Using Artificial Intelligence (AI) for Monitoring and

Diagnosing Electric Motor Faults Based on Vibration

Signals
Van – Nam Pham* Quang - Huy Do Ba Duc-Anh Tran Le Quang - Minh Nguyen
Faculty of Engineering Faculty of Engineering Electrical School of Electrical & Electronic School of Information and
Electrical & Automation & Automation Engineering Communications Technology
Hanoi University of Industry Hanoi University of Industry Hanoi University of Science and Hanoi University of Science and
Hanoi, Vietnam Hanoi, Vietnam Technology Technology
nampv@haui.edu.vn dohuy9379@gmail.com Hanoi, Vietnam Hanoi, Vietnam
ducanhtran11082002@gmail.com qminhnguyen2211@gmail.com
Thanh - Lam Bui Alberto Ernesto Coboi
Foreign Language Specialized Faculty of Engineering Electrical
School & Automation
Vietnam National University Hanoi University of Industry
Hanoi, Vietnam Hanoi, Vietnam
thanhlam6900@gmail.com AlbertoCoboi@haui.edu.vn

Abstract— Detecting bearing faults in electric motors is highly it is imperative and practical to conduct research and develop
crucial for improving production efficiency and reducing systems for examining, monitoring, and predicting electric
accidents in complex mechanical systems, which poses significant motor faults using IoT (Internet of Things technology)[18]
challenges for current fault diagnosis technology. This paper within factory environments. Furthermore, data encompassing
investigates and applies Artificial Intelligence (AI) to enhance the electrical, mechanical, and thermal parameters, among others,
monitoring and diagnosis process of electric motor faults based on will be gathered and integrated with intelligent software for
vibration signals. The research aims to construct a model for diagnostic purposes, leveraging artificial intelligence (AI), a
collecting sample data from motors with three common types of hallmark of the fourth industrial revolution
bearing faults and utilizes the Resnet-50 network to assess the
accuracy of monitoring and diagnosing faults. The study conducts The general operating principle of these systems includes the
a vibration signal analysis to identify potential indicators of faults following steps:
in electric motors. The survey results presented in the paper
demonstrate the accuracy of using the Resnet-50 network in
monitoring and diagnosing electric motor faults. The paper also
provides essential insights into the performance of AI networks
and their practical applicability in the field of industrial
equipment maintenance and management.

Keywords— Bearing fault diagnosis, industrial equipment

maintenance, Resnet.
Fig. 1. General operating principle of the electric motor fault alert system.
I. INTRODUCTION
In most modern factory settings, the majority of engine • Collect data from sensors: Gather information from
monitoring and alert systems are installed separately, focusing various sensors, including temperature, vibration,
on only a limited set of electrical parameters. They often lack pressure, and current sensors affixed to electrical
substantial data collection for in-depth analysis. Regular online machinery. These sensors continuously collect data on
checks and continuous monitoring of engine parameter analyses the machines' operational conditions.
are infrequently conducted, and there is a deficiency in
automated engine maintenance planning capabilities. This • Transmit data to the central control system: Transmit
deficiency results in the inability to detect unforeseen the data from these sensors to the central control system
breakdowns and engine malfunctions, which, when not through network connections, such as the internet or
adequately monitored, can lead to significant machinery local area networks. The data is directed to servers or
damage, necessitating costly repairs or even engine central hubs for analysis and processing.
replacements. Conversely, when conditions are continuously • Analyze data: Within the central control system, apply
monitored, potential issues can be identified early, leading to machine learning algorithms, artificial intelligence, and
simpler, quicker, and more cost-effective repairs. Consequently, data analysis techniques to scrutinize the sensor data.

