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

net/publication/305033264

Condition Monitoring of Bearing Damage in Electromechanical Drive Systems by

Using Motor Current Signals of Electric Motors: A Benchmark Data Set for Data-
Driven Classification

Conference Paper · July 2016

CITATIONS READS
193 4,626

4 authors:

Christian Lessmeier James Kimotho

Universität Paderborn Jomo Kenyatta University of Agriculture and Technology
6 PUBLICATIONS   217 CITATIONS    13 PUBLICATIONS   294 CITATIONS   

SEE PROFILE SEE PROFILE

Detmar Zimmer Walter Sextro

Universität Paderborn Universität Paderborn
54 PUBLICATIONS   819 CITATIONS    196 PUBLICATIONS   1,650 CITATIONS   

SEE PROFILE SEE PROFILE

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

Energy Harvesting View project

Modular electromechanical drive concepts for industrial applications View project

All content following this page was uploaded by Christian Lessmeier on 08 July 2016.

The user has requested enhancement of the downloaded file.

 Condition Monitoring of Bearing Damage in Electromechanical Drive
Systems by Using Motor Current Signals of Electric Motors: A
Benchmark Data Set for Data-Driven Classification
Christian Lessmeier1, James Kuria Kimotho2, Detmar Zimmer3 and Walter Sextro4
1,3
Chair of Design and Drive Technology, Faculty of Mechanical Engineering,
Paderborn University, Pohlweg 47 – 49, 33098 Paderborn, Germany
christian.lessmeier@uni-paderborn.de
detmar.zimmer@uni-paderborn.de
2,4
Chair of Mechatronics and Dynamics, Faculty of Mechanical Engineering,
Paderborn University, Pohlweg 47 – 49, 33098 Paderborn, Germany
james.kuria.kimotho@uni-paderborn.de
walter.sextro@uni-paderborn.de

ABSTRACT
1. INTRODUCTION
This paper presents a benchmark data set for condition
monitoring of rolling bearings in combination with an According to statistics, 40-70% of electro-mechanic drive
extensive description of the corresponding bearing damage, systems and motor failures are caused by rolling bearing
the data set generation by experiments and results of data- damages, which can lead to high costs in applications
driven classifications used as a diagnostic method. The because of downtimes (Bonnett & Yung, 2008; Djeddi,
diagnostic method uses the motor current signal of an Granjon, & Leprettre, 2007). Thus, high-risk applications or
electromechanical drive system for bearing diagnostic. The those with high maintenance costs are continuously
advantage of this approach in general is that no additional monitored. Detection of bearing damage is typically
sensors are required, as current measurements can be monitored by vibration analysis using acceleration sensors.
performed in existing frequency inverters. This will help to These additional acceleration sensors are widely utilized,
reduce the cost of future condition monitoring systems. A especially in large applications such as wind power turbines
particular novelty of the present approach is the monitoring or cement mills. In industrial applications which use a great
of damage in external bearings which are installed in the number of inexpensive small electric motors with an
drive system but outside the electric motor. Nevertheless, approximate power consumption of around 1 kW or less, the
the motor current signal is used as input for the detection of cost for additional sensors is financially not feasible. For
the damage. Moreover, a wide distribution of bearing this reason, several research projects focus on detecting
damage is considered for the benchmark data set. The bearing damage by using existing signals such as motor
results of the classifications show that the motor current currents, which can be measured by the already existing
signal can be used to identify and classify bearing damage frequency inverters. Researchers prefer data-driven
within the drive system. However, the classification classification methods that are based on machine learning
accuracy is still low compared to classifications based on algorithms to detect damage states by using the motor
vibration signals. Further, dependency on properties of those current signals (MCS). These methods are often examined
bearing damage that were used for the generation of training in case studies of special applications and damages such as
data are observed, because training with data of artificially broken rotor bars or bearing damages. The condition
generated and real bearing damages lead to different monitoring (CM) methods based on MCS are still being
accuracies. Altogether a verified and systematically investigated and not yet prevalent in industrial applications
generated data set is presented and published online for as they continue to have restrictions. (Bellini, Filippetti,
further research. Tassoni, & Capolino, 2008; Herold, Piantsop Mbo’o, &
Hameyer, 2013; Paschke et al., 2013; Picot et al., 2014;
Christian Lessmeier et al. This is an open-access article distributed under
the terms of the Creative Commons Attribution 3.0 United States License,
Stack, Habetler, & Harley, 2003)
which permits unrestricted use, distribution, and reproduction in any
medium, provided the original author and source are credited.
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

For MCS-based methods, it has to be distinguished between the development of reliable CM systems based on MCS.
bearing damage in the motor itself (internal bearings) and (Bellini et al., 2008; Stack et al., 2003).
bearings in the remaining drive train (external bearings).
Thus, it is no surprise that some authors report a high
Damage of the internal motor bearings or other faults in the discrepancy between the proposed methods in science and
motor itself, as for example broken rotor bars or rotor their application in industry. Stack assumes that one reason
eccentricity, directly influence the airgap of the motor and for this discrepancy is the use of the simple artificial defects
induce vibrations at characteristic frequencies in the motor in the form of single point damages in research. For a more
current. Detection of damage in external bearings is more detailed consideration, bearing damage is nowadays
complex as the damage signature has to be transmitted separated into single point and distributed damage. (Nandi,
indirectly through torque variations along the drive train. Toliyat, & Li, 2005; Stack et al., 2003; Tandon &
Therefore, it is damped and superimposed with disturbances Choudhury, 1999)
from the powered process, leading to noisy and hard
Single point damages are characterized by their small
detectable signals. (Blödt, Granjon, Raison, & Rostaing,
extend at a localized position, for example a crack or a small
2008; Herold et al., 2013; Schoen, Habetler, Kamran, &
pit. When the rolling elements run across these defects,
Bartfield, 1995).
shock pulses stimulate vibrations at the characteristic
Schön et al. (1995) discuss both situations, damage in bearing frequencies (Randall, 2011). Distributed damages,
internal and external bearings, and derive formulas for the also called generalized roughness (Stack et al., 2003) or
theoretically expected frequencies and sidebands in each extended faults (Randall, 2011), induce broadband
situation. However, these frequencies can only be detected vibrations which are often not easy to separate from the
reliably in special cases because they are influenced by noise of the signal. In general, these damages are caused by
operating conditions, machine design, external noise etc. wear, corrosion or plastic deformation, but also by extensive
and have usually only been investigated for large scale pitting (fatigue) damages (Nandi et al., 2005). The authors
damages. (Bellini et al., 2008; Mbo'o, Herold, & Hameyer, of the present paper also observed the combination of
2004). Hence, additional investigations are required. different damage types in one bearing, which leads to
superimposed signals for CM.
It is assumed that good results for diagnostics can be
achieved with data-driven classification methods using While the bearing damage signature of vibrations is well
machine learning (ML) (Bellini et al., 2008; Kankar, known, there is little experience with the MCS of external
Sharma, & Harsha, 2011; Paschke et al., 2013). The low bearing damage and the correlation of the artificial damages
availability of training and testing data with the MCS of used in research with real damages in the industry (Bellini et
external bearing for ML procedures limits further research, al., 2008). Especially for MCS-based condition monitoring,
so that there is insufficient knowledge about the capability the authors of this paper think a more detailed description of
of industrial usage. To overcome this limitation, research damage is required as the separation into single point and
has to address the systematic generation of training data distributed damage is not sufficient. Moreover, a precise
with mechanical bearing damage. (Nectoux et al., 2012; description of specific test damages and test bearing
Stack et al., 2003; Zarei & Poshtan, 2009) specifications for experimental examinations are needed.
Therefore, a systematic methodology is necessary to
This results in two subjects of activity, which should be
characterize the appearance of damage for CM, as existing
considered in more detail for the generation of training data:
methodologies are not very specific. Given the need for
a systematic generation of the bearing damages themselves
further research the present paper focuses on the systematic
and their systematic specification.
generation of training data for ML classification algorithms
In the industry, a wide variety of bearing damages occur. based on MCS.
Especially the development of fatigue damages or damages
To create a systematic data set, first a method for the
caused by solid particles is randomly influenced; in
systematic description of the rolling bearing damage is
addition, the appearance of damages can change over time.
described (section 2). Afterwards, the generation of artificial
In the literature, six main damage modes and more than 20
and real bearing damages to be used for training data
different damage symptoms in bearings are treated. (ISO
generation in the experiments (section 3) and the
15243, 2010; Bartz, 1985; Schaeffler Technologies AG &
experimental set-up itself (section 4) are explained. Section
Co. KG, 2015) Recent research studies on CM address only
5 gives details on the database and their availability for
some of these damages. Furthermore, many of the available
further research.
research papers focus only on artificial bearing damages
because these are easy to generate. (Pacas, Villwock, & To validate the data set and as a proof of concept for the
Dietrich, 2009) Publications show that in most cases only general method, the paper points out some damage detection
single point damages are used for research, which inhibits results as well. In section 6 the established method of
envelope analysis is used to prove the proper data

