This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

FaultNet: A Deep Convolutional Neural Network

for bearing fault classification
Rishikesh Magar1, Lalit Ghule1, Junhan Li1,2, Yang Zhao1,3, and Amir Barati Farimani1
1
Mechanical Engineering Department, Carnegie Mellon University, Pittsburgh, PA 15213 USA
2
Electrical and Computer Engineering Department, Carnegie Mellon University, Pittsburgh, PA 15213 USA
3
Civil and Environmental Engineering Department, Carnegie Mellon University, Pittsburgh, PA 15213 USA
Corresponding author: Amir Barati Farimani (e-mail: [email protected]).

ABSTRACT The increased presence of advanced sensors on the production floors has led to the collection
of datasets that can provide significant insights into machine health. An important and reliable indicator of
machine health, vibration signal data can provide us a greater understanding of different faults occurring in
mechanical systems. In this work, we analyze vibration signal data of mechanical systems with bearings
by combining different signal processing methods and coupling them with machine learning techniques to
classify different types of bearing faults. We also highlight the importance of using different signal
processing methods and their effect on accuracy for bearing fault detection. Apart from the traditional
machine learning algorithms we also propose a convolutional neural network FaultNet which can
effectively determine the type of bearing fault with a high degree of accuracy. The distinguishing factor of
this work is the idea of channels proposed to extract more information from the signal, we have stacked the
‘Mean’ and ‘Median’ channels to raw signal to extract more useful features to classify the signals with
greater accuracy.

INDEX TERMS Convolutional Neural Network, FaultNet, Featurization, Machine Learning

I. INTRODUCTION another paper on signal processing on vibration data, Zoltan

et al., have demonstrated that signal processing techniques
With the advent of the 4th industrial revolution, industries such as Discrete Wavelet and Wavelet Packet Transform are
across the globe are using artificial intelligence (AI) to effective in extracting features from the frequency domain for
improve their processes and increase efficiency to meet the fault detection [4]. Their simulated results proved the
ever-rising customer demands. In this rapidly changing techniques are even capable of predicting abnormalities
landscape of technology, organizations across the globe, have exploring long-term tendencies of the detected signals. Signal
increased the presence of sensors on the production floor with processing techniques have been performed on acoustic
the motivation of gathering data that can give them valuable signals as well. Adam et al. have proposed a signal processing
insights into their processes[1]. This sensory data contains technique named MSAF-RATIO-24-MULTIEXPANDED-
rich information about the machine and its effective analyses FILTER-8. This technique is used on acoustic signals
using AI can contribute significantly towards preventive captured from electric motors and extracted features are used
maintenance, quality control, and increased process to classify motor faults [5].
efficiency[2]. Realizing these obvious benefits of cost- In this work, we will be focusing on the vibration signal data
effective tools like AI, organizations across the world are and will be analyzing the different methods of fault detection
turning towards smarter technologies. in bearings using vibration signals. Vibration data has many
Driven by the keenness of the industry to embrace advanced applications in the areas of structural weakness or looseness,
digital technologies, many researchers are using different rotating component looseness, and validating the presence of
signal processing methods and coupling them with machine resonance. The optimal monitoring of vibration signals can
learning algorithms to address some of the complicated thus help the analysis of machine performance more
research problems. For example, Thomazella et al. used effectively, improve efficiency, and more importantly give us
digital signal processing techniques such as short-time insights about machine health. With bearing failure being one
Fourier transform (STFT) and the ratio of power (ROP) to of the major contributors to the downtime of industrial
extract features from vibrations signals captured to monitor machines, it is very important to address this problem with
chatter phenomenon during the grinding process [3]. In high reliability and reduce the break-down of machines