979-8-3503-3094-6/24/$31.00 ©2024 IEEE

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE 723 ICOIN 2024
 The primary objective is to identify trends, patterns, and In recent years, there has been a growing focus on utilizing
deviations that may signify potential incidents or faults. deep learning (DL) methods [7-11] to address the mentioned
issues, with many DL approaches applied to bearing fault
• Predict issues and generate alerts: Based on data diagnosis. While Convolutional Neural Networks (CNN) are
analysis, the system predicts potential incidents or classic DL structures for image classification [7,8,19], various
undesirable situations in the future. When a potential DL models [9,10,11,20] have found wide applications in fault
incident is identified, the system generates alerts and detection tasks. However, deeper DL models often face the
notifies managers or technicians for timely intervention. problem of gradient vanishing, requiring performance sacrifices
• Optimize maintenance: Provide in-depth insights during training. To overcome these challenges, Residual
regarding anticipated issues and the machinery's current Network (ResNet) was introduced [12]. ResNet employs
status. This information aids in fine-tuning maintenance residual connections, enabling the learning of residual functions
schedules and practices, resulting in time and resource from input rather than complex mappings from input to output.
savings. This innovation has significantly improved the performance of
deep neural networks, leading to the introduction of various
• Monitor and provide feedback: Following the issuance ResNet-based models [13,14]. In this research paper, we
of alerts, the system continues to monitor the propose the use of ResNet-50 to train with a dataset containing
machinery's condition and assess the outcomes of three common bearing fault types constructed from our
interventions. This ongoing evaluation helps gauge the experimental model.
solution's effectiveness and make necessary
adjustments. III. METHODOLOGY
The remaining part of this article is organized as follows: A. Data Collection
Section 2 discusses prior related works on this issue, presenting Conducting real-time, automated diagnostics on engines
several relevant studies to support this research. Section 3 operating in practical conditions presents significant limitations
explains experimental data collection and the development of
due to the complexity of the problem. Vibration signals and
anomaly detection algorithms. Section 4 presents the main
machinery noise during production often exhibit high levels of
results of algorithm development and a performance comparison
study. Finally, Section 5 concludes the article and suggests noise and variability influenced by environmental conditions.
feasible directions for future research. Careful selection of appropriate feature values for signal
recognition can enhance the efficiency of the recognition
II. RELATED WORK model.
Recent literature surveys in [2-4] have highlighted the
challenges and opportunities for developing robust predictive
maintenance techniques based on machine learning, particularly
for rotating equipment such as bearings, motors, gearboxes, and
pumps. Many challenges and opportunities still await
exploration in this field to enhance the accuracy of machine Fig. 2. The general model of the engine fault detection system
learning models and increase the flexibility of proposed In this paper, we have developed a motor model to address
predictive maintenance methods in the future. the problem of identification and training a deep learning
Tuan A. Z. Rahman et al. [5] proposed an intelligent network. Here is a basic description of this model:
anomaly detection method for electric motors based on vibration • Mechanical Component: A 3-phase, 2HP motor with 2
signals combined with AI algorithms. They developed an bearings and an electromagnetic brake assembly to
unsupervised learning model for two different types of motors simulate load and torque effects.
within the same category: a new experimental motor and an old
industrial motor. The model's performance in anomaly detection • Control Component: A 2.5KW inverter controls the
for both types of motors was extensively studied, and the results motor's rotational speed, ranging from 1500 to 1750
showed that it had the highest anomaly detection capability for RPM.
standardized motor conditions using mapped features. However,
• Vibration Measurement Equipment: Vibration data is
they currently utilized only data from normal motor conditions
collected, stored on an SD card, and transmitted using
due to a lack of information about fault conditions. WIFI communication standards for AI model training,
M. Masood Tahir et al. [6] presented a solution using with sampling frequencies of 6 KHz and 12 KHz.
vibration signal features like RMS, Mean, Variance, skewness,
The identification states are divided into four categories:
kurtosis, median, range, etc., for model training, similar to many
other studies. During the data preprocessing phase, this paper • N (normal state): The motor operates normally without
introduced an approach called Median-based Outlier Detection any damage.
(MOD) to detect outliers (data samples in which features are
affected by external factors, not due to faults, and exclude them
from the training process to improve model performance).
However, this paper did not address the classification of similar
fault types with different fault sizes.

