Design of Fault Diagnosis Model for A Motor

Ho-Si-Hung Nguyen1, a), Thi-Kim-Truc Nguyen1, Thi-Hoang-Giang Tran2, Manh-

Tien Luu1, Thanh-Trung Nguyen1, Tuan-Phong Nguyen1, Duc-Anh-Bao Huynh3
1
Faculty of Electrical Engineering, University of Science and Technology, The University of Danang, 54 Nguyen
Luong Bang Street, Lien Chieu Distric, Da Nang city, Viet Nam
2
Faculty of Project Management, University of Science and Technology, The University of Danang, 54 Nguyen
Luong Bang Street, Lien Chieu Distric, Da Nang city, Viet Nam
3
Faculty of Information Technology, University of Science and Technology, The University of Danang, 54 Nguyen
Luong Bang Street, Lien Chieu Distric, Da Nang city, Viet Nam

a)
Corresponding author: [email protected]

Abstract. In modern production systems, the motor plays an important role in supplying energy and driving production processes.
However, things don't always go smoothly and engine failures are one of the common challenges encountered in the production
process. Misalignment of drive shaft is a common engine fault resulting from improper installation or damage of the machine
components. Vibration analysis is an effective method to diagnose motor faults because the high frequency of vibration will limit
the normal noise. This study designs a data collection model of motor's vibration signals and presents a novel approach to classify
motor faults (misalignment of drive shaft) based on the vibration signals. The motor faults is classified by a proposed deep
Convolutional Neural Network (CNN) after the collected signal is denosied by using wavelet transform. Experimental results show
that classifiers based on CNN easier and more stable accuracy rates than those based on time domain.

Keywords: Multiscale entropy, Wavelet transform, Induction motor, Fault detection

In recent years, with the rapid development of

1. Introduction
IoT technology in smart factories, the amount of
data collected is increasing rapidly and in
Electric motors are an important component addition, the breakthrough development of
in industrial systems. While most motors operate artificial intelligence (AI) methods, fault
stably and reliably, they may encounter diagnosis for motor based on AI methods is
unforeseen issues in harsh or demanding work getting more attention from academia. AI
environments. Different motor faults have been methods can be classified in to three categories
stu-died in literature, including coil winding being statistical, machine learning and deep
faults, rotor and stator imbalances, broken rotors, learning. Statistical analysis methods include
air gap eccentricity, and bearing faults [1-6]. Principal Component Analysis (PCA) and Partial
These faults can decrease the motor's Least Squares (PLS) [22]. Common machine
performance, leading to inefficient energy learning methods used in motor diagnosis are
consumption and potential breakdowns. principal component analysis (PCA) [23],
Therefore, it is necessary to regularly or even Hilbert–Huang Transform and Sup-port Vector
continuously monitor the motor's operational Machine (SVM) [24], Hilbert Huang Transform
condition to ensure efficiency and productivity. with k-Nearest Neighbors (k-NN) [25], Singular
Various measurements from different sensors are Value Decomposition (SVD) [26]. Deep learning
used for online motor condition monitoring, and methods, Convolutional Neural Networks
the collected data is analyzed to identify potential (CNNs) is an effective method in practical
motor faults [7-8]. Common measurements for applications. The use of CNNs in detecting motor
motor fault diagnoisis include voltage and current faults has been in development in recent years.
of the stator [9-10], air gap eccentricity and [27] used generator torque data for CNN fault
external flux density [11], rotor position and detection in wind turbines based on benchmark
speed [12], output torque, internal and external model, while [28] used a CNN with an adaptive
temperature [13-16], sound [17-18], and learning rate for fault diagnosis with the CWRU
vibration [19-21]. dataset. Other authors have used their own
datasets for fault detection and classi-fication pro-ximity sensor is placed close to the motor and
purposes; used an adaptive 1­D CNN [29] on functions as a keyphasor.
motor current data and [30] on vibration data. To create dynamic imbalance, a heavy disc is
This study focuses on detecting and attached to the shaft between the bearings. This
classifying motor faults in shaft misalignments. heavy disc has a diameter of 75 mm and a
The causes of misalignment can include incorrect thickness of 25 mm.
installation, wear, and damage to components due
to machine vibrations. Shaft misalignments can
lead to damaged shaft bearings, housing,
bending, or wearing of the crankshaft, and worn
couplings. Therefore, maintaining precise
alignment is crucial to ensure safe and reliable
operation. Specifically, this study proposes a
method to detect motor shaft misalignment by
measuring the motor's vibrations. The proposed
method calculates the difference between the
original vibration signal and the signal after Figure 1: The experimental model.
denoising using wavelet transform. Experimental (A) Motor; (B) Shaft; (C) Bearings; (D) Heavy
results show that this difference can effectively disc; (E) Keyphasor; (F) Inductive proximity
dis-tinguish between aligned motors and motors sensor; (H) electrical cabinet.
with misalignments.
2.2. Data Collection Circuit
The rest of this paper is organized as follows.
Section 2 describes the data collection model and 2.2.1. Systerm Block Diagram
the experimental setup. Section 3 presents the
proposed method for detecting motor shaft
misalignment. Section 4 describes vibration data
collected by sensors in time domain. After that,
analysis results and evaluation present in section
5. Finally, Section 6 concludes this paper.
2. Experimental