2
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

acquisition and to allow an initial evaluation of the bearing important when model based approaches are used and, for
damages. Section 7 shows that machine learning example, the geometry is needed to calculate the
classification algorithms are able to identify the damaged characteristic frequencies.
bearings in industrial drive trains using MSC, but require
Criteria in the third category “application specific
sophisticated training data, as classification accuracies differ
information” list the specific information of the individual
depending on the training data sets. Therefore, the
bearing in the form of distinct identification codes and
dependency of the classification results on different damage
information concerning the place of operation and the
types and manifestations of damage is examined.
corresponding operating conditions. This information is
especially valuable to help figure out the damage causes in
2. CATEGORIZATION OF BEARING DAMAGE
industry applications. Further on in this paper the four-digit
For systematic approaches of the examination of bearing code for each specific bearing is used for identification. It
damages, a comprehensive methodology for categorizing consists of two letters and two numbers, e.g. KA01, KB23.
bearings and their specific damages is needed. The proper
The damage itself in the fourth group is described by its
description and categorization of bearing damages and their
type and subtype according to ISO 15243 (2010), its
cause is not an easy task as bearing damages occur as an
interaction of different causes and conditions. Moreover, the Table 1. Categorization of bearing damage.
patterns of damages may vary widely and often occur in
combinations. Bearing damages can have different Category and sub-category Criterion Example
Bearing Type ball bearing
development states and not all of them lead to a sudden Bearing designation
failure. General info
(dimension series, bore 6203
(bearing)
code)
ISO 15243 gives a methodology for the classification of Suffix n/a
bearing damage and failures. The damages are categorized Diameter of inner raceway 24.0 mm
Diameter of outer raceway 33.1 mm
into six main damage modes and their sub-modes. The six Pitch circle diameter 28.55 mm
main damage modes are: fatigue, wear, corrosion, electrical Geometry Number of rolling elements 8 pc.
Manufacturer
erosion, plastic deformation, and fracture and cracking. Rolling element diameter 6.75 mm
specific
Length of rolling element 6.75 mm
information
This existing methodology of categorization is helpful to Nominal pressure angle 0° N
about bearing
Static load rating 4750 N
describe damages and to figure out their causes. Parameters Dynamic load rating 9500 N
Nevertheless, it only considers the damage modes in general Speed limit 12000 rpm
Manufacturer FAG
and does not describe the detailed physiognomy of the Bearing code (used for
damage. Identification datasets)
KB24
Sample number 12-01
The authors do not know of any extensive and established Installation site 01
method to describe the physiognomy of bearing damages in Place of Installation type (system KAt - lifetime
operation type) test rig
detail. Therefore, a method for general categorization and Application Operator KAt
detailed specification was developed. It focuses on the specific Number of load cycles 2769500
information Lifetime 15:01 h:min
detailed description of the damage as well as on the
Load 3800 N
corresponding bearing and its application. Description Operation Dynamic equivalent load 3800 N
criteria were gathered using brainstorming methods and conditions Rotational speed 2900 rpm
sorting them into a hierarchic structure of categories. The Load direction 0°
categories and criteria are shown in Table 1. Comment n/a
Mode fatigue
Type of
The criteria were grouped into four main categories, the first Damage
Sub-mode n/a
Symptom pitting
three giving information about the bearing and the fourth Component outer ring
providing detailed information about the damage. According Position of damage raceway
to these criteria, a detailed profile (or fact sheet) can be Damage Damage combination multiple
location Arrangement of the
created for any damaged bearing. repetitive and multiple no repetition
damages
The criteria of the first group “general info” name the Damage Length 9.4 mm
bearing type and the standardized code according to the Extent of damage 3
Geometry Width total raceway
standards of each bearing series. Depth n/a
Characteristic of damage distributed
The criteria in the second group “manufacturer specific acc. lifetime
information” give information about the internal geometry Damage method
test
Damage
and parameters of the bearing as these are not standardized. occurrence Cause of damage (category)
operating
conditions
Consequently, these parameters differ from manufacturer to Cause of damage (detailed) load, lubricant
manufacturer. Especially for CM, these parameters may be

3
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

location, geometry, and occurrence. If there is more than a Extent of damage: The extent of the damage describes the
single damage to one bearing, the criteria of the last group size of the damage in normalized levels, which are
are repeated to get a detailed and full description of all independent of the bearing size. The levels are based on the
occurring damages. length of the damage, as this is a determining factor for the
signal output in CM and the intensity of the damage from
As this paper focuses on CM of bearing damages, some of
the point of view of a machine operator. For this the
the most important criteria concerning the description of
percentage of length relative to pitch circumference is
bearing damages for CM are explained in detail. For a better
calculated and then assigned to five levels according to
understanding, an example is given in the last column of
Table 2.
Table 1.
Table 2. Damage levels to determine the extent of damage.
Damage combination: This criterion characterizes the
occurrence of combined or repetitive damage based on the Assigned
Damage Limits for bearing
damage symptoms. This is important for real damages percentage
level 6203
which are not plain single point damages. It is described by values
the following three options: 1 0-2 % ≤ 2 mm
2 2-5 % >2 mm
 Single damage: One single component of the rolling
bearing is affected by a single damage, for example a 3 5-15 % >4.5 mm
single pitting on the inner ring. 4 15-35 % >13.5 mm
5 >35% >31.5 mm
 Repetitive damage: Identical damage symptoms are
repeated at several places on the same bearing
component, for example several, non-continuous Characteristic of damage: The characteristic of damage
pittings on the inner ring raceway. assigns the damages into the already mentioned groups of
 Multiple damage: Different damage symptoms occur in single point and distributed damages. For a clear and easy
the bearing or identical damage symptoms occur on division between these groups the dividing value is set to
different bearing components. This option can include the diameter of the rolling elements, to which the damage
repetitive damages. length is compared. This guarantees that for the single point
defects only one rolling element has contact to the damage
Arrangement of the repetitive and multiple damages: This in all cases and the extent of the damage is small in
criterion characterizes the arrangement of the damage comparison to the bearing size.
symptoms on each component (e.g., the inner ring) for the
repetitive and multiple damages (see above). This criterion It should be noted that the appearance of the damages can
is described by the following options: have a simultaneous influence on several of the critera.
Therefore the criteria are not independent. The dependency
 Regular: The damage symptoms recur in a regular is tolerated in order to gain a detailed description of
pattern on the component. damages, due to the facts, that the damage diversity is high
 Random: Random distribution of the local damage and the real phenomena intermix.
symptoms.
Altogether, this method of categorizing gives a very detailed
 No repetition: The damage occurs only once and this description of the regarded bearings and their damages. The
criterion does not apply. difficulty is to collect all the information and parameters, as
Geometrical size: The geometrical size of the damage is some of them are not obvious and hard to obtain. The
described by the length, width, and depth of the damage, method was developed using a wide approach and can thus
according to the directions quoted in VDI 3832 (2013) and be used to investigate the damage causes or other bearing
depicted in Figure 1. damage related issues as well.
position of rolling element rolling element
3. GENERATION OF BEARING DAMAGE
length
A main focus of this research paper is to generate systematic
rolling direction data of measurement signals for condition monitoring with a
width

broad variety of bearing damage. As an intermediate step,

the bearing damages had to be generated, requiring a special
damage preparation for the bearings.
damage depth
To keep the amount of experiments manageable, research
Figure 1. Parameters for describing the was limited to ball bearings of type 6203. The types of
geometry of bearing damages. damage were processed and selected respecting the
technical possibilities of their manufacturing and their