724
 • MI (Misalignment damaged): Misalignment occurs
when the shaft is not aligned properly or not in the correct
plane. Misalignment can generate uneven torque and
cause vibration.
• IR (inner raceway damage): The bearing's inner raceway
is damaged.
• OR (outer raceway damage): The bearing's outer
raceway is damaged. (a)-N (b)-MI

The shaft bearings had inner and outer raceway damage at a

2 mm diameter point. Vibration signals were recorded using an
accelerometer at the specified location, as shown in Figure 3.
The signals were recorded for approximately 2 minutes at a 12
kHz sampling frequency.
Vibration
measuring
(c)-IR (d)-OR
Motor shaft
device

Fig. 4. Representation of the collected data sample set in the time domain (s).
Electrom
Motor agnetic
brake
B. ResNet-50
ResNet-50 is a widely recognized and influential deep
Driveshaft Support
learning architecture, particularly in the field of computer vision.
Bearings It belongs to the family of ResNet (Residual Neural Network)
models, which are designed to address the challenge of training
very deep neural networks by introducing residual learning.
Conventional neural networks encounter challenges when it
comes to training extremely deep structures because of the
vanishing gradient issue and performance degradation with
increasing network depth. The vanishing gradient problem
hinders effective learning, particularly in the initial layers of the
Fig. 3. Test model, device placement, and bearing faults location. network [15].

The measuring device is attached to the motor casing. All ResNet was introduced by Kaiming He et al. in their paper
data files are in MATLAB Data (*.mat) format, collected at rates "Deep Residual Learning for Image Recognition" in 2015 [15].
of 6000 and 12000 samples per second. Speed and torque data The key innovation was the introduction of residual blocks,
are obtained using a torque encoder and recorded manually by a which enable the training of extremely deep networks by
incorporating shortcut connections or skip connections. They
motor controller.
have introduced the concept of residual learning, which is
The collected sample dataset, detailed in the table below: applied to multiple layers within the ResNet framework.
Residual blocks in ResNet are effective when the input and
TABLE I. DATA COLLECTED FROM THE EXPERIMENTAL output data have the same dimensions. Moreover, ResNet blocks
MODEL come in two varieties, with two layers for ResNet-18 and
Type of error Sampling Number of Time per sample ResNet-34 networks and three layers for ResNet-50 and ResNet-
frequency (Hz) samples (minutes) 101 networks. The initial two layers of the ResNet architecture
N 12000 4 2 are reminiscent of GoogleNet, involving a 7x7 convolution
MI 12000 4 2 operation and 3x3 max-pooling with a stride of 227 [16].
IR 12000 4 2 ResNet-50 utilizes a bottleneck design in each of its residual
OR 12000 4 2 blocks. This design comprises three convolutional layers with
kernel sizes of 1x1, 3x3, and 1x1, which serves to reduce
computational complexity while maintaining representational
The following is an image of the collected data set after capacity. The name 'ResNet-50' stems from the fact that this
being converted to time domain(s). architecture consists of 50 layers. It's a deep structure that
incorporates multiple residual blocks, each with varying filter
counts. A residual block is composed of two 3x3 convolutional
layers and a shortcut connection. This shortcut connection
bypasses one or more layers and directly combines the input
with the output, forming what is referred to as the residual. A
visual representation of the ResNet-50 architecture is provided
in Figure 5.

725
 𝑛𝑛

Υ(𝑘𝑘) = ∑ 𝑋𝑋(𝑗𝑗)𝑊𝑊𝑛𝑛 (𝑗𝑗−1)(𝑘𝑘−1) (1)

𝑗𝑗=1

Where 𝑊𝑊𝑛𝑛 = 𝑒𝑒 (−2𝜋𝜋𝜋𝜋)/𝑛𝑛 Fourier transform is represented as:

+∞
Υ(𝜔𝜔) = ∫ 𝑥𝑥(𝑡𝑡)𝑒𝑒 −𝑖𝑖𝑖𝑖𝑖𝑖 𝑑𝑑𝑑𝑑 (2)
−∞

With 𝑡𝑡/𝑇𝑇 = (𝑗𝑗 − 1)/𝑛𝑛 and 𝜔𝜔𝜔𝜔 = 2𝜋𝜋(𝑗𝑗 − 1)(𝑘𝑘 − 1)/𝑛𝑛, we

can represent the sampled signal data using the sinc function:
𝑛𝑛

𝑥𝑥(𝑡𝑡)𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐(𝑡𝑡) = ∑ 𝑋𝑋(𝑗𝑗)𝛿𝛿(𝑡𝑡 − (𝑗𝑗 − 1)∆𝑡𝑡) (3)