2.1. Experimental model Figure 2: System Block Diagram

Figure 1 depicts the model of the Components in the system:
experimental model used for this study. The Arduino Nano: Receives control signals from
experimental model is relatively simple and the control panel and from Arduino Mega to
includes 1 DC motor, 1 rotating shaft, 2 bearings, control the motor speed.
1 coupling, 1 heavy disc2 mounted on the shaft, Analog Proximity Sensor: This sensor is used
oscillation measurement sensors, 1 keyphasor, to measure the oscillation of the shaft. It
and a control and signal acquisition system. generates an analog signal, usually a voltage
In this experimental model, the rotating shaft range, to represent the measured value. This
can reach a maximum speed of 12,000 signal is sent to the Arduino for processing
revolutions per minute. The acceleration of the through a converter.
shaft can be easily adjusted depending on the Keyphasor (Digital Proximity Sensor): This is
specific requirements of each experiment. The another type of proximity sensor, but the signal
rotating shaft, with a diameter of 10 mm, is obtained from this sensor is a digital pulse.
connected to the motor via flexible coupling and Keyphasors are often used to determine the phase
is supported by two bearings. The distance angle of a rotating device, and they are typically
between these bearings can be easily adjusted used as a reference or base signal with a phase
using two rails placed underneath. To measure angle of 0º for the system.
the oscillation of the shaft, an inductive proximity Arduino Mega: This is a control board that
sensor is installed vertically. In this laboratory receives signals from Arduino Nano, the analog
setup, we use the inductive proximity sensor to proximity sensor, and the keyphasor. It then
assess the shaft's oscillation. Another inductive processes these signals to send to Blynk.
 NodeMCU: NodeMCU allows Wi-Fi In this study, we propose a deep Convolutional
connectivity and communication via the TCP/IP Neural Network (CNN) with multiple
protocol, enabling the transmission and reception convolutional layers to optimize the image
of data over the Internet. This facilitates remote classification task as illustrated Figure 5.
monitoring and control of the system.
Blynk is a platform with smart applications
that allow you to easily interact with
microcontrollers over the Internet. The Blynk
App creates an interface for remote control and
monitoring of systems on any device.
Additionally, it stores collected data for signal
analysis purposes. Figure 5: The proposed CNN Network Model
The CNN model consists of convolutional layers,
pooling layers, and fully connected layers. In this
CNN architecture, we use three convolutional
layers.
The convolutional layer is a crucial component in
the CNN model, as it helps to learn local features
from the input image. Each convolutional layer
applies a set of filters to perform convolutional
operations on the input image. The result of the
Figure 3: Real-life image of the signal convolutional operation is a feature map, where
collection circuit local features are encoded.
2.2.2. Signal Collection from the Data Collec- In this CNN model, we use three convolutional
tion Model. layers with the following parameters:
The proximity sensor is placed 20cm away  Convolutional Layer 1: 16 filters, filter
from the bearing, and the Keyphasor is positioned size 3x3, ReLU activation function.
near the small groove on the disk, close to the  Convolutional Layer 2: 32 filters, filter
motor's location. In the normal operating state, no size 3x3, ReLU activation function.
heavy objects are placed on the heavy disk.  Convolutional Layer 3: 16 filters, filter
However, to induce shaft misalignment in the size 3x3, ReLU activation function.
experiments, a heavy object will be placed at After each convolutional layer, we use a pooling
positions 1 to 8 as shown in Figure 4. layer to reduce the size of the feature map. In this
report, we employ the max pooling layer with a
pooling window size of 2x2 and a stride of 2.
Finally, after the convolutional and pooling
layers, we use a fully connected layer to perform
the classification task.
 Flatten Layer: This layer is used to flatten
the input of the model. It is an essential step
to convert data from a multidimensional
Figure 4: Positions of the Misalignment Masses format (e.g., images) into a one-dimensional
on the Rotating Disk vector to feed into the subsequent Dense
The signals from the four sensors and the layers.
keyphasor will be collected in each experiment.  Dense Layer 1: This is a fully connected
 Experiment 1: No additional weight layer with 1024 units (neurons) and a ReLU
placed on the heavy disk corresponding to the (Rectified Linear Unit) activation function.
normal operating condition. This layer performs linear computation on the
 Experiment 2: A mass with a weight of input and applies the ReLU function to
m=14g placed at positions 1 to 8 on the heavy generate a non-linear output.
disk.  Dense Layer 2: This layer is also a fully
connected layer with 128 units and a ReLU
2.3. The proposed deep Convolutional Neural
activation function. It continues the linear
Network.
 computation on the input from the previous the oscillation frequency is also higher compared
layer and applies the ReLU function. to the normal state.
 Dense Layer 3: This is the final layer of 250
the model, consisting of 9 units and a softmax
activation function. This layer performs the 200
classification of the output into nine different
classes, using the softmax function to 150
calculate the probability of classification for
100
each class.
The fully connected layers receive input from the 50
flattened feature map and apply the softmax
activation function to calculate the class 0