4
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

representation of bearing damage in industrial applications. between the damage and the undamaged raceway areas is
Two groups of damages in ball bearings were used in the apparent. To examine this correlation, the third type of
experiments: artificial and real damages. Both types of artificial damage caused by a manual electric engraver is
damage exist at the inner and outer ring of the ball bearing prepared, which has an irregular surface structure and lower
6203. For this paper bearings of the manufacturers FAG, depth, thus resembling real pitting damage.
MTK and IBU/IBB are used. All these bearings have eight
Table 3. Artificial bearing damage.
rolling elements and the geometrical sizes are nearly
identical, so that the characteristic kinematic frequencies do
not vary more than 1-2%.
The artificial damages were introduced manually using
machining tools (see 3.1). For the generation of the real
bearing damages an apparatus for accelerated lifetime tests
Sharp trench by EDM Artificial pitting by
was used (see 3.2). - KA01
Drilling - KA09
electric engraver - KA03

3.1. Artificial Damage

Details about the size and a categorization of the available
The use of artificial damages to develop CM methods is test bearings according to the developed criteria (see
very common, as reported in several research papers. Often section 2) are listed in Table 4.
rather extensive bearing damages such as holes in the outer
race of the bearing are used. (Blödt et al., 2008; Nandi et al., Table 4. Test bearings with artificial damage.
2005; Zarei & Poshtan, 2009)

Component

Extent of
Damage

Damage
Bearing

Method
(level)
Code
Some researchers use only outer ring damages, even though
inner ring damages are more likely to occur because of the
higher Hertzian stress. Furthermore, inner ring damage is
not as easy to detect as outer ring damage because in most KA01 OR 1 EDM
situations, the signal is disguised by a modulation because KA03 OR 2 electric engraver
of the rotation of the inner ring. (Obaid, Habetler, & Stack, KA05 OR 1 electric engraver
2003; Pacas et al., 2009) KA06 OR 2 electric engraver
Most commonly used methods for artificial damage KA07 OR 1 drilling
preparation are: KA08 OR 2 drilling
KA09 OR 2 drilling
 Trenches generated by electrical discharge machining KI01 IR 1 EDM
(EDM). (Niknam, Thomas, Hines, & Sawhney, 2013), KI03 IR 1 electric engraver
(Villwock, 2007; Yang, Mathew, & Ma, 2005; Zoubek,
KI05 IR 1 electric engraver
Villwock, & Pacas, 2008), (Niknam et al., 2013; Patil,
KI07 IR 2 electric engraver
Mathew, Rajendrakumar, & Desai, 2010)
KI08 IR 2 electric engraver
 Drilling holes into the rings: The bore diameter, the OR: outer ring; IR: inner ring;
orientation, and the position depend on the bearing type
and size. (Amirat, Choqueuse, & Benbouzid, 2013; It has to be mentioned that all the artificial damages are
Blödt et al., 2008; Djeddi et al., 2007; Silva & Cardoso, single point damages without a repetition or combination
2005; Zarei & Poshtan, 2009) with other damages (compare section 2 - Categorization of
The artificial damages used in this paper were caused by Bearing Damage)
three different methods:
1. electric discharge machining (trench of 0.25 mm length 3.2. Generating Real Bearing Damage Samples by
in rolling direction and depth of 1-2 mm), Accelerated Lifetime Tests
2. drilling (diameter: 0.9 mm, 2 mm, 3 mm), and There are two options to test and develop CM methods with
3. manual electric engraving (damage length from measurement data of real bearing damages: to measure
1-4 mm). bearing damages in real or in scientific test rigs.
Table 3 shows three examples of damages caused by each of The successful use of data from industry applications is
the three methods. The first two methods are very precise quite difficult, as it is complicated to receive systematical
and easy to reproduce. Therefore, these artificial damages and comparable training data for different damages. This is
are appropriate for comparing research results with other because of the long lifetime of most bearings and, if damage
studies. A lack of correlation to real bearing damage is is recognized, bearings are replaced before failure, so that
assumed though, because a very abrupt and sharp transition

5
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

defect states are seldom. Therefore, often only a small characterized according to the damage with the highest
number of damage states are available. Moreover, there are extent.
many different bearing types, sizes and machine types, and
Table 5. Test bearings with real damages caused by
the operating conditions may change irregularly as they
accelerated lifetime test.
depend on the application. Therefore, measurements are
expected to be influenced by various factors. This is helpful

and symptom)

Characteristic
Arrangement
Combination
Bearing code

(main mode

of damage
at later development stages to increase the robustness, but

Extent of
Damage

Bearing
element

damage
distracting for the development of the basic methods. To
reduce the external influences, scientific test rigs are used,
which enable the generation of realistic bearing damages by
accelerated lifetime tests. Furthermore, damages for the no single
KA04 fatigue: pitting OR S 1
needed bearing types and geometries can be systematically repetition point
generated at reproducible conditions. The main Plastic deform.: no single
KA15 OR S 1
Indentations repetition point
disadvantage is that this consumes a lot of time and single
resources. (Nectoux et al., 2012; Qiu, Lee, Lin, & Yu, 2006) KA16 fatigue: pitting OR R random 2
point
no single
For the present research paper, ball bearings with real KA22 fatigue: pitting OR S
repetition
1
point
damages were obtained from an accelerated life time test. Plastic deform.:
KA30 OR R random 1 distributed
The accelerated life time test rig consists of a bearing Indentations
housing and an electric motor, which powers a shaft with IR single
KB23 fatigue: pitting M random 2
(+OR) point
four test bearings of type 6203 in the housing (Figure 2).
IR no
The test bearings rotate under a radial load which is applied KB24 fatigue: pitting
(+OR)
M
repetition
3 distributed
by a spring-screw mechanism. The applied radial force is Plastic deform.: OR +
KB27 M random 1 distributed
higher than in usual bearing applications to accelerate the Indentations IR
appearance of fatigue damages, but still low enough to not no single
KI04 fatigue: pitting IR M 1
repetition point
exceed the static load capacity of the bearing. Moreover, no single
low viscosity oil was used, which leads to improper KI14 fatigue: pitting IR M 1
repetition point
lubrication conditions and favors the appearance of no single
KI16 fatigue: pitting IR S 3
damages. repetition point
single
KI17 fatigue: pitting IR R random 1
point
no single
KI18 fatigue: pitting IR S 2
repetition point
no single
KI21 fatigue: pitting IR S 1
repetition point
OR: outer ring; IR: inner ring;
S: single damage; R: repetitive damage; M: multiple damage

Figure 3 shows two examples of bearing damages generated

in the accelerated lifetime test.
Figure 2. Apparatus for accelerated life time test.
Several damaged bearings were obtained by the lifetime test outer ring
and categorized according to the developed criteria (see inner ring
section 2). 33 damages were recognized in 18 bearings out
of a test amount of 108 bearings used in the lifetime test.
Around 70 % of the occurred damages were fatigue
damages, which arise in the form of pittings. The rest of the
bearings, except for one fracture, were damaged by plastic
deformation, i.e. in the form of indentations caused by inner ring outer ring
debris. Pitting damages occurred both on the inner and outer
ring of the bearings. Indentations were found at the outer Figure 3. Indentation at the raceway of the outer ring (left);
ring only. Damage at the rolling elements was not observed. small pitting at the raceway of the inner ring (right).
The extent of the damage was categorized by the length of
Bearing KB27 (left) shows indentations from debris
the damaged surface in rolling direction into the levels 1 to
particles at the outer ring. As the damage is repeated along
3 (compare Table 2 and Table 5). Multiple damages were
the raceway only some of them are shown. Bearing KI14
(right) has a very small pitting at an early stage (level 1) at

6
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

the inner ring. For the experiments, representative samples modulation of the inverter. (Lessmeier, Piantsop Mbo'o,
with different properties were selected (see Table 5). Coenen, Zimmer, & Hameyer, 2012)
Figure 5 shows the schema of the measurement procedure
4. EXPERIMENTAL SET-UP
and the recorded measurands. The motor phase currents are
To generate the experimental data for the development of measured by a current transducer of the type LEM CKSR
CM methods of damaged bearings by using motor current 15-NP with an accuracy of 0.8 % of IPN = 15 A. The MCS
signals, a specific test rig was designed and operated at the are then filtered by a 25 kHz low-pass filter and converted
Chair of Design and Drive Technology, Paderborn from an analogue to a digital signal with a sampling rate of
University. The test rig is a modular system to ensure 64 kHz. The current transducers are used instead of the
flexible use of different defects in an electrical driven internal ammeters of the inverter because of their easy
mechanical drive train. Defects in mechanical components, signal access as the currents can be measured externally
as they occur in gearboxes or electrical machines, are between motor and inverter.
experimentally reproduced to generate failure data using the
test rig. For the generation of the measurement data, the measuring bearing
current signals of the electric motor are recorded. motor
shaft module
load motor

