanical
Engineering
261