IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 1

Thermal Imaging and Vibration Based Multi-Sensor

Fault Detection for Rotating Machinery
Olivier Janssens, Mia Loccufier, and Sofie Van Hoecke

Abstract—In order to minimize operation and maintenance 4) Feature selection/fusion: possibly not all features are
costs and extend the lifetime of rotating machinery, damaging relevant, or features share information. Hence, they can
conditions and faults should be detected early on and auto- be eliminated or fused together.
matically. To enable this, sensor streams should continuously be
monitored, processed and interpreted. In recent years, infrared 5) Classification: to let the system automatically decide if
thermal imaging has gained attention for said purpose. However, a fault or condition is present, a classification (machine
the detection capabilities of a system that uses infrared thermal learning) algorithm is required. This algorithm deter-
imaging is limited by the modality captured by this single mines the condition using the extracted features.
sensor, as is any single sensor-based system. Hence, within this
article a multi-sensor system is proposed which not only uses
infrared thermal imaging data, but also vibration measurements Region of
Feature
Feature
Image interest selection or Classification
for automatic condition and fault detection in rotating machinery. extraction
extraction
fusion
It is shown that by combining these two types of sensor data,
several conditions/faults and combinations can be detected more Fig. 1: General steps in an image processing/machine learning
accurately than when considering the sensor streams individually. pipeline.

Index Terms—Machine learning, fault detection, preventive For automated infrared thermal imaging-based fault de-
maintenance, feature extraction tection and condition monitoring, the focus of the field has
been on data-driven approaches [10]. Such approaches exploit
phenomena in the data that can be related to the conditions
I. I NTRODUCTION and faults. This is necessary because pre-existing model-based
knowledge that can be extracted from the infrared thermal

I N order to prevent machine failure, prolong the lifetime of

machines, reduce operational costs and enhance operational
uptime, condition monitoring and fault detection are required.
imaging data is currently not available.
In related literature, it has been demonstrated that several
types of faults and conditions in rotating machinery are
This is done by monitoring a set of physical properties of detectable to a certain extent using infrared thermal imaging,
a machine using different types of sensors. In recent years, such as rotor imbalance, misalignment, coupling looseness,
infrared thermal imaging based approaches for online fault lubricant inadequacy and rolling element bearing damages [2]–
detection and condition monitoring have started to appear [13]. For the detection of certain conditions, it is easy to see
as is discussed in the recent review by Touret et al. [1]. why infrared thermal imaging is advantageous, such as for
Infrared thermal imaging enables non-contact, non-intrusive lubricant inadequacy. One of the main purposes of lubricant
and fine-grained temperature measurements, which is ideal is friction control. If there is too little lubricant in the bearing
for automatic fault detection and condition monitoring. In housing, excessive friction will occur, which results in heat
order to let a system autonomously detect conditions and that is observable by an infrared thermal camera. Additionally,
faults, infrared thermal imaging is mostly used in combination if there is too much lubrication in the bearing, churning will
with image processing and machine learning [2]–[11]. Such a occur, also resulting in additional heat that can be observed.
solution often consists of five steps (see also Fig. 1): However, for certain conditions, infrared thermal imaging-
1) Image extraction: from an infrared thermal video, based approaches provide sub-optimal results, such as for
single frames are extracted. the detection of damages to a rolling element bearing. For
2) Region of interest extraction: per frame, regions from example, as demonstrated in [10], the outer-raceway damages
which features have to be extracted are segmented. In can only be detect 45 % of the time using an infrared thermal
this step irrelevant parts in the image are also removed. imaging-based system.
3) Feature extraction: features are extracted from the seg- Fortunately, other commonly used techniques such as
mented areas. These features are designed so that they acoustic emission [14], motor current signature analysis
contain information related to conditions that should be (MCSA) [15] and vibration analysis [16] have proven their
detectable. merits w.r.t. damage detection of rolling element bearings.
Therefore, within this article, additional to infrared thermal
O. Janssens and S. Van Hoecke are with imec - IDLab, Department imaging we focus on vibration analysis as it is a well es-
of Electronics and Information Systems at Ghent University, e-mail: (od- tablished, mature –and arguably the most frequently used–
[email protected])
Mia Loccufier is with the DySC Research Group, Ghent University, technique for bearing fault detection [17], [18]. Vibration
Technologiepark 914, 9052 Zwijnaarde, Belgium analysis, which uses accelerometer data, is often done based
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 2

on the physics of the faults and the dynamics of the system. TABLE I: Summary of the 8 conditions in data set one. The
Therefore, several properties in the frequency spectrum of weights in the column headers indicate the weight of the bolt
the vibration signal have been related to specific conditions added to the rotor.
and damages [16]. This knowledge allows to extract specific No imbalance Imbalance: 13 g or 17.3 N
model-based features with well established techniques. Once Healthy bearing (HB) Condition 1 Condition 2
Outer-raceway fault (ORF) Condition 3 Condition 4
processed, faults can be detected and machine’s conditions can Mildly inadequately lubricated bearing Condition 5 Condition 6
be assessed. For example, raceway damages will cause a peak Extremely inadequately lubricated bearing Condition 7 Condition 8