0
15
30
45
60
75
90
105
120
135
150
165
180
195
210
probability for the output classes.
The model is built based on the feedforward
neural network architecture, where the neuron
units in the Dense layers are fully connected to
each other. The ReLU and softmax activation
functions are used to generate non-linear output
Figure 7: Oscillation graph of the motor shaft
and classify the probabilities for the output
when a heavy mass with a weight of m=14g is
classes, respectively.
placed on the heavy disk at positions 1 to 8.
2.4. Vibration Data Collection Results
2.5. Analysis Results and Evaluation
2.4.1. Case without any additional weight
2.5.1. Data Preprocessing
placed on the heavy disk.
The data collected through sensors using the
The case without any additional weight placed on
proposed measurement method is divided into the
the heavy disk can be considered equivalent to the
following cases: no heavy object on the hard disk;
normal operation of the motor.
heavy objects placed at each deviation point (8
deviation points) with masses of m=14g. Details
about the data collection process are as follows:
At each deviation point, measurements are taken
using a specialized measuring device, which
returns 100,000 analog signals (ranging from 50
to 210) with a frequency of 3ms per signal. After
passing through the noise reduction step, the
Figure 6: Oscillation graph of the motor shaft
signals are converted into graphs for each faulty
when no additional weight is placed on the
and non-faulty point of the motor. These graphs
heavy disk.
are then segmented into portions with an analog
Clearly, even when the motor operates under
range of 192ms. Finally, from the segmented
normal conditions, there is still a small imbalance
graph frames, we transform them into 2D images.
oscillation on the motor shaft. There are several
The result is a dataset consisting of 24,000
reasons that lead to this imbalance, such as
images of faulty points (3,000 images per faulty
uneven motor shaft design or uncertain motor
point, with 1,000 fault images corresponding to
bearing during rotation, resulting in increased
each heavy object mass m), and 600 images of
imbalance oscillation.
non-faulty points. In total, there are 24,600
2.4.2 Case with a heavy mass of m=14g placed images.
on the heavy disk at positions 1 to 8. In this proposed model, the dataset is divided into
In this case, a heavy mass with a weight of m=14g three parts to facilitate the training and validation
will be placed at positions 1 to 8. The total processes. The percentage of data for training is
number of samples is 350,000, and the collected 70%; the percentage for validation is 20%, and
graph results will be compared with the case the percentage for testing is 10%.
when no heavy disk is placed.
Clearly, when the additional weight is placed, the
oscillation amplitude increases significantly, and
 Figure 8: The process of converting analog data into 2D image format
The process of converting vibration signals
3. Results and discussion
into 2D images is applied to all samples in the
After training and evaluating the model on the
dataset. Figure 4.6 displays sample 2D images in
train and valid datasets with 20 epochs, Figure 10
the constructed dataset for each type of fault at
illustrates the results of the model's loss and
different positions of the heavy object with
accuracy. The results obtained are as follows:
varying masses. The labels are numbered from 0
 Loss value on the Train dataset: 0.12
to 8, corresponding to the labels in Table 1:
Table 1: Legend for Label Types Used in the  Loss value on the Valid dataset: 0.27
Model.  Accuracy on the Train dataset: 0.95
Lable Explanation  Accuracy on the Valid dataset: 0.92
0 Fault at unbalance position 1
1 Fault at unbalance position 2
2 Fault at unbalance position 3
3 Fault at unbalance position 4
4 Fault at unbalance position 5
5 Fault at unbalance position 6
6 Fault at unbalance position 7
7 Fault at unbalance position 8
8 Normal

