Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

Optimizing Induction Motor Fault

Detection with Transfer Learning: A Comparative
Analysis of Deep Learning Models
Hakeem Issah1; Asante Prince Kwabena2; Boateng Kelvin Osei3; Elvis Afful4; Norbert Awuah5; Alhassan Osumanu6
1,2,3,4,5,6
Department of Electrical and Electronic Engineering, Kwame Nkrumah
University of Science and Technology, Kumasi, Ghana

Abstract:- With the advancements of Industry 4.0, I. INTRODUCTION

condition monitoring maintenance has become essential
for preventing equipment failures and operational The most well-known method of converting electrical
disruptions. Motor Current Signature Analysis (MCSA) to mechanical energy is the use of an induction motor, also
is commonly utilized for condition monitoring to detect called an asynchronous motor. This is because it runs at a
and diagnose various faults in Induction Motors (IMs). speed lower than the synchronous speed. The operation of
Despite its popularity, there is limited research the induction motor is based on the principle of induction of
comparing deep learning models for Induction Motor EMFs and currents in the rotor that is not directly connected
fault detection and classification with traditional to any power supply.
approaches. This study explores the detection and
classification of Induction Motor faults using three Induction motors are the most widely used type of
Transfer Learning (TL) models: InceptionV3, electric motor in residential, commercial, and industrial
ResNet152, and VGG19. applications. They are favored for their affordability,
durability, and dependable performance, even under
The research began by modeling a Squirrel Cage challenging environmental conditions, including potentially
induction motor in MATLAB to simulate healthy, single- explosive atmospheres. These qualities make induction
phasing, and double-phasing conditions, capturing time- motors a preferred choice over other motor types in a variety
domain stator current signatures (current spectrum) to of settings. This motor plays an important role in modern
identify fault characteristics. The data were then used to industrial plants. They are widely used due to a large number
assess the effectiveness of the TL models in detecting and of favourable features such as low price, reliability, rugged
classifying motor faults. Around 500 datasets were construction, and low maintenance costs [1]. With the
created from these simulated conditions, labeled increasing evolution in industrial processes, induction
accordingly, and used to train and validate the TL motors have replaced 90 percent of the actuators altogether
models, each incorporating additional convolutional exercised in the production line and were surveyed to be
layers to enhance performance. Model evaluation more fault tolerant [2].
utilized metrics such as the multiclass confusion matrix,
precision, recall, and F1-score across various fault While induction motors are known for their reliability,
scenarios. they can be affected by environmental conditions,
operational demands, and installation issues, leading to
Results indicate that stator current signatures can various failures that shorten their expected lifespan. Faults
effectively reveal individual faults, with ResNet152 in induction motors progress through three stages [3]. The
outperforming the other models in classification first stage, known as the incipient fault, marks the initial
accuracy. These findings highlight that applying transfer degradation of internal components. At this stage, despite
learning techniques with a limited amount of current some damage, the motor typically continues to function
signature data can support predictive maintenance in without noticeable issues. The second stage, or developed
industrial settings, potentially reducing costly equipment fault, involves more extensive damage that significantly
shutdowns and disruptions in production. impacts motor performance, though it remains operational.
In the final stage, called a catastrophic fault, the damage has
Keywords:- Convolutional Neural Network, Transfer spread, affecting multiple components and causing the
Learning, Simulink. motor to cease functioning altogether.

IJISRT24NOV003 www.ijisrt.com 398

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

 Faults: 5–10 percent. Mechanical faults, including issues like

Induction motor (IM) faults generally fall into two bearing and eccentricity faults, account for roughly 40–50
primary categories: mechanical and electrical, as illustrated percent of motor breakdowns. Given the significant impact
in Fig. 1. According to data from the Institute of Electrical of faults in core components such as stators, rotors, and
and Electronics Engineers (IEEE) and the Electric Power bearings, this study focuses on the most prevalent motor
Research Institute (EPRI), stator electrical faults represent faults.
30–40 percent of issues, while rotor faults contribute about

Fig 1: Classes of Faults in Induction Motors

Failures of motor core components such as stators, considerable energy savings and lowered running costs. In
rotors and bearings account for a large percentage of motor sum, early detection of motor faults is essential for
breakdowns. Hence in this setting, the most common motor maintaining consistent production and remaining
faults would be examined. competitive in the industry.

In terms of stator faults, the stator includes a laminated II. LITERATURE REVIEW
core, external frame, and insulated windings, all of which
experience electrical and environmental stresses that can lead Research into fault detection in induction motors
to failures. Stator faults are typically classified based on their employs various methodologies, each providing insights into
location: they may occur in the stator frame, the winding, or challenges in scalability, accuracy, and adaptability. [5]
the laminations of the stator core. Among these, winding utilizes cyclostationarity to capture the periodic
failures are particularly serious, often resulting from characteristics of electrical signals, enhancing early fault
insulation breakdown. This leads to localized overheating, detection by identifying subtle statistical changes. However,
which, if not detected, can cause further insulation damage, its adaptability is limited when applied to motors under
potentially resulting in a catastrophic short circuit inter-turn varying loads. [6] implements Convolutional Neural
fault [4]. Networks (CNNs) for three-phase induction motor
diagnostics, excelling in fault pattern recognition. Yet, CNNs
Understanding the behavior of induction motors (IMs) encounter computational constraints, making real-time
under fault conditions and diagnosing these issues has posed applications challenging. Similarly, [7] employs Two-
a longstanding challenge for researchers in electrical dimensional Time-Domain Gray Coded Image (TDGCI)
machinery. Common motor faults are often associated with coupled with CNNs to diagnose rotor faults. While effective
key components like stators, rotors, and bearings. If these in identifying visual fault patterns, the reliance on Fast
faults are not identified in their initial stages, motor Fourier Transform (FFT) preprocessing limits TDGCI’s
performance can deteriorate, potentially leading to complete effectiveness in low-severity conditions where FFT is noise-
failure. Detecting faults early in IMs brings significant sensitive.
benefits to industrial operations by enabling cost-effective
failure prediction and proactive maintenance planning. This [8] explores random multi-frequency resonant sparse
approach allows for timely preventive measures, reducing noise power spectrum (rMFRSNPS) in conjunction with
the need for costly part replacements and preventing probability vector resonance analysis (PVRA), effectively
unplanned production halts and downtime [1]. Additionally, identifying faults in isolated conditions. However, the
early fault detection contributes to motor efficiency by method struggles to generalize to complex fault scenarios,
addressing operational inefficiencies, resulting in limiting its applicability in environments with multiple