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

[6][7][8]. As determined by Zhang et al., for rotating machine are highly compatible and effective when addressing the
health monitoring, vibration signal is very important as it bearing fault diagnostics problem.
contains rich information regarding machine health [9]. Most of the deep learning architectures used for bearing fault
Therefore, the analysis of vibration data may help us in the diagnosis are based on Convolutional Neural Network
detection and prevention of faults in bearing. In their study, (CNN). Guo et al. propose a hierarchical adaptive deep
Samanta et al. used time domain statistical features extracted convolution network for bearing fault size prediction. In their
from vibrations signals to classify faults using an artificial paper, they convert the signal data into a 32x32 array and use
neural network[10]. This study was one of the earliest CNNs to accomplish the task. However, their work does not
attempts to utilize the capabilities of deep learning for bearing use other information available from signal data like
fault detection using vibration signal data. Apart from signal skewness, kurtosis, impulse factor, RMS value[16]. Another
featurization, some researchers have also used wavelet work done by Pham et al. proposes a method that converts the
decomposition to extract relevant information from the signal. signal data into its spectrogram which is then fed to VGG16
In a study conducted on wavelet transformation of vibrations for classification[17], [18]. In their paper, they used only four
signals for fault diagnosis, Sun W et al., use a combination of classes and achieved 98.8% accuracy. When compared to
discrete wavelet transforms and envelope analysis using their work, we achieved a comparable accuracy on 10 classes
which they extract the characteristic spectrum of rolling with computationally inexpensive architecture. Pan et al.,
bearing vibration data. Subsequently, a spectrum cross- employed 1D CNN and LSTM, in order to take advantage of
correlation coefficient is then applied to identify different the signal data, in their paper one-dimensional CNN and
operating conditions of rolling bearings[11]. Based on this LSTM are combined into one unified structure by using the
coefficient, different vibration signals are then classified. CNNs output as input to the LSTM to identify the bearing
Building upon the previous works, we use different signal fault types[19]. They also compare the usage of nine different
featurization methods to extract 14 features from the raw featurization techniques and using them with different
vibration signals to classify bearing faults using machine traditional machine learning algorithms. However, Pan et al.,
learning and deep learning approaches. In order to do not use stacked median and mean channels in their work
comprehensively analyze the signal data, we also and use a more computationally heavy framework by
implemented wavelet decomposition on the raw signal and combining the CNN and LSTM approaches. Most of the
couple it with machine learning approaches to evaluate its state-of-the-art works report an accuracy of more than 98% in
performance for bearing fault classification. With the bearing fault detection. Guo et al., in their paper, used
motivation of developing a generalized model, we evaluate Stacked denoising Autoencoders have obtained an accuracy
our machine learning and deep learning approaches on two of 99.83%. However, they separate the data as per the size of
major publicly available datasets for bearing fault the fault and then make predictions and have only 4 classes in
classification. their predictions[20]. As the dataset created by the CWRU
The first dataset that we analyze has been developed by Case bearing center is able to mimic the actual operating
Western Reserve University (CWRU) bearing center[12]. conditions the dataset contains some noisy signals as is
The dataset from the CWRU bearing center will be referred expected in the actual environment. Therefore, the use of
to as the CWRU dataset throughout the paper. The CWRU SDAE has been made particularly by researchers to make
dataset is one of the important datasets in this research area their predictions more resilient to the noise in the
and has been widely used by researchers to benchmark the dataset[21][22]. Another approach used by Li et al. combines
performance of their models. In their study, Smith et al., have the convolutional neural network and Dempster-Shafer
proposed a benchmark for the CWRU dataset using three theory-based evidence fusion. In their work, they demonstrate
different techniques. They have carefully analyzed the adaptability to different loads and report an accuracy of
different ball faults and compared the signal data amongst the 98.92% [23]. LiftingNet by Pan et al. proposes split, predict
faults. Thus, articulating the difference among signals data and, update blocks that are accurately able to predict the
when different types of fault occur. However, they do not use bearing faults and are adaptable to different motor speeds and
the signal featurization techniques that we have employed and loads. However, their approach is not able to gauge the size
don’t use any deep learning models[13]. Many researchers of the fault[24]. Our FaultNet can not only predict the type of
recently have used different deep learning models on the fault but also the size of the fault based on the input signal.
CWRU bearing dataset. In their review paper, Zhang et al., Wang et al. propose the creation of a time-frequency image of
have compiled a comprehensive list of different methods used the signal and classifying them with AlexNet based
by researchers working in this area. Based on their review of architecture[25][26]. When compared to AlexNet which has 5
different methods, it is evident that the best performing deep convolutional layers ours is a relatively inexpensive
learning models have accuracies in the range of (97%- architecture computationally. Roy et al. propose an
99%)[14]. Another recent review paper by Neupane et al., autocorrelation-based methodology for feature extraction
also discusses different bearing fault classification datasets, from a raw signal and then use the random forest classifier for
signal feature extraction techniques, and some of the highly fault classification. They achieve comparable accuracies to
accurate deep learning architectures[15]. Based on both the the deep learning methods discussed earlier[27].
review papers we can conclude that deep learning methods

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

The second dataset that we have considered in this work is II. DATASET PREPROCESSING
the Paderborn University Data Center bearing dataset[28].
From here onwards, the dataset from Paderborn University A. Case Western Reserve University Bearing Dataset
will be referred to as the Paderborn dataset throughout the The test rig to generate the dataset consists of a 2 hp electric
paper. The dataset has vibration as well as motor current motor to the left, driving a shaft on which a torque transducer
signal captured on the test-rig. In the paper proposing the and encoder are mounted in the middle coupled to a
dataset, to extract the features, Fast Fourier Transform (FFT) dynamometer in the right. The torque is applied to the shaft
and power spectral density (PSD) are performed on vibration via a dynamometer and electronic control system. The test
and motor current signal. After feature extraction and feature rig also includes bearings at both the drive end (DE) and fan
selection,18 features emerge for motor current signals, and 15 end (FE) of the motor. The bearing at the DE and FE are
features are extracted for the vibration signal data[29]. Using 6205-2RS JEM and 6203-2RS JEM, respectively. The 6205-
conventional machine learning approaches, Karatzinis et al, bearing used for data collection is a Single Row Deep
achieved the highest accuracy of 98%. However, they do not Groove Radial Ball Bearing with an inner diameter of 25mm,
use advanced deep learning techniques which may possibly an outer diameter of 52mm, and 15mm in width. To collect
increase the accuracy. In another study Zhong et al., the vibration signal data single point faults were artificially
transform the signal using Short-time Fourier Transform
induced using electro-discharge machining (EDM) with fault
(STFT) and use CNN to classify the bearing fault [30]. On
diameters from 7 to 28 mils (0.18 to 0.71mm). The motor
the transformed signal domain, they apply CNN. The average
loads varied from 0 to 3 hp (approximate motor speeds of
accuracy achieved is 97.4%. Compared to their work, our
model yields better results by directly using the raw signal. 1720 to 1797 rpm). The vibration data was collected using
Bin Li et al., have implemented 1 dimensional CNN accelerometers, which were attached to the housing with
architecture and the best result achieved by them is 98.3 % magnetic bases. The data was collected with two sampling
accuracy in fault classification. However, they have not frequencies, one with 12,000 samples per second, and 48,000
explored the 2D CNN method to improve the results. In samples per second, and was processed using MATLAB®.
another study, Pandhare et al., have implemented 2D CNN In their study, the DE & FE bearing data for the normal (N),
for the bearing fault classification on the Paderborn inner race fault (IF), outer race fault (OF), and the rolling
dataset[31]. In their work, they have demonstrated 2D CNN element(ball) fault (BF) conditions was acquired for fault
on 3 different signal types – raw time domain signal, pattern classification where the fault diameters were selected
envelope spectrum, and spectrogram. The maximum accuracy to be 7 mils, 14 mils, and 21 mils.
achieved is with a spectrogram. For raw signals, the achieved TABLE I: CWRU BEARING HEALTH CONDITIONS AND CLASS LABELS
accuracy of 95% is slightly lower when compared to the other Health Fault size Total class
Condition (mm) dataset labels
studies. Another group of researchers, Wang et al, have
proposed a method to use 1D CNN as well as 2D CNN Normal - 280 0
together to predict the fault class in the Paderborn dataset ball fault 0.18 280 1
[32]. They have concatenated the 1D CNN output with 2D ball fault 0.36 280 2
CNN output before passing it on to a fully connected neural ball fault 0.53 280 3
network for classification. Their resultant accuracy for the inner race fault 0.18 280 4
classification task is 98.58%. However, their approach is inner race fault 0.36 280 5
computationally expensive and hence may not be very inner race fault 0.53 280 6
suitable for online deployment. outer race fault 0.18 280 7
In this paper, we propose FaultNet, a CNN based model to outer race fault 0.36 280 8
determine different types of bearing faults with high outer race fault 0.53 280 9
accuracy. The aim of this paper is to set a benchmark for Ten different conditions are investigated to verify the
bearing fault detection using conventional machine learning accuracy of the proposed method in consideration of multiple
algorithms and deep learning techniques on CWRU and fault patterns. The vibration signals of ten health conditions
Paderborn datasets. It is important to note that the base are referred to in table 1. In this paper, we used the data from
architecture for both the datasets is the same and the
the drive end of the test rig. The sampling frequency chosen
performance of FaultNet is not dataset specific, suggesting
is 48 kHz with the load condition being 2 HP at 1750 rpm.
wide applicability and deployability of the model to detect
To analyze and classify different bearing faults we do some
different types of bearing faults. We achieve state-of-the-art
accuracies for both datasets while proposing a different preprocessing steps on the dataset. The rotating speed of the
methodology to extract features from the data. We also study shaft is 1750 rpm and the sampling frequency is 48 kHz
different signal processing techniques and compare implies that approximately 1670 data points will be collected
accuracies of the traditional machine learning algorithms for one revolution. Out of 1670 data points, the first 35 points
when combining different types of signal features and our and last 35 points are ignored to account for the noise in the
own 2D CNN model. data. Thus, 467600 data points of each fault class are chosen
and divided into 280 samples, with 1670 data points. Finally,
we have 2800 samples with 10 different classes with 280