𝑗𝑗=1

Accordingly, four frequency spectra are plotted, each

corresponding to one of the three fault charts and one for the
normal chart as shown in Figure 6. The MATLAB script reads
the raw data files, which are then divided into arrays of 256
samples for the signal image transformation. The final step
involves converting this signal data into spectral images,
Fig. 5. Resnet-50 architecture. highlighting key features from the original data.
A neural network gains knowledge by utilizing
backpropagation. In the ResNet50 model, the upper layers
remain adaptable, allowing them to learn through
backpropagation, while the lower layers are kept fixed. The
process of modifying the weights during backpropagation in
these upper layers is known as fine-tuning. Fine-tuning the
upper layers of ResNet50 is necessary because there's no
assurance that their statistical properties, like mean and variance,
will align with those of our specific dataset [17].
IV. EXPERIMENT AND RESULTS
A. Experimental Dataset
This research validates the proposed approach using self-
generated experimental data. The data acquisition system
includes an accelerometer, a measurement module, a chassis,
and LabVIEW software, creating a fully automated data
collection system. The software handles vibration signal
presentation, analysis, and gathering. The dataset includes both Fig. 6. Analyze the raw data using Short-time Fourier Transform (STFT).
typical and defective bearings with internal ring defects, external
ring defects, and shaft misalignment in three failure scenarios. After the conversion of raw data into spectral domain
An accelerometer positioned above the bearing records vibration images, we proceeded with the individual classification and
signals at a 12 kHz sampling frequency. Detecting machine labeling of error image files. This process resulted in an image
vibration serves as an early warning system for unfavorable dataset comprising a total of 11.9k images per faulty class. These
bearing conditions, particularly critical for high-power electric images were then distributed in a ratio of 70% for training, 20%
motors that require a warm-up period. for validation, and 10% for testing, for each class.
B. Identify the Headings C. Training datasets to each Model
Using the self-created dataset, we developed a MATLAB The division of the training dataset into four distinct classes
program to convert raw data into frequency spectra. Initially, we is of paramount importance. Three of these classes are dedicated
constructed a time vector 't' ranging from 0 to 1 second, with to faulty bearings (IR, OR, MI), while one class represents
intervals of 1/Fs, aligned with the sampling frequency (Fs) of normal bearings. It is imperative that each class possesses a
12000 Hz. Subsequently, we utilized the Fast Fourier Transform diverse dataset, and the inclusion of 11.9k images per faulty
(FFT), an algorithm derived from the combination of discrete class is essential. This ensures that the model comprehensively
and continuous Fourier transforms, to analyze the signals and learns variations associated with different faults, necessitating
obtain their corresponding frequencies. The discrete Fourier meticulous data collection.
transform (DFT) can be described as follows:

726
 Similarly, the 'Normal' class requires a sufficient number of
images for effective discrimination between normal and faulty
bearings.
The inclusion of 11.9k images for this class ensures that the
model develops a robust understanding of the typical
characteristics exhibited by normal bearings. The creation of
these four classes, each with a substantial and diverse dataset,
serves as a foundational element for the model's future
performance in bearing classification and prediction tasks.
D. Results of fault diagnosis and comparative analysis post-
training for each model
Table II shows the exact percentage loss of the ResNet-50
model after training.