Additionally, the vibration signal of the housing of the test testbed

bearings are measured as reference. speed - n
surface acceleration - a
radial force- F
torque- M
current temperature - T
4.1. Test Rig transducer
Current
u, v
The test rig consists of several modules: an electric motor A/D
(1), a torque-measurement shaft (2), a rolling bearing test frequency
converter
data
module (3), a flywheel (4) and a load motor (5), see inverter storage
Figure 4. The ball bearings with different types of damage
are mounted in the bearing test module to generate the
Figure 5. Schema of measurement procedure.
experimental data.
At this scientific level of development, a high sampling rate
The rolling bearing module provides the possibility of using
and accuracy are additional advantages of this setup.
a test bearing under a constant radial load, which can be
Nevertheless, the used transducers are similar to the ones
continuously adjusted up to 10 kN before each experiment. commonly used in industry applications, so that few
An adapter gives the possibility to measure the vibration of
difficulties are expected transferring the research outcomes
the inner housing, which holds the test bearing in the main
to industrial CM systems.
direction of the load. The precise design of the bearing
module and additional features, such as the possibility to The acceleration of the bearing housing is measured at the
simulate tilting faults or the use of roller bearings, are adapter at the top end of the rolling bearing module using a
described by Lessmeier, Enge-Rosenblatt, Bayer, & piezoelectric accelerometer (Model No. 336C04, PCB
Zimmer, 2014. Piezotronics, Inc.) and a charge amplifier (Type 5015A,
Kistler Group) with a low-pass filter at 30 kHz. The signal
is digitalized and saved synchronously to the MCS with a
sampling rate of 64 kHz.
The flywheel and the load machine simulate inertia and load
of the driven equipment, respectively. The load motor is a
PMSM with a nominal torque of 6 Nm (power of 1.7 kW).
Figure 4. Modular test rig. To record the operating conditions the following additional
The motor (1) is a 425 W Permanent Magnet Synchronous parameters are measured synchronously to the motor
Motor (PMSM) with a nominal torque of T = 1.35 Nm, a currents and vibration signal but with lower sampling rates:
nominal speed of n = 3,000 rpm, a nominal current of the radial force on the bearings (Compression and Tension
I = 2.3 A and a pole pair number p = 4 (Type SD4CDu8S- Force Sensor Type K11, Lorenz, 10 kN), the load torque at
009, Hanning Elektro-Werke GmbH & Co. KG). It is the torque-measuring shaft, the rotational speed (Torque
operated by a frequency inverter (KEB Combivert 07F5E Transducer Model 305, Magtrol, 2 Nm) and the oil
1D-2B0A) with a switching frequency of 16 kHz. This temperature in the bearing module.
standard industrial inverter is used to provide conditions
similar to motors used in the industry because the current 4.2. Experiments
signals show significant noise due to the pulse-width The test rig was operated under different operating
conditions to analyze the influence of operation parameters

7
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

and to ensure the robustness of the CM methods at different bearing diagnostics employing ML-algorithms.
operating conditions.
Some diagnostic data sets for bearing damages are publicly
The rotational speed of the drive system, the radial force available; the most popular and comprehensive ones are
onto the test bearing and the load torque in the drive train listed below:
are the main operation parameters. To ensure comparability 1
of the experiments, fixed levels were defined for each  CWRU: Bearing Data Center/ Seeded Fault Test Data
2
parameter (Table 6). All three parameters were kept  FEMTO Bearing Data Set
constant for the time of each measurement. At the basic  MFPT Fault Data Sets
3

setup (Set no. 0) of the operation parameters, the test rig 4

 Bearing Data Set IMS
runs at n = 1,500 rpm with a load torque of M = 0.7 Nm and
a radial force on the bearing of F = 1,000 N. Three These data sets focus on different aspects. Some use
additional settings are used by reducing the parameters one artificial damages (CWRU, MFPT), others use real damages
by one to n = 900 rpm, M = 0.1 Nm and F = 400 N (set No. (FEMTO, IMS). CWRU and FEMTO use different
1-3), respectively. For each of the settings, 20 measurements operating conditions by varying load and speed; others use
of 4 seconds each were recorded. Another parameter is the only different load situations (MFPT) or just one condition
temperature, which was kept roughly at 45-50 °C during all (IMS). The FEMTO data set provides run-to-failure data
experiments. with measurements over a long period, but does not give
any information about the properties of the damages.
In total, experiments with 32 different bearings were
performed: 12 bearings with artificial damages and 14 Altogether, the data sets mentioned use the classical
bearings with damages from accelerated lifetime tests (see vibration signals for bearing diagnostics. To the best
Table 4 and Table 5). Moreover, experiments with 6 healthy knowledge of the authors of the present paper, no publicly
bearings and a different time of operation were performed as available data set providing data for diagnostics of external
reference states as shown in Table 7. bearing damages based on MCS is known.
Therefore, the focus was to systematically create a high
5. DATABASE quality data set which takes into account several operating
Nowadays, a huge amount of data is collected in industry conditions and a wide distribution of artificial as well as
and science for different purposes; some of it is made public realistic bearing damages.
in repositories or on websites. But obtaining the appropriate Smith and Randall demand some properties for benchmark
data in the needed quality and quantity for specialized data sets which are based on their experience with different
research often is still challenging, especially, if a wide range diagnostic methods used on a vibration-based CM data set
of different types of damages or the yet rarely used MCS are for bearings. They list three requirements (Smith & Randall,
the target of interest. This also applies to training data for 2015), which can be applied accordingly to MCS data sets:
Table 6. Operating parameters.  Systematic and comprehensive documentation.
Rotational Load Radial Name of  High sampling rates (>40 kHz).
No. speed Torque force Setting  Data verification with established methods before
[rpm] [Nm] [N] publication of the data.
0 1500 0.7 1000 N15_M07_F10 The first two issues are already addressed in this paper by
1 900 0.7 1000 N09_M07_F10 the characterization of damage and the extensive
2 1500 0.1 1000 N15_M01_F10 documentation of the experiments, including a high
sampling rate for the two main signals (64 kHz).
3 1500 0.7 400 N15_M07_F04

Table 7. Operating parameter of healthy (undamaged)

bearings during run-in period. 1
Case Western Reserve University (CWRU), Cleveland, Ohio, USA
http://csegroups.case.edu/bearingdatacenter/home
Bearing Run-in Period Radial Load Speed 2
FEMTO-ST Institute, Besançon, France; http://www.femto-
Code [h] [N] [min-1] st.fr/en/Research-departments/AS2M/Research-groups/PHM/IEEE-PHM-
K001 >50 1000-3000 1500-2000 2012-Data-challenge.php
3
K002 19 3000 2900 Mechanical Failures Prevention Group (MFPT) Society (a Division of the
K003 1 3000 3000 Vibration Institute), Oak Brook, IL, USA,
http://www.mfpt.org/FaultData/FaultData.htm
K004 5 3000 3000 4
J. Lee, H. Qiu, G. Yu, J. Lin, and Rexnord Technical Services, NSF
K005 10 3000 3000 I/UCR Center for Intelligent Maintenance Systems, Milwaukee, WI, USA:
K006 16 3000 2900 http://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-data-repository/