at the bearings’ fault frequencies for which the amplitude can

be used as an indicative feature [16], [19].
As infrared thermal imaging-based approaches suffer from are listed in TABLE III and for the infrared thermal camera
downsides that can be compensated using features extracted in TABLE IV. The rolling element bearing in the housing
from established vibration analysis techniques, a combination at the right-hand side in the set-up was changed in-between
of both approaches in a multi-sensor system will improve test runs, hence this is the housing that was monitored by the
accuracy. Up until now, very little research has been done thermal camera. Additional to the infrared thermal camera, a
on combining vibration data with temperature data. One ex- thermocouple was placed in the room to measure the ambient
ample where a multisensor system is developed is [20]. Their temperature. Furthermore, two accelerometers were mounted
proposed system is designed to detect a cracked rotor, rotor on this bearing housing to measure the accelerations in the
rub or coupling misalignment in a rotating machine, using x-direction and y-direction.
both vibration data and temperature (thermocouple) measure-
ments. It should be noted that no thermal camera is used. A
second relevant example is provided in [21], where MCSA
is combined together with infrared thermography for fault
detection. Their focus lies on faults such as bearing damage,
misalignment and imbalance, but not on the combination of
faults. It should also be noted that in contrast to vibration
analysis, when using MCSA only rotating machinery using
electrical actuators can be considered.
In this paper we propose a system that uses both infrared
thermal imaging data and vibration data, which is the first
of its kind to the best of our knowledge. The system extracts
features from both sensor streams and fuses them together. The
fused features are subsequently provided to a classification
algorithm. We show that not only can the system detect a Fig. 2: 3D image of the set-up. The labels are: 1. servo-motor
multitude of conditions/faults but also combinations of faults with a speed controller; 2. coupling; 3. bearing housing; 4.
and conditions. bearing; 5. disk; 6. shaft; 7. accelerometers; 8. metal plate.
For this research, two data sets were constructed both The red square indicates what the infrared thermal camera
consisting of vibration measurements and infrared thermal captures
imaging measurements. In Section II, the set-up to create these
data sets and the data sets themselves are discussed. Next, To imitate outer-raceway faults (ORF) in the bearing, three
in Section III, an overview is provided of the preprocessing, small shallow grooves were added mechanically on the bear-
feature extraction and classification steps of our system. Sub- ings’ outer-raceway by a field expert (see Fig. 3 for an example
sequently, in Section IV, the results of the proposed system of such a groove). For the respective data sets, two locations
applied on our data sets are discussed. Finally, in Section V, for the ORF faults were chosen. For data set one, the ORF is
a conclusion is provided. located at the 10 o’clock position in the housing. This is close
to the top of the housing facing the infrared thermal camera
II. DATA as to better see the thermal effects of the raceway fault. For
Two data sets were created –both containing vibration mea- data set two, the ORF fault is located at the 6 o’clock position,
surements and thermal imaging data– of conditions and faults which is the loaded zone, as to maximize the impact of the
we induced in a rotating set-up. The conditions present in each raceway fault. A cross-sectional view containing the locations
data set are listed in TABLE I and TABLE II respectively. of these ORFs can be seen in Fig. 4.
Note that a condition in the data sets always exists out of a Both the healthy bearings (HB) and those with an ORF are
combination of two faults/conditions, e.g.: condition 4 in the placed in a housing that contains a grease reservoir enclosing
first data set consists of (1) a bearing with an outer-raceway 20 g of grease [22]. Additionally, at the start 2.5 g of grease
fault while (2) the machine is imbalanced. was added to these bearings as recommended by the manufac-
turer [9]. For the bearings with reduced lubricant in data set
A. Set-up one, i.e. mildly inadequately lubricated bearings (MILB) and
The set-up used to create the two data sets can be seen in extremely inadequately lubricated bearings (EILB), no grease
Fig. 2. A detailed list of the relevant properties of the set-up reservoir is present. For the MILBs, the grease on each individ-
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 3

TABLE II: Summary of the 12 conditions in data set two. The weights in the column headers indicate the weight of the bolt
added to the rotor.

No imbalance Imbalance: 4.1 g or 5.5 N Imbalance: 9.3 g or 12.4 N Imbalance: 13 g or 17.3 N

Healthy bearing (HB) Condition 1 Condition 2 Condition 3 Condition 4
Outer-raceway fault (ORF) Condition 5 Condition 6 Condition 7 Condition 8
Hard particle contamination (HP) Condition 9 Condition 10 Condition 11 Condition 12

TABLE III: Test set-up details

Property Value
Accelerometer type CCDL
Brand Brüel & Kjaer
Accelerometer product type 4533-B
Bearing code FAG 22205-E1-K
Bearing type: Spherical roller bearing with ta-
pered bore and adapter sleeve
Housing code SNV052-F-L
Housing type Closed plummer block
Grease Molykote BR 2 plus
Chosen rotation speed 25 Hz which is 1500 rpm
Eigenfrequency of the set-up deter- 17.8 Hz
mined using an impact test
Accelerometer sample frequency 51,200 Hz
Motor type Single phase asynchronous induc-
tion motor
Motor power 1.1 kW
Motor brand SEW Eurodrive
Motor product ID DRE90M2/FI/LN
Phase three phase
Number of poles 2
Voltage supply 220 – 242 V (star)
and 380 – 420 V (delta)
Motor speed controller MC LTE-B0015-2B1-1-4

Fig. 3: Three shallow grooves in the outer-raceway of a bearing

TABLE IV: Infrared thermal camera details simulating an outer-raceway fault.
Property Value
Thermal camera FLIR SC655
Capture speed 6.25 frames per second (fps) a
Resolution 640 x 480 pixels
Distance: camera - housing 38 cm
Emissivity 0.9
Lens Macro lens
Spectral range 7.5 - 13 µm

ual bearing is superficially reduced (1.5 g reduction). Similarly,

b
for the EILBs the grease in the bearings is decreased more
(0.75 g reduction). For the hard particle faults (i.e. lubricant Fig. 4: Section view of where the outer raceway faults for data
contamination) in data set two, 0.02 g of iron particles are set one (a) and two (b).
mixed in the lubricant of the bearings. Experiments showed
that the chosen lubricant conditions support the objectives of
the research well. B. Data set
To complete the data sets, all the different bearing condi- Every condition in both data sets is created for five different
tions are also tested during imbalance, this is done by adding bearings. By using multiple bearings, variability is introduced
bolts to the rotor at a radius of 5.4 cm. The weight of the in the data set due to manufacturing, mounting and grease
bolts can be found in TABLE I and II. The weight of the bolts distribution. Each bearing is run for one hour, and for the last
results in imbalance that is below the G1 tolerance determined 10 minutes —when steady-state is reached— infrared thermal
according to the ISO 1940. This is a strict tolerance, but en- video is captured together with accelerometer measurements.
sures that the proposed approach will also work for scenarios For data set one, in total, 5 bearings × 8 conditions = 40
where a less strict tolerance is suitable. recordings are made. For data set two, 5 bearings × 12
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 4

conditions = 60 recordings are made. An example of an 1) Vibration measurements: Using expert knowledge and
infrared thermal image of the bearing housing can be seen signal processing, it is possible to detect imbalance by ob-
in Fig. 5. serving an unusually high amplitude at the rotation frequency
of the machine. Hence, the amplitude at the rotation frequency
is extracted as feature1 . The first step to extract this feature
is windowing. The goal of windowing is to minimize the
60 C influence of noise and outliers in the feature extraction step. A
window contains one minute of vibration data, and overlaps by
50 % with its neighbouring window. The size of a window was
determined empirically and with the purpose of being able to
detect the condition of the machine every 30 seconds. Next, the
Discrete Fourier Transform is calculated using the windowed
35 C
signals. A single spectrum covers 25.6 kHz and the resulting
frequency resolution is 0.0167 Hz. Finally, from this frequency
spectrum, the amplitude at the rotation frequency, i.e. 25 Hz,
is extracted as feature. A plot showing the peak at 25 Hz is
shown in Fig. 6. This feature is extracted from the vibration
10 C signals of both accelerometers. As this is done per window,
in the end there are 19 samples per 10 minute recording, each
containing two features. The number of windows is determined
using Eq. 1, where s is the length of the signal –which is 600
Fig. 5: Example of an infrared thermal image of the bearing
seconds–, w the window length –which is 60 seconds– and o
housing.
the amount of overlap between the windows expressed as a
percentage –which is 50 %–.
s−w
III. M ETHODOLOGY number of windows = +1 (1)
w∗o
In our data sets, each vibration and infrared thermal record-
ing is assigned two labels, i.e. one for the machine condi-
tion and one for the bearing condition. Hence, we regard
the fault/condition detection task as a multi-label classifica- 0.175
tion problem and propose a system with two classification
0.150
pipelines. The first pipeline identifies the gradation of imbal-
ance in the rotating machine and the second pipeline identifies 0.125
Amplitude