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

samples each. Further details, which introduce the test set-up paper, the features extracted from raw signal data include
and other data collected, can be found at the CWRU Bearing multiple time and time-frequency domain features. Some of
Data Center website. the statistical time domain features that we extract include
mean, variance, standard deviation, root mean square (RMS).
B. Paderborn University Dataset Moreover, features such as kurtosis and skewness are also
This dataset is generated using 32 bearings. The bearing type extracted as these signals are not stationary. In their paper,
used for this dataset generation is 6203, which stands for Caesarendra et al., give us some physical insights into the
Deep Groove Ball Bearings with dimensions (inner diameter, features as they report the approximate values of kurtosis and
outer diameter, and width) – 17X40X12mm. Out of 32, 6 skewness for a normal bearing to be 3 and 1 respectively
bearings are healthy, 12 bearings have artificially created [34]. Hence, for bearings that are faulty, we expect to have
defects and the remaining 14 bearings are naturally damaged. kurtosis and skewness values shifted from 3 and 1. Another
The artificial defects have been created by using drilling, important observation we made was that for faulty bearings,
EDM, and electric engraving machine. The artificial defects the bearing signal amplitude undergoes abrupt changes when
are produced on both, inner and outer race. The natural rolling elements pass over the defective region of the bearing.
damages are produced by accelerated lifetime tests. A These abrupt changes are responsible for disturbing the
detailed description can be found in the paper[28]. overall distribution of signal and therefore can act as an
important clue in detecting faulty bearings. Generally, the
Further, the bearings’ samples can be divided into 3 classes, value of kurtosis increases and skewness may change to the
healthy, inner race fault, and outer race fault. By this negative or positive side for faulty bearings. Apart from these
classification, there are 6 healthy bearings, 11 inner race fault features, dimensionless features such as crest factor, shape
bearings, and 12 outer race fault bearings. This amounts to factor, impulse factor are also extracted. The shape factor is
29 bearings in total. The Remaining 3 bearings are omitted affected by the shape but is independent of the dimension.
due to their nature of the fault. These 3 bearings have inner The crest factor is a measure of an impact when a rolling
as well as outer race fault. In the study conducted by element comes in contact with the raceway. Table 2
Paderborn University, the authors have classified these summarizes all the 14 features extracted from the raw signal
bearings on the grounds of the maximum contributing fault. If data along with their mathematical formulae used.
the inner race damage is more compared to the outer race, the TABLE II: FEATURES AND THE MATHEMATICAL FORMULAE USED TO
bearing is classified as inner race fault bearing. For the CALCULATE VALUES FOR EACH SIGNAL
current model, we used 29 bearings data which can be No Feature Formula
classified distinctively. The data set is generated with 1 Mean Mean =
multiple combinations of rpm, torque, and load. For the 2 Absolute Abs Mean =
purpose of this study, we use the following combination. mean
3 Maximum
N=1500 rpm, load torque=0.7 Nm and Radial force=1000 N.
4 Minimum
Each bearing is used 20 times to generate 20 signals with 5 Peak to Peak Maximum - Minimum
one fixed combination. The signal generated is a vibration 6 Absolute
signal for 4 sec with a sampling frequency of 64kHz. That max
means, in a signal, there are 256,000 data points. To avoid 7 Root Mean
initial and ending noise and disturbance, the sample signal is Square
clipped off for the first 1/16th part and the last 1/16th part.
Eventually, the signal used has 2,24,000 data points which 8 Variance
are used further for featurization. In total 2320 signals have
been used for classification.
9 Clearance
factor
III. FEATURIZATION
In every machine learning process, feature engineering plays
a very important role and can significantly affect the 10 Kurtosis
performance of an algorithm. Feature engineering can directly 11 Skewness
help the machine learning algorithm to identify the
underlying patterns and effectively improve the accuracy of
the model. For signal data, featurization includes deriving
different domains’ features from raw signals such as time 12 Impulse
domain, time-frequency domain, etc. The vibration signals Factor
from machinery components are in general considered to be
non-stationary. The non-stationary signals mean that the 13 Crest Factor
frequencies present in a signal vary with time [33]. Therefore,
it is important to extract features from the time domain as
14 Shape Factor
well as the time-frequency domain to capture the time-
varying nature of frequencies present in a signal. In this