8
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

The examination of the data with established methods will 6. ENVELOPE ANALYSIS FOR VIBRATION SIGNALS
be addressed in the next chapter, so that the provided data
First of all, the bearing damage diagnostic is performed with
set and the corresponding documentation fulfill the
the envelope analysis based on the vibration, as this is a
requirements. Therefore, it is assumed that the data is
well-known and established procedure. The objective of this
beneficial to further research.
examination is to verify the data and check the execution of
To enable and encourage collaboration in the field of experiments for mistakes and errors. Moreover, the
bearing condition monitoring and to allow researchers to use detection of the damages provides a first estimate of the
the generated data as benchmark data sets for further attitude of signals and corresponding damages to the
research, the data is published online. The data is licensed authors.
under the Creative Commons Attribution-NonCommercial
It can be demonstrated that the results of the envelope
4.0 International License.5 Noncommercial academic use of
analysis correspond to the description in literature. The
the data is explicitly allowed, but a citation of the origin is
bearing damages cause typical characteristic kinematic
required and expected.6 For commercial use, please contact
frequencies that can be observed in the signals. These
the author. The download page is available at the KAt-
frequencies can be calculated for localized damages when
DataCenter website of the Chair of Design and Drive
the position of the damage (e.g. outer or inner ring) and the
Technology, Paderborn University, Germany:
geometrical parameters of the bearing are known. (Randall,
http://mb.uni-paderborn.de/kat/datacenter 2011)
The data consists of measurements from 32 different Two single point damages, one at the inner ring (KA04) and
bearing experiments. The bearings belong to three main one at the outer ring (KI18) are considered exemplarily. The
groups: envelope spectra in Figure 6 clearly shows the ballpass
frequency of the outer raceway (fo) and their harmonics. The
 Undamaged (healthy) bearings (6x), see Table 6. envelope spectra of the damage at the inner raceway in
 Artificially damaged bearings (12x), see Table 4. Figure 7 shows the fundamental rotation frequency of the
 Bearings with real damages caused by accelerated shaft and its harmonics (fn), the ballpass frequency of the
lifetime tests, (14x) see Table 5. inner race (fi), and its sidebands as well as its corresponding
harmonics.
The specifications of the bearings are listed in the tables
above. 0
fn=25Hz;

In summary, the main characteristic of the data set are: -10 fo

2fo
fo=76,25Hz
Amplitude [dB]

3fo
-20

 Synchronously measured motor currents and vibration -30

signals with high resolution and sampling rate of 26 -40

damaged bearing states and 6 undamaged (healthy) -50
states for reference.
-60
0 50 100 150 200 250 300
 Supportive measurement of speed, torque, radial load, data set: N15_M07_F10_KA04_1 Frequency[Hz]

and temperature.
Figure 6. Envelope spectra of vibration signal
 Four different operating conditions (Table 6). for bearing damage at the outer ring.
 20 measurements of 4 seconds each for each setting,
saved as a matlab file with a name consisting of the 0
fn=25Hz;
fi=123,3Hz
code of the operating condition and the four-digit -10
fn fi-fn fi
Amplitude [dB]

2fn fi+fn
fi-2fn
bearing code (e.g. N15_M07_F10_KA01_1.mat). -20
fi-3fn 3fn
4fn 2fi-3fn
fi+2fn 2fi-2fn 2fi-fn
fi+3fn fi+4fn
fi 2fi+fn

-30 fi-4fn

5fn
Systematic description of the bearing damage by -40
uniform fact sheets (according to the categorization in -50
section 2 - Categorization of Bearing Damage). -60
0 50 100 150 200 250 300
data set: N15_M07_F10_KI18_1 Frequency [Hz]

Figure 7. Envelope spectra of vibration signal

for bearing damage at the inner ring.

5 7. CLASSIFICATION ALGORITHM AND RESULTS

To view a copy of this license, visit
http://creativecommons.org/licenses/by-nc/4.0/ The most common analysis for bearing damages involving
6
Please cite this paper and give name of the author, institute and link to the
Kat-DataCenter: Christian Lessmeier et al., KAt-DataCenter: http://mb.uni-
MCS is to convert the time domain signals into the
paderborn.de/kat/datacenter, Chair of Design and Drive Technology, frequency domain and check for the bearing characteristic
University Paderborn.

9
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

Figure 8. Frequency spectrum from MCS for (a) healthy bearing, (b) outer ring damage and (c) inner ring damage.

frequencies (Blödt et al., 2008). However, in most cases the (CART), random forests (RF), Boosted Trees (BT), neural
raw signals are masked by external noise, what makes this networks (NN), support vector machines with parameters
method very difficult to apply. Figure 8 shows exemplary optimally tuned using particle swarm optimization (SVM-
plots of the power spectral density of a healthy bearing, a PSO), extreme learning machine (ELM), and k-nearest
bearing with outer ring damage, and a bearing with inner neighbors (kNN).
ring damage. From Figure 8, only the electrical supply
frequency (𝑓𝑒 ) and its harmonics are easily observable. This
is mainly due to the masking of bearing characteristic
frequencies by external noise and also the presence of
distributed damages which are difficult to detect using
characteristic frequency approaches (Yang, Merrild, Runge,
Pedersen, & Hakon Børsting, 2009).
However, by extracting features from the raw MCS, it is
possible to observe the clustering of various damage classes
of the bearing through a feature plot as seen in Figure 9.
This shows that machine learning (ML) algorithms can be
trained to identify various damages on the bearings from
extracted features.
Figure 10 shows the workflow of the application of machine
learning for the classification of various bearing damages.
The machine learning algorithm learns to map the input
features to the corresponding target which consists of a class
label representing the type of damage. A classification
model is obtained which can be used to predict the type of
damage for a given set of input features. In this work, the
bearings were categorized in three classes as either healthy,
having inner ring damage, or having outer ring damage.
Figure 9. Clustering of bearing health states for (a) features
Seven state of the art algorithms and an ensemble of the
from MCS and (b) vibration signals.
algorithms using majority voting were implemented. These
algorithms include: classification and regression trees

Target TRAINING

Feature ML Classification
Data Extraction and
Acquisition Algorithm Model
Selection

Data Extraction of Fault Predicted

Acquisition Selected Features Classification Class
Model

TESTING/ONLINE DIAGNOSIS

Figure 10. Application of machine learning (ML) for fault diagnosis.

10
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

7.1 Feature Extraction and Selection In the current study, a selection criterion of 𝑑𝑗̅ ≥ 0.7 was
Feature extraction involves deriving time-, frequency,- and found to yield the best results. In this case, 9 out of 23
time-frequency-domain features from the raw signals. features are selected from each of the two MCS per sample
Signals acquired from machinery components such as faulty and 15 features from the vibration signals.
bearings are normally considered non-stationary, meaning
that frequency components of the signal vary over time. 7.2 Training with Data from Bearings with
Therefore, the extraction of time-frequency features is Artificially Induced Damages
necessary. In this work, wavelet packet decomposition At first, the algorithms were trained with features extracted
(WPD) is employed for the extraction of the time-frequency from measurements with artificially induced bearing
features. The raw signal is decomposed up to 3 levels. The damages (see Table 8). The objective was to use a
detailed coefficients and approximate coefficients of level 1 combination of healthy bearings and artificial damaged
to 3 are obtained and from which the wavelet energy is bearings to identify bearings with real damages and classify
computed. Fast Fourier Transform (FFT) and power spectral them as either healthy, inner ring damage or outer ring
density (PSD) are used to extract the frequency domain damage (3-class approach). In the testing part, features
features. A total of 23 features are extracted from each extracted from data with real damages (Table 8) were used
signal. A list of the extracted features can be found in as the input to the classification model. The output was the
(Kimotho & Sextro, 2014). Since not all features are class label.
suitable for fault classification, it is necessary to select
features that contain most information on the health status of Table 9 shows the classification accuracy of the machine
a component to avoid over-fitting and to improve accuracy. learning algorithms used with features extracted from MCS
Suitable features for fault or health state classification and vibration signals under operating conditions in setting 0
should provide a good separation between different classes. (see Table 6), which was found to produce better
In this study, a feature selection method based on maximum performance. Features from vibration signals result in a
separation distance between different health states was better classification accuracy than those from MCS.
employed. Table 8. List of data sets used for training and
Given a feature set of 𝑗 = 1, 2, . . . 𝑄 features in 𝑐 = testing the machine learning algorithms.
1, 2, . . . 𝑁𝑐 classes or health states, the feature selection is Class Training Testing
performed as follows: 1 Healthy K002 K001
1. Normalize the features between 0 and 1. KA22
KA01 KA04
2. Compute the mean (𝑚𝑗𝑐 ) of each feature 𝑗 within 2 OR Damage KA05 KA15
class 𝑐 as follows KA07 KA30
1
𝑚𝑗𝑐 = ∑𝑛𝑖=1 𝑥𝑖𝑗𝑐 , KA16
𝑛 KI14
where 𝑥 is the feature and 𝑛 is the number of KI01 KI21
samples. 3 IR Damage KI05 KI17
KI07 KI18
3. Compute the mean of the squared Euclidean
KI16
distance (𝑑𝑗 ) between each feature data point 𝑖 and
the mean of the same feature in each class
1 2 Table 9. Performance of various algorithms trained with
𝑑𝑗 = ∑𝑁 𝑁𝑐 𝑛
𝑘=1 ∑𝑐=1 ∑𝑖=1(𝑥𝑖𝑗𝑘 − 𝑚𝑗𝑐 ) .
𝑐
𝑛𝑁𝑐2 features of bearings with artificially induced damages and
4. Normalize the separation distance with the tested with features of bearings with real damages.
maximum feature separation distance to produce a Algorithm Classification Accuracy [%]
performance evaluation criteria MCS Vibration Signals
𝑑𝑗 CART 26.8 65.9
𝑑𝑗̅ = .
max(𝑑) RF 45.0 64.1
5. Select the distance with a performance greater than BT 38.6 62.3
a predetermined threshold. The threshold can be NN 45.5 65.5
defined by evaluating the classification accuracy SVM-PSO 60.9 65.5
for each combined set of features. ELM 45.5 65.9
kNN 45.5 63.2
Ensemble 45.9 75.0