the bearing’s condition. In the end, the results of both pipelines 0.100
are combined to identify the present condition of the rotating 0.075
machine. This approach reduces the number of classes to be
identified by a single classifier. For example, in data set two, 0.050
there are 12 unique classes. When split into a multi-label 0.025
problem, 3 unique classes remain for the first classifier and 0.000
4 for the second classifier. 0 5 10 15 20 25 30
In order to combine vibration data with infrared thermal Hz
data, feature fusion is employed. Feature fusion entails that
Fig. 6: Example of the peak at the rotation frequency. Note
features are first extracted from the respective data streams
that no spectral leakage occurs.
and are subsequently combined in feature vectors which are
provided to the classification algorithm. Therefore, in this 2) Infrared thermal imaging: In order to extract features
section, first, the features are discussed per pipeline. Features from the infrared thermal imaging data to detect imbalance in
are extracted from the vibration data as well as the infrared the set-up, first windowing is applied similar to the vibration
thermal imaging data. Next, the classification step is discussed. data processing. A window contains one minute of infrared
thermal video and 30 seconds overlap. In the second step,
consecutive frames are differenced, as such an infrared thermal
A. Features: Pipeline one video will reveal imbalance introduced oscillations. To detect
The goal of pipeline one is to detect the gradation of 1 It should be noted that a high amplitude at the rotation frequency can
imbalance regardless of the bearing condition. First, the also be attributed to for example misalignment. However, we solely consider
features extracted from the accelerometer measurements are imbalance in this article and designed the experiment as such. If other faults,
that cause a rise in the amplitude at the rotation frequency, should be detected,
discussed and subsequently, the features from the infrared additional features should be engineered (i.e. amplitude at three times the
thermal imaging data. rotation frequency for misalignment)
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 5

the degree of imbalance, the difference between frames needs Additionally, three statistical features are also extracted,
to be quantified resulting in features. In the third step, for i.e.: kurtosis, crest factor and the root-mean-square (RMS).
each differenced frame, along each column of pixels, the sum Both the kurtosis and the crest factor are widely used statics
of the pixels is taken. Similarly, along each row of pixels, the in bearing fault detection [23]. Additionally, RMS has been
sum of the pixels is taken. The operation on the rows can be shown to be indicative of the amount of separation between
expressed as y = Xj and the operation on the columns as the rolling elements and the raceways due to the lubricant in a
x = XT i; where X ∈ Rmxn represents a frame with m rows linear bearing [24]. Note that all the features, i.e. the amplitude
and n columns; i = [1, 1, ..., 1]T ∈ Rm ; j = [1, 1, ..., 1]T ∈ at the BPFO, kurtosis, crest factor and the RMS are again
Rn and thus x ∈ Rn and y ∈ Rm . Examples of the output calculated on windows extracted from the signal similar to as
vectors displayed in line graphs along the image’s axis are what is done for pipeline one.
shown in Fig. 7. As can be seen, the width (i.e. range) of these 2) Infrared thermal imaging: First, in order to increase
graphs indicate the amount of movement in the differenced robustness against environmental temperature changes, the
frame. Generally, if there is imbalance, the width will be larger. ambient temperature, measured by the thermocouples, is sub-
Hence, the standard deviation, calculated using these output tracted from every pixel’s temperature value as to end up
vectors, is chosen as a feature. In the end, the extracted features with relative temperatures in the infrared thermal videos.
are averaged per window to remove the effect of noise and Second, the relevant components of the set-up are segmented
possible outliers. from the image using a threshold determined by the Otsu
algorithm [25], a commonly used background segmentation
algorithm.
B. Features: pipeline two
Third, similar to pipeline one, the infrared thermal video is
The goal of pipeline two is to identify the specific con- subdivided into overlapping windows. Next, from the prepro-
dition of the bearing regardless of the imbalance gradation cessed infrared thermal imaging data, data-driven features are
in the machine. Again, first the features extracted from the extracted using image processing. These features are extracted
accelerometer measurements are discussed and subsequently, from the individual frames, but averaged over the extracted
the features from the infrared thermal imaging data. windows. As a first feature, the standard deviation (σ) of the
1) Vibration measurements: When a rolling element hits pixel values is chosen. As the temperature distribution varies
an aberration in the outer raceway, the natural frequency of for the different bearing conditions, the standard deviation is a
the raceway is excited. This action results in a high frequency suitable descriptor, e.g. when a grease reservoir is present the
burst of energy which decays rapidly. Afterwards, the natural standard deviation will be higher. The standard deviation of
frequency is excited again as the next rolling element hits the pixel values is not sufficient to discriminate between the
the fault. This high frequency impulse is superimposed, that bearing conditions. Therefore additional features are used.
is, amplitude modulated, on a carrier signal which originates The second feature is a measurement of concentration re-
from the rotating machine. To identify a fault, it is necessary lated to the spatial temperature distribution. The chosen feature
to identify the frequency of occurrence of these high energy is called the Moment of Light or the second-order moment
bursts. Therefore, an established technique, i.e. envelope de- of the pixels collectively containing 20 % of the brightest
tection, is employed [23]. First, a high pass filter is applied. pixels (M20 ) [26]. The M20 is influenced by the presence of
All frequencies below 1 kHz (determined empirically), such as a grease reservoir. Hence, it is very useful to indicate if a
the carrier frequency, are removed. Next, the envelope signal is grease reservoir is present or not. Furthermore, if the M20 and
determined which will have a frequency equal to the frequency the standard deviation are combined, a clear distinction can
of occurrence of the high energy bursts. The envelope is be made between the samples created when a grease reservoir
constructed by taking the magnitude of the analytical signal, is present, i.e. healthy bearings (HB) and the ones with an
which is computed by the Hilbert Huang transform. When outer-raceway fault (ORF), and the samples without a grease
there is an outer-raceway fault, the frequency of the envelope reservoir, i.e. extremely and mildly inadequately lubricated
signal will manifest itself at the ball pass frequency of the bearings (EILB and MILB).
outer raceway (BPFO). Based on these two features, it is not possible to detect the
The BPFO is calculated using Eq. (2), where n is the level of lubricant degradation as the EILB(-IM) and MILB(-
number of rolling elements in the rolling element bearing, f IM) samples overlap. Therefore, the Gini coefficient is added
the rotation frequency, d the diameter of the rolling elements, which proves to be useful to detect the different levels of
D the diameter of the rolling-element cage and α the contact lubrication inadequacy. However, none of these features clearly
angle. This results in a BPFO at 150.41 Hz for the chosen help to distinguish between a bearing with an ORF and a HB
bearings. This envelope detection method is applied on the as is also observable when plotting the standard deviation of
vibration signal coming from the accelerometer on top of the pixels and the Gini coefficient in Fig. 8. The scatter plots
the housing. From the resulting frequency spectrum of the show that there is some overlap between the ORF and HB
envelope signal, the amplitude at the BPFO is chosen as samples2 . This illustrates the weakness of the infrared thermal
feature. imaging-based approach, which can be solved by combining