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

Time-frequency domain representation methods such as CWRU dataset (figure 1(b)). Similar to CWRU multiple
short-time Fourier transform (STFT), wavelet transform, and combinations of different features
Wigner-Ville distribution (WVD) are commonly used for the
non-stationary or transient signal. These methods implement
a mapping of one-dimensional time-domain signals to a two-
dimensional function of time and frequency. The objective is
to provide a true time-frequency representation of a signal.
Similar, to the methods presented in the review [35]
conducted by Feng et al., on time-frequency analysis
methods for machinery fault diagnosis, we decomposed these
raw signals by employing wavelet decomposition package
(WPD) using Haar wavelet as a mother wavelet for the
extraction of time-frequency domain features. The wavelet
decomposed signal consists of approximation coefficients
and detailed coefficients. In this work, we use the
approximation coefficients as they are more sensitive
towards bearing conditions as suggested in [28] for the
extraction of statistical features mentioned in table II.

IV. RESULTS WITH SIGNAL FEATURIZATION

(a)

A. Case Western Reserve University Bearing Dataset

We used all the 14 features in Table 2 and evaluated the
performance of different shallow learning algorithms. The
train-test split used was 80-20% and average 5-fold cross-
validation accuracy has been reported. Amongst all the
models we tried, the random forest yielded the highest
accuracy (figure 2(a)). An important functionality of the
random forest algorithm is that it provides feature importance
which gives the user important insights about the features.
The feature importance score of the top 5 features calculated
using Random Forest has been demonstrated in figure 1(a). It
was observed that absolute mean, variance, RMS, shape
factor, and the mean are the 5 most important signal features.
Subsequently, to analyze the effects of these important
features on the accuracy multiple combinations of these
features were used to classify the bearing faults. As expected,
the accuracy improves when a model is trained with a greater
number of features (figure 2(a)). We also implemented (b)
wavelet decomposition on the raw signal data for the CWRU FIGURE 1. Feature importance based on Random Forest results. (a)
dataset extracted the same 14 features from Table 2 on the shows top 5 important features obtained on CWRU bearing dataset.
decomposed signal. Random forest was the best performing Similarly (b) shows the results obtained for Paderborn University
dataset. For both the datasets, out of 5 important features.
algorithm, and it was observed that wavelet decomposition
level 2 and level 3 had a slightly higher performance (figure were tried and the accuracy of different shallow learning
3(a)). After testing multiple algorithms and different signal methods was evaluated (figure 2(b)). We also tested for three
featurization techniques it was observed that the performance different decomposition levels to check the effect of wavelet
did not improve beyond 90% accuracy, to further improve decomposition on the overall accuracy of the model. As
the accuracy we then tried deep learning approaches. demonstrated in figure 3(b), it is observed that there is a
slight increase in accuracy with the decomposition level.
B. Paderborn University Dataset
Similar to the CWRU dataset we decided to use deep
Like the CWRU dataset, we used a train-test split of 80-20% learning to improve the accuracy further.
and report the average five-fold cross-validation accuracy. After analyzing the results from conventional machine
For the Paderborn data, the random forest algorithm gave the learning approaches with signal featurization, we realized
highest accuracy. Utilizing the feature importance that the best accuracy achieved was not comparable with the
functionality of random forest, the top five important features state-of-the-art results discussed earlier. Therefore, we
were calculated. It is important to note that, out of the top decided to build a Convolutional Neural Network (CNN) that
five important features, four features are common in the is computationally inexpensive and also achieves higher
accuracy.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

(b)
FIGURE 3. Classification accuracy using all 14 features on different
signal wavelet decomposition level (a) shows accuracy for CWRU
dataset. Decomposing signals at different result in slightly higher
accuracy. Also, for some ML algorithms decomposed signals gives
(a) better accuracy compared to raw signals.
Similarly (b) shows the results obtained for Paderborn University
dataset.

V. CONVOLUTIONAL NEURAL NETWORK

To further improve the classification performance, we
developed FaultNet a CNN based architecture that takes raw
signal data as input without any pre-processing. CNNs,
because they possess a special ability to extract relevant
features from the data, given a task of prediction. In the
FaultNet architecture, there are two parts, the first part being
the convolution part and the second part is fully connected
layers. We have two convolution layers. Each convolution
layer is followed by a max-pooling layer. The activation
function used for both convolution layers is ‘ReLU’. Usage
of max-pooling layers ensures that the most important
features are selected. The addition of pooling has also led to
decreased computational times making the FaultNet
relatively inexpensive architecture.
(b)
FIGURE 2. Classification accuracy with raw data, all features, and five As we know, the raw signals CWRU dataset contains 2800
important features (a) shows accuracy for CWRU dataset for five signals of 1600 data points. The signal data is converted into
different classification algorithms. It is evident from the figure that
accuracy improves as number of features increases. Only using raw
a 2D array of shape 40 × 40. Therefore, we have 2800 signals
data yields the lowest accuracy. in the form of 2D arrays of shape 40 × 40. Similarly, for the
Similarly (b) shows the results obtained for Paderborn University Paderborn University dataset, signals of 250,000 datapoints
dataset.
are split into 100 smaller signals of shape 50 × 50 2D arrays.
The convolution operation is performed on the 2D data. The
convolution operation is defined by:

After the convolution layers, the output is flattened and fed to

fully connected layers. The fully connected layer has an input
layer with 5184 neurons for the Paderborn dataset. There is
only one hidden layer with 256 units. According to the
dataset, the output layer has either 3 (Paderborn dataset) or
(CWRU dataset) 10 neurons. For fully connected layers as
well, the activation function used is ‘ReLU’. Besides, for the
final layer, a drop-out of 0.25 is added to prevent the
overfitting in the neural network. Soft-max activation is
applied to the outputs of the neurons in the final layer[36]
[37]. The overall network architecture is shown in figure 4.
(a)

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

As the problem’s nature is classification, we use Cross-

entropy loss. For training, we use the ‘Adam’ optimizer with
a constant learning rate of 0.001. The whole network is
trained for 100 epochs with a batch size of 128 on NVIDIA
RTX2080 GPU. At the end of the training, the loss value for
the Paderborn dataset is 0.0003.