11
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

Using features extracted from MCS, the support vector Vibration Signals
machine approach with parameters optimally tuned by using 1
20 0 0 100%
particle swarm optimization (SVM-PSO) shows the best 9.1% 0.0% 0.0% 0.0%
performance of all algorithms with 60.9% classification

Predicted Class
2 0 45 0 100%
accuracy. Figure 11 depicts the confusion matrix for SVM- 0.0% 20.5% 0.0% 0.0%
PSO. On the confusion matrix, the rows correspond to the
0 55 100 64.5%
predicted class and the columns to the true class. The 3 0.0% 25.0% 45.5% 35.5%
diagonal cells represent the correctly predicted class, while
100% 45.0% 100% 75.0%
the off-diagonal cells show misclassifications. The left 0.0% 55.0% 0.0% 25.0%
column displays the false positives (healthy samples
1 2 3
classified as damaged) and the upper row the false negatives
Target Class
(samples with damages classified as healthy). The bottom
row indicate the accuracy and misclassification rate for each Figure 12. Confusion matrix for ensemble of all
true class while the right column shows the accuracy and algorithms trained with features extracted from vibration
misclassification rate of each predicted class. The cell in the signals of bearings with artificially induced damages and
bottom right gives the overall accuracy. Bearings with inner tested with features of bearings with real damages.
ring damages (class 3) record the highest misclassification
rate, with the majority of the cases being classified as 7.3 Training with Data of Bearings with Real
having outer ring damages (class 2). Damages
MCS
The data sets in Table 10 of healthy bearings and those with
1
20 0 4 83.3% real damages were used for training and testing in a 5-fold
9.1% 0.0% 1.8% 16.7% cross-validation manner. For each combination, three data
sets from each class were used for training and the other two
Predicted Class

0 78 60 56.5%
2
0.0% 35.5% 27.3% 43.5% for testing, resulting in 10 combinations.

3
0 22 36 62.1% Table 10. Categorization of data sets for healthy
0.0% 10.0% 16.4% 37.9%
bearings and bearings with real damages.
100% 78.0% 36.0% 60.9%
0.0% 22.0% 64.0% 39.1% Healthy Outer ring damage Inner ring damage
1
(Class 1) (Class 2) (Class 3)
2 3
K001 KA04 KI04
Target Class
K002 KA15 KI14
Figure 11. Confusion matrix for SVM-PSO trained with K003 KA16 KI16
features extracted from MCS of bearings with artificially K004 KA22 KI18
induced damages and tested with MCS features of bearings K005 KA30 KI21
with real damages.
Using features extracted from vibration signals, the Similarly, features from both MCS and vibration signals
ensemble algorithms with majority voting shows the best were used separately. The mean classification accuracy for
performance of all algorithms with 75% classification the 10 combinations was computed and is presented in
accuracy and no false positives or false negatives (Figure Table 11. Random Forest and the ensemble algorithm have
12). However, bearings with outer ring damages have the the highest classification rates of 93.3% using the MCS data
highest misclassification rate, with the majority of samples while CART, RF, and ensemble have the highest
being classified as having inner ring damages. classification rate of 98.5% using the vibration data.
However, the overall performance of other algorithms
remains relatively the same regardless of the sensor data
used.
Figure 13 is the confusion matrix of the ensemble
predictions using features from MCS. There are no false
positives or negatives. However, the highest
misclassification rate is recorded with the outer ring damage
(class 2), as is the case with the ML methods trained with
data of bearings with artificially induced damages. A closer
look at the results indicates that the misclassified samples
belong to the data sets KA04 and KA22 which have
damages of level 1 (see Table 5).

12
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

7.4 Training with Data of Bearings with Multiple

Damages
Table 11. Performance of various algorithms trained and
tested with features of bearings with real damages. Further tests were carried out by including data sets from
bearings with multiple damages at both raceways, at inner
Algorithm Classification Accuracy [%] and outer ring (KB23, KB24 and KB27). According to the
MCS Vibration Signals dominant damage type KB23 and KB24 were assigned to
CART 66.7 98.3 the inner ring damage (class 3), while KB27 was assigned to
RF 83.3 98.3 the outer ring damage (class 2).
BT 81.7 83.3
NN 65.8 44.2 Training and testing was done in a 5-fold cross validation
SVM-PSO 56.7 75.8 manner based on the KAxx and KIxx data sets as described
ELM 69.2 60.8 in section 7.3. In addition, KB23 and KB24 were used
kNN 68.3 62.5 interchangeably for training and testing in a way that when
one was used for training, the other was used for testing.
Ensemble 93.3 98.3
KB27 was only used for testing since it is the only data set
where the outer ring damage is dominant. Table 13 shows
Figure 14 is the confusion matrix of the ensemble of the average performance of the algorithms for this case. For
algorithms using features from vibration signals. A few the neural networks, parameters such as the number of
misclassification instances are observed in samples with hidden layers and neurons were not tuned and this explains
outer ring (KA04) and inner ring (KI21) damages also with the poor performance.
damages of level 1.
Table 13. Performance of various algorithms trained and
MCS tested with features of bearings with real damages.
400 0 0 100%
1 33.3% 0.0% 0.0% 0.0% Classification Accuracy [%]
Algorithm
MCS Vibration Signals
Predicted Class

0 320 0 100% CART 86.3

2
0.0% 26.7% 0.0% 0.0%
91.3
RF 68.1 91.3
0 80 400 83.3% BT 63.7 79.4
3 0.0% 6.7% 33.3% 16.7%
NN 66.3 33.2
100% 80.0% 100% 93.3% SVM-PSO 80.6 70.8
0.0% 20.0% 0.0% 6.7%
ELM 81.9 71.9
1 2 3 kNN 61.3 70.0
Target Class Ensemble 86.3 91.3
Figure 13. Confusion matrix for ensemble of
algorithms trained and tested with MCS features Figures 15 and 16 are the confusion matrices for the
of bearings with real damages. ensemble of algorithms with the inclusion of the data sets of
bearings with multiple damages. The inclusion of these data
Vibration Signals
sets reduces classification accuracy, with the highest
misclassification rate falling within the KB27 and with most
400 0 0 100%
1
33.3% 0.0% 0.0% 0.0%
of the samples being classified as having inner ring
damages.
Predicted Class

0 390 10 97.5%
2
0.0% 32.5% 0.8% 2.5%

0 10 390 97.5%
3
0.0% 0.8% 32.5% 2.5%

100% 97.5% 97.5% 98.3%

0.0% 2.5% 2.5% 1.7%
1 2 3
Target Class

Figure 14. Confusion matrix for ensemble of

algorithms trained and tested with vibration features
of bearings with real damages.