1 d
BP F O = nf (1 − cos α) (2) 2 For more information regarding these feature the reader is referred to [9].
2 D
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 6

0 0

100 100
y-coordinate

y-coordinate
200 200

300 300

400 400

0 1 2 3 4 5 6 7 0 50 100150200250300350
Sum Sum
14 80
70
12 60
10
Sum

Sum
8 50
40
6 30
4 20
2 100
0
0 50 100 150 200 250 0 50 100 150 200 250
x-coordinate x-coordinate
(a) (b)
Fig. 7: Examples of the line graphs representing the sums along the x axis and y axis for a healthy bearing when the set-up
is in balance (a) and imbalance (b). Note that in (a), the differenced frame is almost entirely black due to the fact that there
is no movement in the infrared thermal video.

the infrared thermal features with the features extracted from (vibration feature) together with the standard deviation along
the vibration measurements. the y-axis (infrared thermal imaging feature) are used.
TABLE V: Features used for the multi-sensor system for both
C. Classification data sets and pipelines. Note that pipeline one detects the
At this point, the features which are extracted from the different imbalance gradations and pipeline two the different
vibration measurements and the infrared thermal imaging bearing conditions.
data in each pipeline have been presented and discussed. Data set Pipeline Vibration features Infrared thermal imaging features
Next, per pipeline, these features are provided to classification One One Amplitude at the rotation frequency Standard deviation along the y-axis
algorithms. As mentioned at the start of this section, feature One Two Amplitude at the BPFO Gini coefficient
fusion is used to combine the two types of sensor data. Feature Amplitude at the rotation frequency Standard deviation
M20
fusion is chosen to enable a single classifier to use information
Two One Amplitude at the rotation frequency Standard deviation along the y-axis
from both modalities. Therefore, both the features extracted
from the vibration measurements and the features extracted Two Two Amplitude at the BPFO frequency
RMS
Standard deviation

from the infrared thermal imaging data are combined in feature Kurtosis
Crest
vectors and provided to classification algorithms. To illustrate: Amplitude at the rotation frequency
one trial using the set-up results in 10 minutes of vibration
measurements and 10 minutes of infrared thermal video. After As classification algorithm, in this paper, the random for-
applying windowing and feature extraction we end up with est classifier is used [27]. A random forest classifier is an
19 samples, each containing two features for the 10 minutes ensemble approach that combines several decision trees. The
vibration signal for pipeline one. For the 10 minutes infrared decision trees are built, i.e. learned, using a subset of features
thermal video we end up with 19 samples each containing two and a subset of samples which are chosen at random with
features (also for pipeline one). By fusing (i.e. concatenating) replacement. A decision tree learns to optimally split the
these two sets of samples together, we end up with 19 samples training data into regions by minimizing the gini score using
containing four features. For dataset one and pipeline one there an optimal feature and feature value. A classification can
are in total 760 samples (i.e.: 19 samples × 8 conditions × 5 subsequently be done by traversing the decision tree from
bearings) each containing 4 features. top to bottom and assigning the sample to the class of
In order to remove features that do not have a positive im- the final node in the decision tree. It should be noted that
pact on the classification results, (exhaustive) feature selection these decision trees can be created in parallel, making the
was done. In TABLE V, the selected features per pipeline algorithm very efficient. Random forest classifiers have many
are listed. In this table, it can for example be seen that for additional advantages, such as they are not prone to overfit,
imbalance detection the amplitude at the rotation frequency require little preprocessing, can deal with both continuous
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 7

60 60
EILB EILB
55 MILB 55 MILB
HB HB
ORF
50 50

45 45

40 40
SD

SD
35 35

30 30

25 25

20 20
0.660 0.665 0.670 0.675 0.680 0.685 0.690 0.660 0.665 0.670 0.675 0.680 0.685 0.690
Gini Gini

(a) (b)
Fig. 8: 2D plots of the Gini coefficient with and without the ORF samples. (a) No overlap exist between the extremely and
mildly inadequately lubricated bearing (EILB and MILB) samples. (b) Some overlap exists between the samples from the
healthy bearings and the ones with an outer-raceway fault (HB and ORF).