TABLE II. SUMMARY OF RESNET-50

The Accuracy of Training Model

Fig. 7. Resnet50 Confusion matrix.
Model Test
Val_accuracy Val_loss Test loss
Accuracy In conclusion, the ResNet50 model demonstrates optimal
Resnet-50 96.527% 93.381% 0.1953 0.3530
capabilities in image classification tasks. It exhibits high test
accuracy and low validation and test loss values, indicating its
ability to generalize well and make accurate predictions on
The provided table offers a comprehensive evaluation of test unseen data. These results suggest that the model is well-trained
and validation performance metrics [21] for various computer and can be considered reliable for various image recognition
vision. The training and evaluation outcomes of the ResNet50 applications.
model are reported in the table. The model exhibits strong
performance in various aspects, indicating its optimal E. Evaluate the model's accuracy once more using a different
capabilities. dataset:
Test Accuracy: The model achieves an impressive test The chart below describes the number of misclassified
accuracy of 96.527%. This metric reflects the model's ability to images in a test set of 2000 images using the ResNet50 model,
correctly classify images in an unseen dataset, which is a crucial showcasing its performance in different categories:
measure of its generalization capacity.
Validation Accuracy: The validation accuracy, at 93.381%,
is another positive indicator of the model's performance. This
metric is crucial during training as it helps monitor the model's
performance on a separate dataset, making it an essential tool for
preventing overfitting.
Validation Loss: The low validation loss value of 0.1953
suggests that the model generalizes well during training. Lower
validation loss values indicate that the model is not overfitting
and is effectively learning from the training data.
Test Loss: The test loss is 0.3530, which is slightly higher
than the validation loss. This difference can be expected, as the
test loss measures the model's performance on a completely
unseen dataset. The proximity of the test loss to the validation
loss suggests that the model maintains its performance on new,
unseen data.
Following the training of the dataset using Resnet50, the Fig. 8. The chart displays the count of misclassified images.
subsequent step involved the utilization of the test dataset
The ResNet50 model was evaluated on a diverse test set of
comprising four distinct classes, each consisting of up to 2000
2000 images, and the results are presented in the chart. The
images per class. The primary objective was to evaluate the
model's performance varies across different domains, as
accuracy of the AI models post-training. The evaluation results
indicated by the misclassification counts in the respective
are conveyed through the presentation of the confusion matrix
categories. Notably, the ResNet50 model demonstrates a
below:
relatively strong performance in the 'Normal' category, with