Figure 10: Graph of Loss and Accuracy

values of the model on the Train and Valid
datasets.
From the graphs, the lines in the loss graph tend to
gradually decrease towards 0 and converge with
each other on both the train (blue line) and increase the accuracy by up to 20% compared to
validation (yellow line) datasets, indicating that when the data is not preprocessed.
the model has a good learning capability and does
4. Conclusion
not encounter overfitting issues.
To assess the model's accuracy, we performed
In this study, the authors presented the steps to
classification on the test dataset. The resulting
build a data collection model for diagnosing
confusion matrix is shown in Figure 11, with the
motor faults. Once the model was constructed,
following results:
data was collected under two conditions: (i)
 Accuracy (Overall Accuracy): 92% Without any additional weight placed on the
 Precision (Accuracy for positive heavy disk, corresponding to normal motor
classes): 98% operation, and (ii) with a heavy mass of m=14g
 Recall (Coverage for positive class placed at different positions on the heavy disk.
predictions): 99% The vibration data collected by sensors are
 F1-Score (Harmonic mean of precision converted into graphs for each faulty and
and recall): 99% nonfaulty point of the motor. These graphs are
The confusion matrix represents the number of then segmented into portions with an analog
predicted labels by the model compared to the true range of 192ms. Finally, from the segmented
labels. We observed that the model's accuracy rate graph frames, they are transformed into 2D
on the test dataset is at a good level of 92%. The images. After that, all the data was trained by a
most common misclassification occurs when the CNN (Convolutional Neural Network) model.
model misclassifies label 3, with 42 instances of The corresponding unbalance faults for different
misclassifying label 3 as label 2. po-sitions were labeled to identify motor faults
for positions 1 to 8. Experimental results show
that classifiers based on CNN get the high
accuracy.
The study has limitations due to the uneven
design of the motor shaft and the uncertain motor
bearing, resulting in increased unbalance
oscillations when the motor operates.
Acknowledgements

In this section you should acknowledge

sponsors (funding agency and grant number) and
those who have helped with your research.