VI. RESULTS WITH CONVOLUTIONAL NEURAL

NETWORK
Initially, we tried to predict the bearing fault class with only
raw signals. However, the 5-fold accuracy maxed out at
95.27%. With the state-of-the-art model achieving accuracies
in the range of 97-100%. We devised a methodology to
incorporate more signal information so that the model can
learn the signal features better. Thus, we came with the idea (b)
of using mean and median channels to augment the raw FIGURE 5. Comparison of accuracies obtained for different CNN
signal in a bid to improve the accuracy. In order to generate models.
(a) CWRU dataset with one-channel, two channel and three channel
new channels, a sliding window with a length of 10 was used approach.
as a filter. For every given sample signal data, the filter scans (b) Paderborn dataset with one-channel, two-channel and three channel
approach
through the whole sample data from the front to the end. To
get the same quantity of data points as the sample data, nine Apart from the improved accuracy, we observed that the
‘0’ were replenished at the end of the sampled signal data deviation of accuracy was lowered over the five folds of test
and the filter has been set to shift by length 1 for each time. datasets. Leading us to conclude that channel addition not
Consequently, a new channel, which had the same shape size only improves accuracy but also generates a more stable
as the original channel of sample signal data, is generated by model in terms of accuracy. In order to augment accuracy as
all outputs while the filter completed the data traversal. For an evaluation metric, we also calculate the precision, recall,
the first additional channel, a mean filter was applied to and F1 score. The high values of the F1 score for both
create a mean channel. For the second channel, the mean datasets (Table 3) indicate that the FaultNet architecture is
filter is substituted with a median filter to generate the robust and captures the faults with high precision. When
median channel. We combined new channels with the comparing the performance of FaultNet with CNN
original channel as the new input for the 2D convolution architecture proposed by Zhang et al, we observe that the
model. With an increasing number of channels, the accuracy precision, recall, and F1 score for their method are 0.8155,
improves simultaneously. The average accuracy has already 0.8105, and 0.8129 [38]. However, these values are only
improved to 98.50% as evident from figure 5(a). Similar reported with 90 training data points. When we trained the
improvements in performance were seen on the Paderborn FaultNet with the same number of datapoints we get the
dataset in figure 5(b). precision, recall, and F1 score as 0.799, 0.7924, and 0.7956
respectively. The performance of FaultNet is comparable to
the architecture proposed by Zhang et al., even in a low
data setting considering the fact that they propose a deeper
CNN that has five convolutional layers, and FaultNet has
two convolutional layers. Pham et al., have also calculated
the precision and recall score on different motor settings
and their average scores are 0.9826 and 0.991 respectively
[17]. When compared to their architecture FaultNet has a
slightly higher precision by 0.0034 and slightly lower recall
by 0.0053. The performance of our architecture FaultNet is
comparable to other state-of-the-art methods in all
performance metrics. Despite being a lightweight CNN
architecture, FaultNet is able to achieve comparable results
on all performance metrics because of the novelty in the
way in which we make use of median and mean channels.
Apart from that, we also plot the confusion matrix for both
(a) datasets (Figure 6). We observe that for the CWRU dataset,
the model fails to classify datapoints belonging to ball fault
with 0.18 mm fault size. The model is confused between ball
fault with 0.18 mm fault size and outer race fault with 0.36
mm. This leads us to conclude that the model has a slightly

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

FIGURE 4. Convolutional Neural Network Architecture. 2D CNN architecture used for Paderborn University dataset. Different colors represent different
operations. There are 2 convolution and 2 max-pool layers along with fully connected neural network with 3 outputs. After each convolution layer, a
‘Batch-norm’ and ‘ReLU’ activation is applied.

lower accuracy for outer race faults. Interestingly, the model

is able to classify smaller size faults with high accuracy. It is
evident that class 0 has better accuracy compared to other
classes. Class 2 has the highest number of misclassifications.
The model has difficulty in distinguishing between the inner
race and outer race faults. In general, it is difficult to classify
outer race faults as observed from confusion matrixes of both
datasets.
TABLE III: PERFORMANCE METRICS FOR CWRU AND PADERBORN
DATASET
Metric CWRU Paderborn
Precision 0.9860 0.9906
Recall 0.9857 0.9918
F1 score 0.9857 0.9915
(a)
VII. PERFORMANCE EVALUATION FOR NOISE
ROBUSTNESS
To evaluate the robustness of the FaultNet architecture and
different conventional machine learning algorithms to noise
we added white Gaussian noise to the vibration data and
assessed the fault classification performance. We chose
seven different signal to noise ratios (SNR) to understand
how noise affects the performance of the different
algorithms. The results for the CWRU and Paderborn
datasets are demonstrated in Table IV and Table V,
respectively. Among the conventional machine algorithms,
we observed that random forest had the highest accuracy
and was more robust to noise for both the datasets (CWRU
(b) and Paderborn). FaultNet achieves a high accuracy of
FIGURE 6. Confusion matrix for comparing the inter-class performance.
(a) CWRU dataset: Classes from 0 to 9 correspond to the labels given in
97.77% (CWRU) and 98.8% (Paderborn) when the SNR is
table I. The test accuracy corresponding to this confusion matrix is 10. It can be observed that the accuracy is slightly lower
98.57%. when compared to the original vibration signal. For all the
(b) Paderborn dataset: Class 0 corresponds to healthy bearing. Class 1
and class 2 represent inner race fault bearing and outer race fault conventional machine learning algorithms and CNN-based
bearing, respectively. Overall accuracy is 99.14%. FaultNet, accuracy increases as the SNR goes up. It must be
noted that for the noisiest signal with SNR equal to –4,