and discrete data and require very few hyperparameters to machine learning system for testing. This is done n times,
be tuned. Furthermore, in two very recent comparisons of a where n equals the number of bearings in the data set (i.e. 5).
plethora of machine learning algorithms, the random forest To average out the fluctuations in the results due to randomness
classifier always ended up as one of the best performing which are inherent to random forest classifiers, the five-fold
approaches currently available [28], [29]. cross-validation is executed 10 times.
To optimize the hyperparameters (e.g.: number of trees In order to put the results of the multi-sensor system in
and maximum depth of the individual trees) grid search perspective, we also include the results of two single-sensor
was applied. This is done by defining a set of plausible systems that use either features extracted from vibration data
hyperparameters and testing every combination of these hy- or infrared thermal imaging data.
perparameters. For brevity, only the results obtained using the
optimal hyperparameters are reported next.
C. Results: data set one
IV. R ESULTS The results for the two single sensor systems and the
Within this section, the results are presented and discussed multi-sensor system can be seen in TABLE VI. A perfect
of the proposed multi-sensor approach for both data sets. imbalance detection is achieved with both the single-sensor
systems and the multi-sensor system. However, detecting the
A. Evaluation score specific bearing condition (EILB, MILB, ORF or HB) is more
difficult. When observing the confusion matrix for the single-
In order to evaluate how well the system performs, accuracy
sensor infrared thermal imaging-based system (Fig. 9b) and the
is chosen as evaluation score. This score specifies the ratio
vibration based system (Fig. 9a), it can be concluded that the
between the number of samples that are correctly classified
infrared thermal imaging-based system has difficulty detecting
and all the samples in total. By optimizing the accuracy, the
outer-raceway faults. Conversely, the vibration-based system
false positives (i.e. flagging a sample as containing data from
has more difficulty detecting lubrication related conditions.
a faulty bearing, while no fault or damage is present) and
Both single-sensor systems exhibit a specific weakness that
false negatives (i.e. flagging a sample as containing data from
can be compensated by the strengths of the other sensor.
a healthy bearing, while a fault or damage is present) are
Hence, combining them in a multi-sensor solution results
minimized.
in a better overall system (Fig. 9c). Furthermore, there is
an additional general improvement in all classes due to the
B. Evaluation procedure interaction of features in the machine learning algorithms on
To objectively evaluate the performance of the systems, per- top of eliminating the weaknesses of the respective single
bearing cross validation is employed. This procedure is a type sensor solutions. This way, for example, outer-raceway faults
of cross validation wherein all the data, i.e. extracted features can be classified perfectly (100 % accuracy), while at most
from the measurements, from all bearings but one are used 90 % accuracy could be achieved using vibration data only,
to train the machine learning system and the data collected respectively 74 % using infrared thermal imaging data only.
from trials using the remaining bearing is given to the trained Using a multi-sensor solution, the classification task becomes
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 8

easier as more information is available resulting in a perfect data set one. Conversely, vibration-based bearing condition
classification results. detection achieves a very high accuracy, and the classes do
not get confused very often (Fig. 10a). When features from
TABLE VI: Classification results for data set one. σ is the both systems are combined, the three bearing conditions can
standard deviation. IR = infrared based, VIB = vibration be detected perfectly (Fig. 10c).
based. MILB = mildly inadequately lubricated bearing, EILB
= extremely inadequately lubricated bearing, HB = healthy TABLE VIII: Classification results for data set two. σ is the
bearing, ORF = bearing with an outer-raceway fault. standard deviation. IR = infrared based, VIB = vibration based.
HB = healthy bearing, ORF = bearing with an outer-raceway
IR VIB Conditions Accuracy fault, HP = bearing with hard particles.
IR MILB, EILB, HB, ORF 88.25 % (σ = 8.07 %)
VIB MILB, EILB, HB, ORF 87.25 % (σ = 8.10 %) IR VIB Conditions Accuracy
IR VIB MILB, EILB, HB, ORF 100.00 % (σ = 0.00 %)
IR HP, ORF, HB 65.00 % (σ = 16.16 %)
VIB HP, ORF, HB 91.67 % (σ = 12.91 %)
IR balance and imbalance 100.0 % (σ = 0.00 %) IR VIB HP, ORF, HB 100.00 % (σ = 0.00 %)
VIB balance and imbalance 100.0 % (σ = 0.00 %)
IR VIB balance and imbalance 100.0 % (σ = 0.00 %)
IR Imbalance gradation 88.33 % (σ = 12.47 %)
VIB Imbalance gradation 75.00 % (σ = 9.13 %)
IR All 8 conditions 88.25 % (σ = 8.07 %) IR VIB Imbalance gradation 90.00 % (σ = 6.24 %)
VIB All 8 conditions 87.25 % (σ = 8.10 %)
IR VIB All 8 conditions 100.0 % (σ = 0.00 %)
IR All 12 conditions 55.00 % (σ = 11.31 %)
VIB All 12 conditions 66.67 % (σ = 21.08 %)
By using a random forest classifier it is possible to extract
IR VIB All 12 conditions 90.00 % (σ = 6.24 %)
how important each feature is to the model [27], which is listed
in TABLE VII for the multi-sensor bearing fault detection.
The table illustrates that both the information coming from The detection of the amount of imbalance is more difficult.
the infrared thermal imaging data and the vibration data is Generally, the infrared thermal imaging-based system seems to
important for the decision making of the classification model. outperform the vibration-based system. However, when both
The effects of adding a certain feature are also directly notice- features from both modalities are used, the overall accuracy
able in the results (TABLE VI). For example, the amplitude at increases. The confusion matrices in Fig. 11 indicate that the
the ball pass frequency of the outer raceway is very important classifier can confuse small imbalance gradation differences.
to the model and adding this feature to the infrared thermal Pipeline one detects the different amounts of imbalance and
imaging features results in an accuracy gain of 9.25 %. In pipeline two detects the specific bearing conditions. In total
total, by using the strengths of both sensors, the accuracy rises there are 12 conditions. When solely using infrared thermal
by 11.75 % compared to an infrared thermal imaging-based imaging data the accuracy score is 55.00 % (σ = 11.31 %) and
system and 12.75 % compared to a vibration-based system. when solely using vibration data the accuracy score is 66.67 %
(σ = 21.08 %). However, when both modalities are combined
TABLE VII: Feature importance in the final multi-sensor there is a major improvement as the multi-sensor system
system according to the random forest classifier. The first three achieves an accuracy score of 90.00 % (σ = 6.24 %). When
features are extracted from the infrared thermal imaging (IR) observing the overall confusion matrix (Fig. 12), it can be
data and the last 2 features from the vibration (VIB) data. σ seen that when imbalance occurs together with contamination
is the standard deviation. in the lubricant, it is more difficult to detect the conditions,
hence there is some small interaction between these two types
Feature Importance of the feature of faults.
Gini Coefficient (IR) 29.41 % (σ = 1.07 %) For completeness, a late fusion approach was also created
Standard deviation (IR) 22.22 % (σ = 3.84 %) and tested for data set two. Late fusion entails the creation of a
M20 (IR) 20.29 % (σ = 3.26 %)
Amplitude at the BPFO (VIB) 25.63 % (σ = 1.82 %) machine learning model for each sensor stream. Subsequently,
Amplitude at the rotation frequency (VIB) 2.45 % (σ = 0.40 %) the output of the machine learning models are combined to
determine a final prediction. As the chosen algorithm, i.e.
random forest classifier, provides probabilities of its predic-
tions, we opted to combine the probabilities of the machine
D. Results: data set two learning models created per sensor stream. The results of the
The results of the two single-sensor system and the multi- experiments are listed in TABLE IX. As can be seen, the
sensor system can be seen in TABLE VIII. For pipeline two, a late fusion approach performs worse compared to the feature
relatively low accuracy is achieved due to the overlap between fusion approach. This can be contributed to the fact that late
the ORF and HB samples using infrared thermal imaging fusion does not allow feature interaction, which we determined
features. As the confusion matrix in Fig. 10b illustrates, the to be beneficial.
ORF class is hard to detect. This was also noticeable for
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 9

(a) (b) (c)

Fig. 9: Confusion matrices of the (a) vibration-based fault detection system, (b) the infrared thermal imaging-based fault
detection system and the (c) multi-sensor fault detection system for data set one.