727
only 6 misclassified images out of 2000. This underscores its [9] Gao, Y.; Liu, X.; Huang, H.; Xiang, J. A hybrid of FEM simulations and
robustness in standard image recognition tasks. generative adversarial networks to classify faults in rotor-bearing
systems. ISA Trans. 2021, 108, 356–366.
In the 'IR' category, the model performs moderately well [10] Liu, H.; Zhou, J.; Xu, Y.; Zheng, Y.; Peng, X.; Jiang, W. Unsupervised
with 9 misclassified images, indicating its ability to adapt to fault diagnosis of rolling bearings using a deep neural network based on
different image modalities. However, in more specialized generative adversarial networks. Neurocomputing 2018, 315, 412–424.
domains such as 'MI' , the model shows room for improvement, [11] Wang, R.; Jiang, H.; Li, X.; Liu, S. A reinforcement neural architecture
search method for rolling bearing fault diagnosis. Measurement 2020,
with 21 misclassified images. Furthermore, in the 'OR' category, 154, 107417.
the model exhibits a reasonable performance with 10 [12] He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image
misclassified images, suggesting its suitability for object recognition. In Proceedings of the IEEE Conference on Computer Vision
recognition tasks. and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp.
770–778.
In conclusion, the ResNet50 model excels in standard image [13] Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. In
recognition tasks but may require further fine-tuning or domain- Proceedings of the IEEE Conference on Computer Vision and Pattern
specific adaptations for optimal performance in specialized Recognition, Salt Lake City, UT, USA, 18–23 June 2018; pp. 7132–7141.
domains. Understanding the model's strengths and weaknesses [14] Huang, G.; Liu, Z.; Van Der Maaten, L.; Weinberger, K.Q. Densely
in different domains is essential for selecting the appropriate tool connected convolutional networks. In Proceedings of the IEEE
for specific image analysis applications. Conference on Computer Vision and Pattern Recognition, Honolulu, HI,
USA, 21–26 July 2017; pp. 4700–4708.
V. CONCLUSION [15] He, K., Zhang, X., Ren, S., & Sun, J. (2015). Deep Residual Learning for
Image Recognition. In Proceedings of the IEEE Conference on Computer
Deep learning algorithms offer numerous advantages for Vision and Pattern Recognition (CVPR) (pp. 770-778).
machinery fault diagnosis, including engines, transformers, and [16] Sarwinda, D., Paradisa, R. H., Bustamam, A., & Anggia, P. (2021). Deep
cutting machines. We anticipate a rapid increase in studies learning in image classification using residual network (ResNet) variants
for detection of colorectal cancer. Procedia Computer Science, 179, 423-
diagnosing roller bearing faults using deep learning. This 431.
paper's primary contribution lies in creating a dataset of three [17] Reddy, A. S. B., & Juliet, D. S. (2019, April). Transfer learning with
types of bearing faults and testing it with various state-of-the- ResNet-50 for malaria cell-image classification. In 2019 International
art deep learning networks. The findings highlight the Conference on Communication and Signal Processing (ICCSP) (pp.
advantages of the ResNet-50 network in this context, opening 0945-0949). IEEE.
new possibilities for classification tasks. In the future, we plan [18] A.Ghasempour, “Internet of Things in Smart Grid: Architecture,
Application, Services, Key Technologies, and Challenges,” Inventions
to expand and refine our dataset to enhance its Journal,vol.4,no.1,pp.1-12,2019,
comprehensiveness. Specifically, we will build and collect data https://doi.org/10.3390/inventions4010022.
with many different bearing sizes. Furthermore, we will find [19] Pham Van Nam and Tran Hoai Linh. "Electrocardiogram (ECG) Circuit
solutions to solve problems related to our big data for real-time Design and Using the Random Forest for ECG Arrhythmia
transmission to IOT systems. This helps engineers remotely Classification." International Conference on Engineering Research and
Applications, 1 Dec. 2022, page range 477-494.
monitor and diagnose machine errors
[20] Vuong Hoang Nam, Hoai Linh Tran, and Van Nam Pham. "Multiple
Neural Network Integration Using a Binary Decision Tree to Improve the
ECG Signal Recognition Accuracy." International Journal of Applied
REFERENCES Mathematics and Computer Science, 2014, Vol. 24, No. 3, 647–655.
[1] Çınar, Z. M., Abdussalam Nuhu, A., Zeeshan, Q., Korhan, O., Asmael, [21] A.Ghasempour and M.Martinez-Ramon, “ Short-Term Electric load
M., & Safaei, B. (2020). Machine learning in predictive maintenance Prediction in Smart Grid using Multi-Output Gaussian Processes
towards sustainable smart manufacturing in industry Regression” IEEE Kansas Power and Energy Conference (IEEE
4.0. Sustainability, 12(19), 8211. KPEC),2023
[2] Rai, A., & Upadhyay, S. H. (2016). A review on signal processing
techniques utilized in the fault diagnosis of rolling element
bearings. Tribology International, 96, 289-306.
[3] Saufi, S. R., Ahmad, Z. A. B., Leong, M. S., & Lim, M. H. (2019).
Challenges and opportunities of deep learning models for machinery fault
detection and diagnosis: A review. Ieee Access, 7, 122644-122662.
[4] Zhao, Z., Wu, J., Li, T., Sun, C., Yan, R., & Chen, X. (2021). Challenges
and opportunities of AI-enabled monitoring, diagnosis & prognosis: A
review. Chinese Journal of Mechanical Engineering, 34(1), 1-29.
[5] Rahman, T. A., Chek, L. W., & Ramli, N. Intelligent Vibration-based
Anomaly Detection for Electric Motor Condition Monitoring.
[6] Tahir, M. M., Khan, A. Q., Iqbal, N., Hussain, A., & Badshah, S. (2016).
Enhancing fault classification accuracy of ball bearing using central
tendency based time domain features. IEEE Access, 5, 72-83.
[7] Chen, Z.; Li, C.; Sanchez, R.V. Gearbox fault identification and
classification with convolutional neural networks. Shock. Vib. 2015,
2015, 390134.
[8] Zhao, J.; Yang, S.; Li, Q.; Liu, Y.; Gu, X.; Liu, W. A new bearing fault
diagnosis method based on signal-to-image mapping and convolutional
neural network. Measurement 2021, 176, 109088.

728