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

FaultNet performs reasonably well with an accuracy of working exactly opposite to the image conversion from RGB
82.12% and 89.3% for CWRU and Paderborn respectively. to grayscale.
When compared with deep learning-based architecture by The novelty of this work is in the concise CNN structure,
Zhang et al., FaultNet achieves comparable accuracy within also, in the concept of augmenting 2D raw signal with its
1% for SNR values of 8 and 10 and considerably mean and median value channels to extract more meaningful
outperforms it when the SNR values are less than 2 on the features for CNN. We have demonstrated that the CNN
CWRU dataset [38]. We would like to note that CNN structure devised here improves upon previous methods and
proposed by Zhang et al., consists of 5 convolutional layers has a highly competitive performance compared with state-
whereas our lightweight architecture FaultNet only has 2 of-the-art methods. We believe this work can pave the way
convolutional layers, making it more suitable for an online for online fault detection in the case of bearings which could
industrial setting. FaultNet is able to achieve high accuracy be extremely beneficial for industries. Our approach can be
because of the novel way in which it is able to use extended to similar types of datasets.
information from signals through mean and median
channels.
TABLE IV: PERFORMANCE EVALUATION OF FAULTNET FOR NOISE
ROBUSTNESS ON CWRU DATASET
SNR RF SVC LR kNN MLP FaultNet
-4 75.16 67.14 68.75 66.25 73.64 82.12
-2 77.27 69.01 70.14 68.32 74.89 84.14
0 79.04 71.52 71.71 69.39 76.03 87.68
2 81.65 73.24 72.17 68.75 76.21 90.74
6 82.38 73.89 72.78 69.25 76.42 93.17
8 84.29 74 74.07 69.10 76.62 96.21
10 85.62 74.78 74.32 69.32 77 97.77

TABLE V: PERFORMANCE EVALUATION OF FAULTNET FOR NOISE

ROBUSTNESS ON PADERBORN DATASET
SNR RF SVC LR kNN MLP FaultNet
-4 87.73 80.72 64.42 80.63 76.12 89.3
-2 89.27 82.47 67.97 81.96 78.41 93.3
0 90.04 83.86 68.68 82.62 79.03 94.1
2 90.65 85 69.6 83.34 82.03 96.5
6 92.38 86.64 70.2 84.12 82.82 97.7
8 94.29 87.44 72.07 84.96 83.5 98.2
10 95.22 88.08 71.67 85.57 84.16 98.8

VIII. CONCLUSION
In this paper, a systematic approach towards a data-driven
vibration-based diagnosis of faults in rolling element bearings
is demonstrated. We have benchmarked the performance of
different machine learning algorithms by using the featurized
signal data and deep learning approaches for the CWRU and
Paderborn datasets. Five-fold accuracies of ~99% are
obtained for both the datasets indicating the state-of-the-art
performance is achieved by the FaultNet architecture. For
classification, it is important to have enough differentiating
features between classes. As we stack a greater number of
channels, the algorithm is able to extract more features
compared to single-channel input. Each additional channel
could be considered as a feature map of the input which
provides more information about the input. This is analogous
to grayscale and RGB images. If we convert an RGB image
to a grayscale image, often, there happens to be information
loss leading to poor performance [39]. Hence, adding more
information improves the performance of our model by

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

REFERENCES: no. 2, pp. 317–328, Mar. 2003, doi:

[1] J. Lee, E. Lapira, B. Bagheri, and H. Kao, “Recent 10.1006/mssp.2001.1462.
advances and trends in predictive manufacturing [11] W. Sun, G. An Yang, Q. Chen, A. Palazoglu, and K.
systems in big data environment,” Manufacturing Feng, “Fault diagnosis of rolling bearing based on
Letters, vol. 1, no. 1, pp. 38–41, Oct. 2013, doi: wavelet transform and envelope spectrum
10.1016/j.mfglet.2013.09.005. correlation,” Journal of Vibration and Control, vol.
[2] Kim, D., Kim, T.J.Y., Wang, X. et al., “Smart 19, no. 6, pp. 924–941, Mar. 2012, doi:
Machining Process Using Machine Learning: A 10.1177/1077546311435348.
Review and Perspective on Machining Industry,” [12] “http://csegroups.case.edu/bearingdatacenter/home.” .
International Journal of Precision Engineering and [13] W. A. Smith and R. B. Randall, “Rolling element
Manufacturing-Green Technology, pp. 555–568, bearing diagnostics using the Case Western Reserve
Aug. 2018, doi: https://doi.org/10.1007/s40684-018- University data: A benchmark study,” Mechanical
0057-y. Systems and Signal Processing, vol. 64–65, pp. 100–
[3] R. Thomazella, W. N. Lopes, P. R. Aguiar, F. A. 131, Dec. 2015, doi: 10.1016/j.ymssp.2015.04.021.
Alexandre, A. A. Fiocchi, and E. C. Bianchi, “Digital [14] S. Zhang, S. Zhang, B. Wang, and T. G. Habetler,
signal processing for self-vibration monitoring in “Deep Learning Algorithms for Bearing Fault
grinding: A new approach based on the time- Diagnostics—A Comprehensive Review,” IEEE
frequency analysis of vibration signals,” Access, vol. 8, pp. 29857–29881, 2020, doi:
Measurement, vol. 145, pp. 71–83, Oct. 2019, doi: 10.1109/ACCESS.2020.2972859.
10.1016/j.measurement.2019.05.079. [15] D. Neupane and J. Seok, “Bearing Fault Detection
[4] Z. Germán-Salló and G. Strnad, “Signal processing and Diagnosis Using Case Western Reserve
methods in fault detection in manufacturing University Dataset With Deep Learning Approaches:
systems,” Procedia Manufacturing, vol. 22, pp. 613– A Review,” IEEE Access, vol. 8, pp. 93155–93178,
620, Jan. 2018, doi: 10.1016/j.promfg.2018.03.089. 2020, doi: 10.1109/ACCESS.2020.2990528.
[5] A. Glowacz, “Acoustic fault analysis of three [16] X. Guo, L. Chen, and C. Shen, “Hierarchical adaptive
commutator motors,” Mechanical Systems and Signal deep convolution neural network and its application
Processing, vol. 133, p. 106226, Nov. 2019, doi: to bearing fault diagnosis,” Measurement, vol. 93, pp.
10.1016/j.ymssp.2019.07.007. 490–502, Nov. 2016, doi:
[6] L. Jiang, T. Shi, and J. Xuan, “Fault diagnosis of 10.1016/j.measurement.2016.07.054.
rolling bearings based on Marginal Fisher analysis,” [17] M. T. Pham, J.-M. Kim, and C. H. Kim, “Accurate
Journal of Vibration and Control, vol. 20, no. 3, pp. Bearing Fault Diagnosis under Variable Shaft Speed
470–480, Nov. 2012, doi: using Convolutional Neural Networks and Vibration
10.1177/1077546312463747. Spectrogram,” Applied Sciences, vol. 10, no. 18, p.
[7] B. Van Hecke, Y. Qu, and D. He, “Bearing fault 6385, Sep. 2020.
diagnosis based on a new acoustic emission sensor [18] K. Simonyan and A. Zisserman, Very Deep
technique,” Proceedings of the Institution of Convolutional Networks for Large-Scale Image
Mechanical Engineers, Part O: Journal of Risk and Recognition. 2014.
Reliability, vol. 229, no. 2, pp. 105–118, Nov. 2014, [19] H. Pan, X. He, S. Tang, and F. Meng, “An Improved
doi: 10.1177/1748006X14558900. Bearing Fault Diagnosis Method using One-
[8] J. Xu, S. Tong, F. Cong, and Y. Zhang, “The Dimensional CNN and LSTM,” Strojniški vestnik -
application of time–frequency reconstruction and Journal of Mechanical Engineering, 2018, doi:
correlation matching for rolling bearing fault 10.5545/sv-jme.2018.5249.
diagnosis,” Proceedings of the Institution of [20] X. Guo, C. Shen, and L. Chen, “Deep Fault
Mechanical Engineers, Part C: Journal of Recognizer: An Integrated Model to Denoise and
Mechanical Engineering Science, vol. 229, no. 17, Extract Features for Fault Diagnosis in Rotating
pp. 3291–3295, Apr. 2015, doi: Machinery,” Applied Sciences, vol. 7, no. 1, p. 41,
10.1177/0954406215584397. Dec. 2016.
[9] X. Zhang, N. Hu, L. Hu, L. Chen, and Z. Cheng, “A [21] C. Lu, Z.-Y. Wang, W.-L. Qin, and J. Ma, “Fault
bearing fault diagnosis method based on the low- diagnosis of rotary machinery components using a
dimensional compressed vibration signal,” Advances stacked denoising autoencoder-based health state
in Mechanical Engineering, vol. 7, no. 7, p. identification,” Signal Processing, vol. 130, pp. 377–
1687814015593442, Jul. 2015, doi: 388, Jan. 2017, doi: 10.1016/j.sigpro.2016.07.028.
10.1177/1687814015593442. [22] F. Xu, W. tai P. Tse, and Y. L. Tse, “Roller bearing
[10] B. SAMANTA and K. R. AL-BALUSHI, fault diagnosis using stacked denoising autoencoder
“ARTIFICIAL NEURAL NETWORK BASED in deep learning and Gath–Geva clustering algorithm
FAULT DIAGNOSTICS OF ROLLING ELEMENT without principal component analysis and data label,”
BEARINGS USING TIME-DOMAIN FEATURES,”
Mechanical Systems and Signal Processing, vol. 17,