(a) (b) (c)

Fig. 10: Confusion matrices of the (a) vibration-based fault detection system, (b) the infrared thermal imaging-based fault
detection system and the (c) multi-sensor fault detection system for data set two.

(a) (b) (c)

Fig. 11: Confusion matrices for (a) vibration-based imbalance detection, (b) the infrared thermal imaging-based imbalance
detection and (c) multi-sensor based imbalance detection for data set two.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 10

100

HB 100.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 %

90
4.1 g bolt - HB 20.00 % 60.00 % 20.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 %

80
9.3 g bolt - HB 0.00 % 0.00 % 100.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 %

13 g bolt - HB 0.00 % 0.00 % 0.00 % 100.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 70

ORF 0.00 % 0.00 % 0.00 % 0.00 % 100.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 %
60

4.1 g bolt - ORF 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 100.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 %
True label

50

9.3 g bolt - ORF 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 80.00 % 20.00 % 0.00 % 0.00 % 0.00 % 0.00 %

40
13 g bolt - ORF 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 100.00 % 0.00 % 0.00 % 0.00 % 0.00 %

30
HP 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 100.00 % 0.00 % 0.00 % 0.00 %

4.1 g bolt - HP 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 80.00 % 20.00 % 0.00 % 20

9.3 g bolt - HP 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 20.00 % 80.00 % 0.00 %
10

13 g bolt - HP 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 0.00 % 20.00 % 80.00 %

0
B

RF

P
B

RF

RF

RF

P
H

-H

-H

-H
H

-H

-H

-H

-O

-O

-O

lt

lt

lt
lt

lt

tl

bo

bo

bo
lt

lt

t
bo

bo

bo

l
bo

bo

bo

g
g

13
1

13

4.

9.
1

13
4.

9.

4.

9.

Predicted label

Fig. 12: Confusion matrix of the multi-sensor system for data set 2 for all 12 condition.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 11