13
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

MCS to improve their generalization, for instance through

400
1 33.3%
0 0 100% selection of suitable kernel functions or optimal parameter
0.0% 0.0% 0.0% tuning.
Predicted Class 2 0 280 30 90.3% Conclusion for application: real industry applications with
0.0% 22.6% 2.4% 9.7%
different electromechanical systems will lead to lower rates,
3
0 140 400 73.6% therefore the development and selection of robust
0.0% 11.3% 31.5% 26.4% algorithms is important. Data sets of the benchmark data
100% 66.7% 92.9% 86.3% will support researchers in developing these algorithms.
0.0% 33.3% 7.1% 13.7%
1 2 3 8. CONCLUSION
Target Class
Application of data-driven classification algorithms for
Figure 15. Confusion matrix for ensemble of vibration-based diagnosis of damages in rolling element
algorithms with the inclusion of data sets with multiple bearings has been researched widely and appears to be a
damages (features extracted from MCS). mature approach. From the present results, it is evident that
MCS also have a great potential for fault diagnosis of
external rolling element bearings in electromechanical drive
Vibration Signals systems. The advantage of this approach in general is that
400 0 0 100% no additional sensors are required, consequently reducing
1
33.3% 0.0% 0.0% 0.0% the cost of the condition monitoring system.
Predicted Class

0 326 0 100% The particular novelties of the present research include the
2
0.0% 26.3% 0.0% 0.0%
consideration of damage in external bearings (positioned
3
0 94 420 81.7% outside the electric motor) and using the MCS as input for
0.0% 7.6% 33.9% 18.3% the detection of the damage. Moreover, a wide variety of
100% 77.6% 100% 92.4% bearing damage is considered.
0.0% 22.4% 0.0% 7.6%
This paper presents important steps in the development of
1 2 3 condition monitoring methods for electromechanical drive
Target Class
systems. The main features of this study can be summarized
Figure 16. Confusion matrix for ensemble of as:
algorithms with the inclusion of data sets with multiple
damages (features extracted from vibration signals).  Development of a systematic categorization method to
describe specific bearing damages in detail.
7.5 Summary of Classification Results  Generation of artificial and realistic bearing damages.
The following conclusions can be drawn from the results  Synchronous acquisition of vibration and motor current
above: signals with a modular experimental setup to gather
data sets for the development and testing of
 Classifications based on vibration signals achieve classification algorithms.
higher classification accuracy than MCS-based  Verification of measurement data by using the
classifications in all cases. established method of envelope analysis based on the
 Training with artificial damage (7.2) and testing with vibration signals. It proves the proper data acquisition
real damages achieve quite low accuracies. and compares it with the state of the art approaches
 Training and testing with real damages (7.3) from proposed in literature.
accelerated lifetime test yield a lot better results.  Application of ML classification algorithms show that
 Conclusion: For a realistic application of classification they are able to identify the damaged bearings using
methods in CM systems the use of training data with MCS, but require sophisticated training data.
artificial damages is apparently not sufficient. The results show that the MCS can be used together with
 Training and testing with an additional focus on machine learning algorithms to identify and classify bearing
multiple damages (7.4) yield even lower accuracies, as damages within the external drive systems. However, the
an example for different damage behavior according to classification accuracy of the machine learning algorithms
the categorization of possible industrial damages. with the MCS is still low in comparison to vibration signal
The discrepancy in the classification results of some of the approaches. Especially the use of signals from artificially
machine learning algorithms could be attributed to induced damages to identify real damages in the drive
overfitting. Therefore, further investigations should be done systems needs to be further investigated as the damage

14
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

recognition rate is quite low. Moreover, significant of Electrical and Electronics Engineers (Ed.), 2007 IEEE
differences are observed using different groups of real International Symposium on Diagnostics for Electric
damage (single and repetitive damages in comparison to Machines, Power Electronics & Drives (pp. 23–28).
multiple damages). Piscataway, NJ, USA: IEEE.
Therefore, more research should be conducted on improving Herold, T., Piantsop Mbo’o, C., & Hameyer, K. (2013).
bearing damage detection from MCS in order to have low Evaluation of the use of an electrical drive as a sensor
cost and accurate condition monitoring system for for the detection of bearing damage. In Conference on
electromechanical drive systems. To promote further Acoustics, AIA-DAGA 2013. Meran, Italy. Retrieved
research, the experimental raw data is published alongside from http://134.130.107.200/uploads/bibliotest/2013TH
with this publication as a benchmark to develop and test Evaluation.pdf
data-driven classifiers or other condition monitoring International Standards Organization (ISO) (2004). Rolling
methods. More detailed examinations concerning the bearings - Damage and failures - Terms, characteristics
behavior of different damages as e.g. with different damage and causes. ISO 15243:2010. Genève, Switzerland:
levels can be carried out with the provided data and the International Standards Organization.
associated damage description. As only a few aspects of the Kankar, P., Sharma, S. C., & Harsha, S. (2011). Fault
classification behavior could be examined in this diagnosis of ball bearings using machine learning
publication, the examination of CM methods should be methods. Expert Systems with Applications, 38, pp.
intensified using the data sets. 1876–1886.
Moreover, some open questions remain, which could be Kimotho, J. K., & Sextro, W. (2014). An approach for
examined using future data sets. Industrial application feature extraction and selection from non-trending data
requires the detection of damages under variable operating for machinery prognosis. Second European Conference
conditions. In addition, other mechanical faults in the drive of the Prognostics and Health Management Society, July
train and other bearing damages should be investigated 8-10, Nantes.
since this work covers only two main modes of damages out Lessmeier, C., Piantsop Mbo'o, C., Coenen, I., Zimmer, D.,
of a possible six. Finally, the implementation into industry & Hameyer, K. (2012). Untersuchung von
applications and related issues have to be discussed and Bauteilschäden elektrischer Antriebsstränge im
tested. Belastungsprüfstand mittels Statorstromanalyse. In K.
Nienhaus & P. Burgwinkel (Eds.): Vol. 81. Aachener
REFERENCES Schriften zur Rohstoff- und Entsorgungstechnik des
Instituts für Maschinentechnik der Rohstoffindustrie,
Amirat, Y., Choqueuse, V., & Benbouzid, M. (2013).
AKIDA 2012. Aachener Kolloquium für Instandhaltung,
EEMD-based wind turbine bearing failure detection
Diagnose und Anlagenüberwachung (1st ed., pp. 509–
using the generator stator current homopolar component. 521). Aachen: Zillekens.
Mechanical Systems and Signal Processing, 41(1-2),
667–678. doi:10.1016/j.ymssp.2013.06.012 Lessmeier, C., Enge-Rosenblatt, O., Bayer, C., & Zimmer,
D. (2014). Data Acquisition and Signal Analysis from
Bartz, W. J. (1985). Wälzlagertechnik // Lagerungen,
Measured Motor Currents for Defect Detection in
Technologie, Berechnungen, Auswahl, Konstruktion und
Electromechanical Drive Systems. Second European
Anwendung. Sindelfingen: Expert Verlag.
Conference of the Prognostics and Health Management
Bellini, A., Filippetti, F., Tassoni, C., & Capolino, G.-A. Society, July 8-10, Nantes.
(2008). Advances in Diagnostic Techniques for Mbo'o, C. P., Herold, T., & Hameyer, K. (2004). Impact of
Induction Machines. IEEE Transactions on Industrial
the load in the detection of bearing faults by using the
Electronics, 55(12), pp. 4109–4126.
stator current in PMSM's. In XXI International
doi:10.1109/TIE.2008.2007527
Conference on Electrical Machines (ICEM) (pp. 1621–
Blödt, M., Granjon, P., Raison, B., & Rostaing, G. (2008). 1627).
Models for Bearing Damage Detection in Induction Nandi, S., Toliyat, H. A., & Li, X. (2005). Condition
Motors Using Stator Current Monitoring. IEEE
Monitoring and Fault Diagnosis of Electrical Motors—A
Transactions on Industrial Electronics, 55(4), pp. 1813–
Review. IEEE Transactions on Energy Conversion,
1822. doi:10.1109/TIE.2008.917108
20(4), 719–729. doi:10.1109/TEC.2005.847955
Bonnett, A., & Yung, C. (2008). Increased Efficiency
Nectoux, P., Gouriveau, R., Medjaher, K., Ramasso, E.,
Versus Increased Reliability. IEEE Industry Applications Morello, B., Zerhouni, N., & Varnier, C. (2012).
Magazine, 14(1), 29–36. doi:10.1109/MIA.2007.909802 PRONOSTIA : An experimental platform for bearings
Djeddi, M., Granjon, P., & Leprettre, B. (2007). Bearing accelerated degradation tests. IEEE International
Fault Diagnosis in Induction Machine Based on Current Conference on Prognostics and Health Management,
Analysis Using High-Resolution Technique. In Institute PHM'12, July 18 -21, Denver.