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

Applied Soft Computing, vol. 73, pp. 898–913, Dec. [32] D. Wang, Q. Guo, Y. Song, S. Gao, and Y. Li,
2018, doi: 10.1016/j.asoc.2018.09.037. “Application of Multiscale Learning Neural Network
[23] S. Li, G. Liu, X. Tang, J. Lu, and J. Hu, “An Based on CNN in Bearing Fault Diagnosis,” Journal
Ensemble Deep Convolutional Neural Network of Signal Processing Systems, vol. 91, no. 10, pp.
Model with Improved D-S Evidence Fusion for 1205–1217, Oct. 2019, doi: 10.1007/s11265-019-
Bearing Fault Diagnosis,” Sensors (Basel), vol. 17, 01461-w.
no. 8, p. 1729, Jul. 2017, doi: 10.3390/s17081729. [33] James Kuria Kimotho, Walter Sextro, “An approach
[24] J. Pan, Y. Zi, J. Chen, Z. Zhou, and B. Wang, for feature extraction and selection from non-trending
“LiftingNet: A Novel Deep Learning Network With data for machinery prognosis,” Second European
Layerwise Feature Learning From Noisy Mechanical Conference of the Prognostics and Health
Data for Fault Classification,” IEEE Transactions on Management Society 2014, 2014.
Industrial Electronics, vol. 65, no. 6, pp. 4973–4982, [34] W. Caesarendra and T. Tjahjowidodo, “A Review of
Jun. 2018, doi: 10.1109/TIE.2017.2767540. Feature Extraction Methods in Vibration-Based
[25] J. Wang, Z. Mo, H. Zhang, and Q. Miao, “A Deep Condition Monitoring and Its Application for
Learning Method for Bearing Fault Diagnosis Based Degradation Trend Estimation of Low-Speed Slew
on Time-Frequency Image,” IEEE Access, vol. 7, pp. Bearing,” Machines, vol. 5, no. 4, p. 21, Sep. 2017.
42373–42383, 2019, doi: [35] Z. Feng, M. Liang, and F. Chu, “Recent advances in
10.1109/ACCESS.2019.2907131. time–frequency analysis methods for machinery fault
[26] A. Krizhevsky, I. Sutskever, and G. E. Hinton, diagnosis: A review with application examples,”
“ImageNet Classification with Deep Convolutional Mechanical Systems and Signal Processing, vol. 38,
Neural Networks,” in Advances in Neural no. 1, pp. 165–205, Jul. 2013, doi:
Information Processing Systems 25, F. Pereira, C. J. 10.1016/j.ymssp.2013.01.017.
C. Burges, L. Bottou, and K. Q. Weinberger, Eds. [36] Vinod Nair, Geoffrey E. Hinton, “Rectified Linear
Curran Associates, Inc., 2012, pp. 1097–1105. Units Improve Restricted Boltzmann Machines,”
[27] S. S. Roy, S. Dey, and S. Chatterjee, ICML, 2006.
“Autocorrelation Aided Random Forest Classifier- [37] Bridle J.S., “Probabilistic Interpretation of
Based Bearing Fault Detection Framework,” IEEE Feedforward Classification Network Outputs, with
Sensors Journal, vol. 20, no. 18, pp. 10792–10800, Relationships to Statistical Pattern Recognition,”
2020, doi: 10.1109/JSEN.2020.2995109. Soulié F.F., Hérault J. (eds) Neurocomputing. NATO
[28] Christian Lessmeier, James Kuria Kimotho, Detmar ASI Series (Series F: Computer and Systems
Zimmer, Walter Sextro, “Condition Monitoring of Sciences), vol 68. Springer, Berlin, Heidelberg, 1990,
Bearing Damage in Electromechanical Drive Systems doi: https://doi.org/10.1007/978-3-642-76153-9_28.
by Using Motor Current Signals of Electric Motors: [38] A. Zhang, S. Li, Y. Cui, W. Yang, R. Dong and J.Hu,
A Benchmark Data Set for Data-Driven “Limited Data Rolling Bearing Fault Diagnosis With
Classification,” Conference: European Conference of Few-Shot Learning,” in IEEE Access Vol. 7, pp.
the Prognostics and Health Management Society, Jul. 110895-110904, 2019, doi:
2016. 10.1109/ACCESS.2019.2934233
[29] G. Karatzinis, Y. S. Boutalis, and Y. L. Karnavas, [39] C. Lau, W. Heidrich, and R. Mantiuk, “Cluster-based
“Motor Fault Detection and Diagnosis Using Fuzzy color space optimizations,” in 2011 International
Cognitive Networks with Functional Weights,” in Conference on Computer Vision, Nov. 2011, pp.
2018 26th Mediterranean Conference on Control and 1172–1179, doi: 10.1109/ICCV.2011.6126366.
Automation (MED), Jun. 2018, pp. 709–714, doi:
10.1109/MED.2018.8443043.
[30] D. Zhong, W. Guo, and D. He, “An Intelligent Fault
Diagnosis Method based on STFT and Convolutional
Neural Network for Bearings Under Variable
Working Conditions,” in 2019 Prognostics and
System Health Management Conference (PHM-
Qingdao), Oct. 2019, pp. 1–6, doi: 10.1109/PHM-
Qingdao46334.2019.8943026.
[31] V. Pandhare, J. Singh, and J. Lee, “Convolutional
Neural Network Based Rolling-Element Bearing
Fault Diagnosis for Naturally Occurring and
Progressing Defects Using Time-Frequency Domain
Features,” in 2019 Prognostics and System Health
Management Conference (PHM-Paris), May 2019,
pp. 320–326, doi: 10.1109/PHM-Paris.2019.00061.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3056944, IEEE Access

FIRST A. AUTHOR Rishikesh Magar received

his B.Tech degree in mechanical engineering from
Vishwakarma Institute of Technology, Pune, India
in 2017 and the M.S (Research) degree in
mechanical engineering from Carnegie Mellon
University, Pittsburgh, USA, in 2020. He is
currently a first year Ph.D. student in mechanical
engineering at Carnegie Mellon University,
Pittsburgh, USA. His research interest lies in
leveraging machine learning to solve research problems pertaining to
mechanical engineering. Particularly he is interested in graph based neural
networks and has previously developed graph-based models for material
property prediction.

SECOND B. AUTHOR Lalit Ghule received his

B.Tech degree in mechanical engineering from
Vishwakarma Institute of Technology, Pune, India
in 2017 and is currently pursuing M.S (Research)
degree in mechanical engineering from Carnegie
Mellon University, Pittsburgh, USA.
His research interest lies in the area of AI in
manufacturing.

THIRD C. AUTHOR Junhan Li received his B.S.

degree in mechanical engineering from Rensselaer
Polytechnic Institute, New York, USA in 2020 and
is currently pursuing M.S degree in electrical and
computer engineering from Carnegie Mellon
University, Pittsburgh, USA.
His research interest lies in area of signal
processing.

FOURTH D. AUTHOR Yang Zhao received his

M.S. degree in civil and environmental
engineering from Carnegie Mellon University,
Pittsburgh, USA in 2018. He later worked in
Mechanical and AI Lab (MAIL) as a research
assistant from August 2019 to May 2020.
His research interest in lies in area of signal
processing.

FIFTH E. AUTHOR Amir Barati Farimani is

currently an Assistant Professor in the mechanical
engineering department at Carnegie Mellon
University. Dr. Barati Farimani joined the
Department of Mechanical Engineering at
Carnegie Mellon University in the fall of 2018. He
was previously a postdoctoral fellow at Stanford
University. He received his PhD in Mechanical
Engineering in 2015 from University of Illinois at
Urbana-Champaign. His lab at CMU focuses on the problems at the
interface of Mechanical Engineering, data science and machine learning.
His lab uses the state-of-the-art deep learning and machine learning
algorithms and tools to learn, infer and predict the physical phenomena
pertinent to mechanical engineering. Currently, he is teaching artificial
intelligence and machine learning to a large class of graduate students at
CMU. He received the Stanley I. Weiss best thesis award from the
University of Illinois in 2016 and was recognized as an Outstanding
Graduate Student in 2015. During his post-doctoral fellowship at Stanford,
Dr. Barati Farimani has developed data-driven, deep learning techniques
for inferring, modeling, and simulating the physics of transport phenomena
and for materials discovery for energy harvesting applications.

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/