TABLE IX: Classification results for data set two using late [10] O. Janssens, V. Slavkovikj, B. Vervisch, K. Stockman, M. Loccufier,
fusion. σ is the standard deviation. IR = infrared based, VIB S. Verstockt, R. V. de Walle, and S. V. Hoecke, “Convolutional neural
= vibration based, HB = healthy bearing, ORF = bearing with network based fault detection for rotating machinery,” Journal of Sound
and Vibration, vol. 377, pp. 331 – 345, 2016.
an outer-raceway fault, HP = bearing with hard particles. [11] T. Bai, L. Zhang, L. Duan, and J. Wang, “Nsct-based infrared image
enhancement method for rotating machinery fault diagnosis,” IEEE
IR VIB Conditions Accuracy Transactions on Instrumentation and Measurement, vol. 65, no. 10, pp.
IR VIB HP, ORF, HB 90.00 % (σ = 6.24 %) 2293–2301, 2016.
IR VIB Imbalance gradation 72.80 % (σ = 12.47 %) [12] O. Janssens, R. V. de Walle, M. Loccufier, and S. V. Hoecke, “Deep
IR VIB All 12 conditions 72.80 % (σ = 12.47 %) learning for infrared thermal image based machine health monitoring,”
IEEE/ASME Transactions on Mechatronics, no. 99, pp. 1–10, 2017.
[13] D. Lopez-Perez and J. Antonino-Daviu, “Application of infrared ther-
mography to failure detection in industrial induction motors: case
V. C ONCLUSION stories,” IEEE Transactions on Industry Applications, 2017.
[14] A.-b. Ming, W. Zhang, Z.-y. Qin, and F.-l. Chu, “Dual-impulse response
In this article, a multi-sensor fault detection system for model for the acoustic emission produced by a spall and the size
rotating machinery is proposed that uses both infrared thermal evaluation in rolling element bearings,” IEEE Transactions on Industrial
imaging and vibration data. In order to achieve this, feature Electronics, vol. 62, no. 10, pp. 6606–6615, 2015.
[15] S. Singh and N. Kumar, “Detection of bearing faults in mechanical sys-
fusion is used wherein model-driven features are extracted tems using stator current monitoring,” IEEE Transactions on Industrial
from the vibration measurements, and data-driven features Informatics, vol. 13, no. 3, pp. 1341–1349, 2017.
from the infrared thermal imaging data. Subsequently, the [16] W. A. Smith and R. B. Randall, “Rolling element bearing diagnostics
using the case western reserve university data: A benchmark study,”
extracted features are fused together and provided to random Mechanical Systems and Signal Processing, vol. 64 – 65, pp. 100 –
forest classifiers for the actual fault detection. We show that 131, December 2015.
by using the multi-sensor approach, the shortcomings, inherent [17] I. El-Thalji and E. Jantunen, “A summary of fault modelling and
predictive health monitoring of rolling element bearings,” Mechanical
to the modality (e.g. vibrations or heat), are compensated Systems and Signal Processing, vol. 60, pp. 252–272, 2015.
by the other modality type. Therefore, the system is able to [18] H. Henao, G.-A. Capolino, M. Fernandez-Cabanas, F. Filippetti,
outperform single sensor based systems on our two data sets. C. Bruzzese, E. Strangas, R. Pusca, J. Estima, M. Riera-Guasp, and
S. Hedayati-Kia, “Trends in fault diagnosis for electrical machines: A
We demonstrate that our proposed multi-sensor approach that review of diagnostic techniques,” IEEE industrial electronics magazine,
includes infrared thermal imaging can provide a significant im- vol. 8, no. 2, pp. 31–42, 2014.
provement in fault detection performance (e.g. 35 % absolute [19] B. P. Graney and K. Starry, “Rolling element bearing analysis,” Mate-
rials Evaluation, vol. 70, no. 1, pp. 78 – 85, 2012.
improvement for data set two). However, it should be noted [20] A. D. Nembhard, J. K. Sinha, A. J. Pinkerton, and K. Elbhbah,
that the financial costs increases as two sensors are required. “Combined vibration and thermal analysis for the condition monitoring
Hence, a cost-benefit analysis is required when considering a of rotating machinery,” Structural Health Monitoring, vol. 13, no. 3, pp.
281–295, 2014.
multi-sensor approach. [21] A. Garcia-Ramirez, L. Morales-Hernandez, R. Osornio-Rios, A. Garcia-
Perez, and R. Romero-Troncoso, “Thermographic technique as a com-
R EFERENCES plement for mcsa in induction motor fault detection,” in IEEE Interna-
[1] T. Touret, C. Changenet, F. Ville, M. Lalmi, and S. Becquerelle, “On the tional Conference on Electrical Machines (ICEM), 2014, pp. 1940–1945.
use of temperature for online condition monitoring of geared systems [22] Schaeffler, “Fag split plummer block housings of series snv,”
– a review,” Mechanical Systems and Signal Processing, vol. 101, no. https://goo.gl/HvzXpN, 2015, [Online; accessed 28-February-2018].
Supplement C, pp. 197 – 210, 2018. [23] E.-T. Idriss and J. Erkki, “A summary of fault modelling and predictive
[2] A. Widodo, D. Satrijo, T. Prahasto, G.-M. Lim, and B.-K. Choi, “Confir- health monitoring of rolling element bearings,” Mechanical Systems and
mation of Thermal Images and Vibration Signals for Intelligent Machine Signal Processing, vol. 6061, pp. 252 – 272, 2015.
Fault Diagnostics,” International Journal of Rotating Machinery, vol. [24] H. Ohta, Y. Nakajima, S. Kato, and H. Tajimi, “Vibration and acoustic
2012, pp. 1–10, 2012. emission measurements evaluating the separation of the balls and race-
[3] H. Fandino-Toro, O. Cardona-Morales, J. Garcia - Alvarez, and ways with lubricating film in a linear bearing under grease lubrication,”
G. Castellanos-Dominguez, “Bearing Fault Identification using Journal of Tribology, vol. 135, no. 4, 2013.
Watershed-Based Thresholding Method,” ser. Lecture Notes in [25] O. Nobuyuki, “A threshold selection method from gray-level his-
Mechanical Engineering, 2014, pp. 137–147. tograms,” IEEE Transactions on Systems, Man and Cybernetics, vol. 9,
[4] V. T. Tran, B.-S. Yang, F. Gu, and A. Ball, “Thermal image enhancement no. 1, pp. 62– 66, Jan 1979.
using bi-dimensional empirical mode decomposition in combination [26] J. M. Lotz, J. Primack, and P. Madau, “A New Nonparametric Approach
with relevance vector machine for rotating machinery fault diagnosis,” to Galaxy Morphological Classification,” Astronomical Journal, vol.
Mechanical Systems and Signal Processing, vol. 38, no. 2, pp. 601–614, 128, pp. 16—182, Jul. 2004.
2013. [27] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
[5] G.-m. Lim, Y. Ali, and B.-s. Yang, “The Fault Diagnosis and Monitoring Learning, ser. Springer Series in Statistics. New York, NY, USA:
of Rotating Machines by Thermography,” in Engineering Asset Man- Springer New York Inc., 2001.
agement and Infrastructure Sustainability, J. Mathew, L. Ma, A. Tan, [28] R. S. Olson, W. L. Cava, Z. Mustahsan, A. Varik, and J. H. Moore,
M. Weijnen, and J. Lee, Eds. Springer London, 2012, pp. 557–565. “Data-driven advice for applying machine learning to bioinformatics
[6] G.-M. Lim, D.-M. Bae, and J.-H. Kim, “Fault diagnosis of rotating problems,” in Pacific Symposium on Biocomputing, 2018, pp. 192–203.
machine by thermography method on suport vector machine,” Journal [29] C. Zhang, C. Liu, X. Zhang, and G. Almpanidis, “An up-to-date
of Mechanical Science and Technology, vol. 28, no. 8, pp. 2948–2952, comparison of state-of-the-art classification algorithms,” Expert Systems
2014. with Applications, vol. 82, pp. 128–150, 2017.
[7] A. Younus and B. Yang, “Wavelet co-efficient of thermal image analysis
for machine fault diagnosis,” in International Conference on Mechanical
Engineering, 2009, pp. 1–7. R EFERENCES
[8] A. M. Younus and B.-S. Yang, “Intelligent fault diagnosis of rotating [1] T. Touret, C. Changenet, F. Ville, M. Lalmi, and S. Becquerelle, “On the
machinery using infrared thermal image,” Expert Systems with Applica- use of temperature for online condition monitoring of geared systems
tions, vol. 39, pp. 2082–2091, 2012. – a review,” Mechanical Systems and Signal Processing, vol. 101, no.
[9] O. Janssens, R. Schulz, V. Slavkovikj, K. Stockman, M. Loccufier, Supplement C, pp. 197 – 210, 2018.
R. Van de Walle, and S. Van Hoecke, “Thermal image based fault
diagnosis for rotating machinery,” Infrared Physics & Technology,
vol. 73, pp. 78 – 87, 2015.
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS 12