15
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

Niknam, S. A., Thomas, T., Hines, W. J., & Sawhney, R. Industrial Electronics Society, 2005. IECON 2005 (pp. 6
(2013). Analysis of Acoustic Emission Data for Bearings pp).
subject to Unbalance. International Journal of Smith, W. A., & Randall, R. B. (2015). Rolling element
Prognostics and Health Management, 2013(015). bearing diagnostics using the Case Western Reserve
Obaid, R. R., Habetler, T. G., & Stack, J. R. (2003). Stator University data: A benchmark study. Mechanical
current analysis for bearing damage detection in Systems and Signal Processing.
induction motors. In IEEE International Symposium on doi:10.1016/j.ymssp.2015.04.021
Diagnostics for Electric Machines, Power Electronics Stack, J. R., Habetler, T. G., & Harley, R. G. (2003,
and Drives (pp. 182–187). August). Fault classification and fault signature
Pacas, M., Villwock, S., & Dietrich, R. Bearing damage production for rolling element bearings in electric
detection in permanent magnet synchronous machines machines. IEEE. Diagnostics for Electric Machines,
(2009). In 2009 IEEE Energy Conversion Congress and Power Electronics and Drives,
Exposition. ECCE 2009 (pp. 1098–1103). Tandon, N., & Choudhury, A. (1999). A review of vibration
Paschke, F., Bayer, C., Bator, M., Mönks, U., Dicks, A., and acoustic measurement methods for the detection of
Enge-Rosenblatt, O., & Lohwe, V. (2013). Sensorlose defects in rolling element bearings. Tribology
Zustandsüberwachung an Synchronmotoren. In F. International, 32(8), pp. 469–480. doi:10.1016/S0301-
Hoffmann & E. Hüllermeier (Eds.), Schriftenreihe des 679X(99)00077-8
Instituts für Angewandte Informatik, VDI 3832 (2013). Measurement of structure-borne sound of
Automatisierungstechnik am Karlsruher Institut für rolling element bearings in machines and plants for
Technologie: Vol. 46. Proceedings / 23. Workshop evaluation of condition. Verein Deutscher Ingenieure
Computational Intelligence. Dortmund, 5. - 6. Dezember e.V., Düsseldorf: Beuth Verlag GmbH.
2013 . Karlsruhe: KIT Scientific Publ.
Villwock, S. (2007). Identifikationsmethoden für die
Patil, M. S., Mathew, J., Rajendrakumar, P. K., & Desai, S. automatisierte Inbetriebnahme und
(2010). A theoretical model to predict the effect of Zustandsüberwachung elektrischer Antriebe
localized defect on vibrations associated with ball (Dissertation). Universität Siegen.
bearing. Special Issue on Advances in Materials and
Yang, H., Mathew, J., & Ma, L. (2005). Fault diagnosis of
Processing Technologies, 52(9), pp. 1193–1201.
rolling element bearings using basis pursuit. Mechanical
doi:10.1016/j.ijmecsci.2010.05.005
Systems and Signal Processing, 19(2), pp. 341–356.
Picot, A., Obeid, Z., Régnier, J., Poignant, S., Darnis, O., & doi:10.1016/j.ymssp.2004.03.008
Maussion, P. (2014). Statistic-based spectral indicator
Yang, Z., Merrild, U. C., Runge, M. T., Pedersen, G. K., &
for bearing fault detection in permanent-magnet
Hakon Børsting. (2009). A Study of Rolling-Element
synchronous machines using the stator current.
Bearing Fault Diagnosis Using Motor’s Vibration and
Mechanical Systems and Signal Processing, 46(2), pp.
Current Signatures. I F A C Workshop Series, pp. 354–
424–441. doi:10.1016/j.ymssp.2014.01.006
359.
Qiu, H., Lee, J., Lin, J., & Yu, G. (2006). Wavelet filter-
Zarei, J., & Poshtan, J. (2009). An advanced Park’s vectors
based weak signature detection method and its
approach for bearing fault detection. Tribology
application on rolling element bearing prognostics.
International, 42, pp. 213–219.
Journal of Sound and Vibration, 289(4-5), pp.1066–
1090. doi:10.1016/j.jsv.2005.03.007 Zoubek, H., Villwock, S., & Pacas, M. (2008). Frequency
Response Analysis for Rolling-Bearing Damage
Randall, R. B. (2011). Vibration-based Condition
Diagnosis. IEEE Transactions on Industrial Electronics,
Monitoring. Chichester, UK: John Wiley & Sons, Ltd.
55(12), 4270–4276. doi:10.1109/TIE.2008.2005020
Schaeffler Technologies AG & Co. KG. (2015).
Wälzlagerpraxis: Handbuch zur Gestaltung und BIOGRAPHIES
Berechnung von Wälzlagerungen (4. Aufl.). Mainz:
Vereinigte Fachverl. Christian Lessmeier (Dipl.-Ing.) was born in
Bielefeld, Germany in 1985. He received his
Schoen, R. R., Habetler, T. G., Kamran, F., & Bartfield, R.
degree in Mechanical Engineering from
G. (1995). Motor bearing damage detection using stator
Paderborn University, Germany, in 2010.
current monitoring. IEEE Transactions on Industry
Since 2010, he has been working as a research
Applications, 31(6), 1274–1279. doi:10.1109/28.475697
assistant with Prof. Zimmer at the Chair of
Silva, J., & Cardoso, A. (2005). Bearing failures diagnosis Design and Drive Technology, Paderborn
in three-phase induction motors by extended Park's University. He has managed and worked in a publicly
vector approach. In 31st Annual Conference of IEEE funded research project, assists different lectures, and is
presently working on his PhD thesis. His current research

16
 EUROPEAN CONFERENCE OF THE PROGNOSTICS AND HEALTH MANAGEMENT SOCIETY 2016

area focuses on condition monitoring in electro-mechanical the research group Mechatronics and Dynamics at the
drive systems, especially artificial damage generation and University of Paderborn.
systematic database creation.
James Kuria Kimotho (MSc.) studied
mechanical engineering at the Jomo Kenyatta
University of Agriculture and Technology,
Kenya and received his Master of Science
degree in 2009. Since 2012, he is with the
research group Mechatronics and Dynamics at
the University of Paderborn. His research
focuses on prognostics and health management of
mechatronic systems.
Detmar Zimmer (Prof. Dr.-Ing.), born in
1958, received his Doctor of Engineering with
honors in 1989 at the Institute for Machine
Design and Gearings at the University of
Stuttgart. His dissertation was commended by
the GfT (Gesellschaft für Tribologie,
Germany).
From 1990 to 2001, he worked for the drive system and
automation supplier Lenze AG, Germany, initially as a
R&D manager for geared motors; later, he was responsible
for the geared motors business unit as an authorized officer.
He represented Lenze in the scientific board of the FVA
(Forschungsvereinigung Antriebstechnik, Germany).
Since July 2001, he has been holding the Chair of Design
and Drive Technology at Paderborn University, Germany.
He is a member of WiGeP (Wissenschaftliche Gesellschaft
für Produktentwickung), the Scientific Community for
Product Development in Germany, and of the Direct
Manufacturing Research Center (DMRC) in Paderborn. In
Paderborn, he is also responsible for the cooperation of
Paderborn University with the Chinese German Technical
Faculty (CDTF) in Qingdao, China.
His main research interests are theoretical and experimental
investigations of drive train concepts and expansion of their
application limits. A further field of interest is the
optimization of parts, assemblies and machines by
systematic, function- and production-oriented design.
Walter Sextro (Prof. Dr.-Ing.) studied
mechanical engineering at the Leibniz
University of Hannover and at the Imperial
College in London. Afterwards, he was
development engineer at Baker Hughes Inteq
in Celle, Germany and Houston, Texas. Back
as research assistant at the University of
Hanover he was awarded the academic degree Dr.-Ing. in
1997. Afterward he habilitated in the domain of mechanics
under the topic Dynamical contact problems with friction:
Models, Methods, Experiments and Applications. From
2004-2009 he was professor for mechanical engineering at
the Technical University of Graz, Austria. Since March
2009 he is professor for mechanical engineering and head of

17

View publication stats