Figure 11: Graph of Loss and Accuracy values of

References
the model on the Train and Valid datasets [1] H. Divdel, M.H. Moghaddam, G. Alipour, A new
From the obtained results, we can see that the diagnosis of severity broken rôtor bar fault -based modeling
model's training accuracy is above 90%. and image processing system. J.Curr.Res. Sci. 1, 771 (2016)
Although the model's accuracy is affected by the [2] C.-C. Kuo et al., Implementation of a motor diagnosis
weight (due to significant vibrations in the system for rôtor failure using genetic algorithm and fuzzy
classification. Appl. Sci. 7(1), 31 (2016)
experimental setup as the weight of the object [3] M.A. Sheikh et al., An intelligent automated method to
increases), the model can still learn the features diagnose and segregate induction motor faults. J. Electrical
quite well if the training data is large and Syst. 13(2), 241–254 (2017)
balanced. If the training data is large and [4] T.A. Garcia-Calva et al., Time-frequency analysis based
on minimum-norm spectral estimation to detect induction
balanced, the model can still learn the features motor faults. Energies. 13(16), 4102 (2020)
quite well, even though the weight affects its [5] Sheikh, Muhammad Arrian, et al. "An unsupervised
accuracy due to significant vibrations in the automated method to diagnose industrial motors faults."
experimental setup as the weight of the object 2018 IEEE International Conference on Electrical Systems
increases.The process of noise reduction and data for Aircraft, Railway, Ship Propulsion and Road Vehicles &
International Transportation Electrification Conference
normalization in the preprocessing stage plays a (ESARS-ITEC). IEEE, 2018.
crucial role in training the model, as it can [6] Alsaedi, Malik Abadulrazzaq. "Fault diagnosis of three-
phase induction motor: A review." Optics 4.1 (2015): 1-8.
[7] Siddiqui, K. M., Sahay, K., & Giri, V. K. (2014). Health [25] Pandya, D. H., Upadhyay, S. H., & Harsha, S. P.
monitoring and fault diagnosis in induction motor-a review. (2013). Fault diagnosis of rolling element bearing with
International Journal of Advanced Research in Electrical, intrinsic mode function of acoustic emission data using
Electronics and Instrumentation Engineering, 3(1), 6549- APF-KNN. Expert Systems with Applications, 40(10),
6565. 4137-4145.
[8] Wu, G., Yan, T., Yang, G., Chai, H., & Cao, C. (2022). [26] Shen, F., Chen, C., Yan, R., & Gao, R. X. (2015,
A review on rolling bearing fault signal detection methods October). Bearing fault diagnosis based on SVD feature
based on different sensors. Sensors, 22(21), 8330. extraction and transfer learning classification. In 2015
[9] Laadjal, K., Antunes, H. R., Sahraoui, M., Bento, F., & prognostics and system health management conference
Cardoso, A. J. M. (2022). On-line Diagnosis and (PHM) (pp. 1-6). IEEE.
Discrimination of Stator Faults in Six-Phase Induction [27] Rahimilarki, R., Gao, Z., Jin, N., & Zhang, A. (2019,
Motors, Based on Voltage Symmetrical Components. IEEE July). Time-series deep learning fault detection with the
Transactions on Transportation Electrification. application of wind turbine benchmark. In 2019 IEEE 17th
[10] Pietrzak, P., & Wolkiewicz, M. (2023). International Conference on Industrial Informatics (INDIN)
Demagnetization Fault Diagnosis of Permanent Magnet (Vol. 1, pp. 1337-1342). IEEE.
Synchronous Motors Based on Stator Current Signal [28] Guo, X., Chen, L., & Shen, C. (2016). Hierarchical
Processing and Machine Learning Algorithms. Sensors, adaptive deep convolution neural network and its application
23(4), 1757. to bearing fault diagnosis. Measurement, 93, 490-502.
[11] Thollon, F., Grellet, G., & Jammal, A. (1993). [29] Ince, T., Kiranyaz, S., Eren, L., Askar, M., & Gabbouj,
Asynchronous motor cage fault detection through M. (2016). Real-time motor fault detection by 1-D
electromagnetic torque measurement. European convolutional neural networks. IEEE Transactions on
Transactions on Electrical Power, 3(5), 375-378. Industrial Electronics, 63(11), 7067-7075.
[12] Piotrowski, J. (2006). Shaft alignment handbook. Crc [30] Lee, J. H., Pack, J. H., & Lee, I. S. (2019). Fault
Press diagnosis of induction motor using convolutional neural
[13] Tavner, P. (2008). Condition monitoring of rotating network. Applied Sciences, 9(15), 2950.
electrical machines (Vol. 56). IET.
[14] Glowacz, Adam. "Thermographic fault diagnosis of
ventilation in BLDC motors." Sensors 21.21 (2021): 7245.
[15] Glowacz, Adam. "Ventilation diagnosis of angle
grinder using thermal imaging." Sensors 21.8 (2021): 2853.
[16] Bonnett, Austin H. "Analysis of winding failures in
three-phase squirrel cage induction motors." IEEE
transactions on industry applications 3 (1978): 223-226.
[17] Delgado-Arredondo, Paulo Antonio, et al.
"Methodology for fault detection in induction motors via
sound and vibration signals." Mechanical Systems and
Signal Processing 83 (2017): 568-589.
[18] R. Lara, R. Jimenez-Romero, F. Perez-Hidalgo and
MD. Redel-Macias, Influence of constructive parameters
and power signals on sound quality and airborne noise
radiated by inverter-fed induction motors, Measurement, 73,
pp. 503-514, 2015
[19] Qian, W.; Li, S. A novel class imbalance-robust
network for bearing fault diagnosis utilizing raw vibration
signals. Meas. J. Int. Meas. Confed. 2020, 156,
[20] Ye, Z.; Yu, J. Multi-scale Weighted Morphological
Network-based Feature Learning of Vibration Signals for
Machinery Fault Diagnosis. IEEE ASME Trans. Mechatron.
2021.
[21] Fan, H.; Xue, C.; Zhang, X.; Cao, X.; Gao, S.; Shao, S.
Vibration Images-Driven Fault Diagnosis Based on CNN
and Transfer Learning of Rolling Bearing under Strong
Noise. Shock Vib. 2021, 2021, 1–16
[22] Yin, S., Ding, S. X., Xie, X., & Luo, H. (2014). A
review on basic data-driven approaches for industrial
process monitoring. IEEE Transactions on Industrial
electronics, 61(11), 6418-6428.
[23] Malhi, A., & Gao, R. X. (2004). PCA-based feature
selection scheme for machine defect classification. IEEE
transactions on instrumentation and measurement, 53(6),
1517-1525.
[24] Wang, Y. S., Ma, Q. H., Zhu, Q., Liu, X. T., & Zhao, L.
H. (2014). An intelligent approach for engine fault diagnosis
based on Hilbert–Huang transform and support vector
machine. Applied acoustics, 75, 1-9.