[2] A. Widodo, D. Satrijo, T. Prahasto, G.-M. Lim, and B.-K. Choi, “Confir- plement for mcsa in induction motor fault detection,” in IEEE Interna-
mation of Thermal Images and Vibration Signals for Intelligent Machine tional Conference on Electrical Machines (ICEM), 2014, pp. 1940–1945.
Fault Diagnostics,” International Journal of Rotating Machinery, vol. [22] Schaeffler, “Fag split plummer block housings of series snv,”
2012, pp. 1–10, 2012. https://goo.gl/HvzXpN, 2015, [Online; accessed 28-February-2018].
[3] H. Fandino-Toro, O. Cardona-Morales, J. Garcia - Alvarez, and [23] E.-T. Idriss and J. Erkki, “A summary of fault modelling and predictive
G. Castellanos-Dominguez, “Bearing Fault Identification using health monitoring of rolling element bearings,” Mechanical Systems and
Watershed-Based Thresholding Method,” ser. Lecture Notes in Signal Processing, vol. 6061, pp. 252 – 272, 2015.
Mechanical Engineering, 2014, pp. 137–147. [24] H. Ohta, Y. Nakajima, S. Kato, and H. Tajimi, “Vibration and acoustic
[4] V. T. Tran, B.-S. Yang, F. Gu, and A. Ball, “Thermal image enhancement emission measurements evaluating the separation of the balls and race-
using bi-dimensional empirical mode decomposition in combination ways with lubricating film in a linear bearing under grease lubrication,”
with relevance vector machine for rotating machinery fault diagnosis,” Journal of Tribology, vol. 135, no. 4, 2013.
Mechanical Systems and Signal Processing, vol. 38, no. 2, pp. 601–614, [25] O. Nobuyuki, “A threshold selection method from gray-level his-
2013. tograms,” IEEE Transactions on Systems, Man and Cybernetics, vol. 9,
[5] G.-m. Lim, Y. Ali, and B.-s. Yang, “The Fault Diagnosis and Monitoring no. 1, pp. 62– 66, Jan 1979.
of Rotating Machines by Thermography,” in Engineering Asset Man- [26] J. M. Lotz, J. Primack, and P. Madau, “A New Nonparametric Approach
agement and Infrastructure Sustainability, J. Mathew, L. Ma, A. Tan, to Galaxy Morphological Classification,” Astronomical Journal, vol.
M. Weijnen, and J. Lee, Eds. Springer London, 2012, pp. 557–565. 128, pp. 16—182, Jul. 2004.
[6] G.-M. Lim, D.-M. Bae, and J.-H. Kim, “Fault diagnosis of rotating [27] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical
machine by thermography method on suport vector machine,” Journal Learning, ser. Springer Series in Statistics. New York, NY, USA:
of Mechanical Science and Technology, vol. 28, no. 8, pp. 2948–2952, Springer New York Inc., 2001.
2014. [28] R. S. Olson, W. L. Cava, Z. Mustahsan, A. Varik, and J. H. Moore,
[7] A. Younus and B. Yang, “Wavelet co-efficient of thermal image analysis “Data-driven advice for applying machine learning to bioinformatics
for machine fault diagnosis,” in International Conference on Mechanical problems,” in Pacific Symposium on Biocomputing, 2018, pp. 192–203.
Engineering, 2009, pp. 1–7. [29] C. Zhang, C. Liu, X. Zhang, and G. Almpanidis, “An up-to-date
[8] A. M. Younus and B.-S. Yang, “Intelligent fault diagnosis of rotating comparison of state-of-the-art classification algorithms,” Expert Systems
machinery using infrared thermal image,” Expert Systems with Applica- with Applications, vol. 82, pp. 128–150, 2017.
tions, vol. 39, pp. 2082–2091, 2012.
[9] O. Janssens, R. Schulz, V. Slavkovikj, K. Stockman, M. Loccufier,
R. Van de Walle, and S. Van Hoecke, “Thermal image based fault
diagnosis for rotating machinery,” Infrared Physics & Technology,
vol. 73, pp. 78 – 87, 2015. Olivier Janssens received his Master degree in
[10] O. Janssens, V. Slavkovikj, B. Vervisch, K. Stockman, M. Loccufier, industrial engineering focussing on information and
S. Verstockt, R. V. de Walle, and S. V. Hoecke, “Convolutional neural communication technology from the University Col-
network based fault detection for rotating machinery,” Journal of Sound lege of West Flanders in 2012. Afterwards he ob-
and Vibration, vol. 377, pp. 331 – 345, 2016. tained his PhD degree in engineering technology
[11] T. Bai, L. Zhang, L. Duan, and J. Wang, “Nsct-based infrared image in 2017 at the Ghent University-imec’s IDLab. His
enhancement method for rotating machinery fault diagnosis,” IEEE research focusses on data-driven multi-sensor condi-
Transactions on Instrumentation and Measurement, vol. 65, no. 10, pp. tion monitoring.
2293–2301, 2016.
[12] O. Janssens, R. V. de Walle, M. Loccufier, and S. V. Hoecke, “Deep
learning for infrared thermal image based machine health monitoring,”
IEEE/ASME Transactions on Mechatronics, no. 99, pp. 1–10, 2017.
[13] D. Lopez-Perez and J. Antonino-Daviu, “Application of infrared ther-
mography to failure detection in industrial induction motors: case
stories,” IEEE Transactions on Industry Applications, 2017. Mia Loccufier received a M.S degree of electrome-
[14] A.-b. Ming, W. Zhang, Z.-y. Qin, and F.-l. Chu, “Dual-impulse response chanical engineer, a M.S. degree of automatic con-
model for the acoustic emission produced by a spall and the size trol engineer and a PhD degree in electromechanical
evaluation in rolling element bearings,” IEEE Transactions on Industrial engineering from Ghent University. She is a profes-
Electronics, vol. 62, no. 10, pp. 6606–6615, 2015. sor at the DySC research group of the Department of
[15] S. Singh and N. Kumar, “Detection of bearing faults in mechanical sys- Electrical Energy, Systems and Automation, Faculty
tems using stator current monitoring,” IEEE Transactions on Industrial of Engineering, Ghent University, Belgium. She is a
Informatics, vol. 13, no. 3, pp. 1341–1349, 2017. lecturer in mechanical vibrations, structural dynam-
[16] W. A. Smith and R. B. Randall, “Rolling element bearing diagnostics ics and systems dynamics. The field of research is
using the case western reserve university data: A benchmark study,” focussed on the dynamics of technical systems. The
Mechanical Systems and Signal Processing, vol. 64 – 65, pp. 100 – main research themes are passive control, especially
131, December 2015. nonlinear tuned mass dampers of mechanical systems and structures, dynamics
[17] I. El-Thalji and E. Jantunen, “A summary of fault modelling and of rotating machinery; stability and bifurcation analysis of nonlinear systems
predictive health monitoring of rolling element bearings,” Mechanical and structures, control of underactuated mechanical systems.
Systems and Signal Processing, vol. 60, pp. 252–272, 2015.
[18] H. Henao, G.-A. Capolino, M. Fernandez-Cabanas, F. Filippetti,
C. Bruzzese, E. Strangas, R. Pusca, J. Estima, M. Riera-Guasp, and
S. Hedayati-Kia, “Trends in fault diagnosis for electrical machines: A
review of diagnostic techniques,” IEEE industrial electronics magazine, Sofie Van Hoecke received her MSc in Computer
vol. 8, no. 2, pp. 31–42, 2014. Science from Ghent University in 2003. Following
[19] B. P. Graney and K. Starry, “Rolling element bearing analysis,” Mate- up on her studies, she achieved a PhD in computer
rials Evaluation, vol. 70, no. 1, pp. 78 – 85, 2012. science engineering at the Department of Informa-
[20] A. D. Nembhard, J. K. Sinha, A. J. Pinkerton, and K. Elbhbah, tion Technology at the same university. Her research
“Combined vibration and thermal analysis for the condition monitoring concentrates on the design of multi-sensor architec-
of rotating machinery,” Structural Health Monitoring, vol. 13, no. 3, pp. tures, innovative ICT solutions for care, and multi-
281–295, 2014. sensor condition monitoring. Currently, she is an
[21] A. Garcia-Ramirez, L. Morales-Hernandez, R. Osornio-Rios, A. Garcia- associate professor at Ghent University and senior
Perez, and R. Romero-Troncoso, “Thermographic technique as a com- researcher at IDLab, Ghent University-imec.