IJISRT24NOV003 www.ijisrt.com 399

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

simultaneous faults. In [9], Discrete Wavelet Transform III. THEORY

(DWT) is used to decompose current signals, aiding stator
fault diagnosis. While advantageous for detailed signal A. Problem Overview
analysis, DWT faces scalability issues, particularly when Detecting faults early is essential to prevent electrical
adapting to real-world, multi-variable conditions. [10] failures, insulation damage, and complete production
further utilizes CNNs, focusing on rotating machinery fault stoppages as faults progress to advanced stages. Early fault
detection through image feature extraction. Though CNNs detection methods based on signal analysis involve capturing
improve recognition of intricate fault patterns, their high one or more physical properties of the induction motor,
computational demand continues to present a barrier to real- processing these signals with appropriate techniques to
time, large-scale diagnostics. identify fault patterns, and analyzing these patterns to
classify the fault type. Recent advancements in
[11] investigates current signature monitoring, a computational technology have enhanced condition
straightforward approach for identifying bearing damage. monitoring methods, integrating signal processing
Despite its simplicity and diagnostic utility, this method techniques in both time and frequency domains with
lacks robustness across varied operational conditions. [12] heuristic approaches like machine learning, genetic
applies Wavelet Packet Transform (WPT) for enhanced algorithms, artificial intelligence, and deep learning. This
signal decomposition, offering insights into complex fault combination offers improved accuracy and effectiveness in
signals. However, WPT’s high computational requirements early fault detection.
hinder its scalability, particularly in larger motor systems
with overlapping fault conditions. Transfer learning, also B. Approach
explored in [12], shows promise in broadening the In our approach, we adopt a quantitative methodology,
applicability of diagnostic models but requires further focusing on data collection and analysis to derive insights
investigation to support diverse motor types and conditions. and make predictions while minimizing bias and ensuring no
[13] presents the Park’s Vector Approach (PVA) for fault key factors are overlooked. Traditional machine learning
detection in low-severity cases. While PVA can identify (ML) techniques classify healthy and faulty motor conditions
subtle abnormalities, its reliability decreases in noisy by using extracted features from signals in various
environments, limiting its utility in real-world scenarios. representations (time, frequency, and time-frequency
domains) as inputs to knowledge-based systems. To
[14] introduces Adaptive Neuro-Fuzzy Inference overcome some limitations of these ML methods, we
Systems (ANFIS), which combine neural network learning propose using a deep learning model, specifically
with fuzzy logic for motor fault diagnosis. Although ANFIS Convolutional Neural Networks (CNNs), to address the
models offer adaptability and precision, their computational following issues:
complexity may hinder scalability. [15] uses Support Vector
Machines (SVMs) to classify fault patterns, demonstrating  Traditional methods rely heavily on hand-crafted feature
solid performance in well-defined data sets but facing engineering and feature selection from raw data.
limitations with more nuanced or overlapping faults. Lastly,  Feature significance can vary with changing conditions.
[16] implements Principal Component Analysis (PCA)  Advanced signal processing is needed to reduce noise in
combined with K-means clustering to identify fault patterns. the signal, adding complexity to feature extraction
While PCA aids in dimensionality reduction, the model’s
effectiveness decreases when confronted with high- C. Design
dimensional, noisy data common in motor diagnostics.
 Induction Motor Model:
In summary, these studies reveal significant Modelling is done in MATLAB Simulink, which
advancements in fault detection methodologies, yet involves creating a simplified representation of a real-world
challenges remain in scalability, adaptability, and real-time induction motor using mathematical equations, specific
application. Many methods, such as CNNs, WPT, and parameters, and assumptions. The model is designed to
ANFIS, excel in controlled environments but struggle under replicate the behavior and characteristics of the actual 3-
varied operational loads or noisy conditions, as seen in [6], phase induced motor being simulated, allowing analysis and
[10], and [12]. Traditional techniques like FFT and PVA understanding of the system’s performance and obtaining the
demonstrate limited sensitivity to low-severity faults ([7] and necessary signatures under different conditions (healthy,
[13]). To enhance these methods, future research must focus single-phasing, phase-phase). MATLAB Simulink would be
on reducing computational demands, increasing model used, where many of the components to be used for modeling
robustness across diverse motor conditions, and improving are found, like the AC Electrical Elements, Induction motor
noise resilience to make these systems viable for broader, block, three-phase source, and others. The blocks needed are
practical applications. easily dragged and dropped unto the Simulink window, and
their parameters like rated power, rated voltage, torque and
rated frequency, are specified. In our experiment, healthy and
stator faults (single-phasing, phase-to-phase) would be
simulated alongside healthy conditions. To simulate a stator
fault, we disconnect one or more phases of the stator
winding.

IJISRT24NOV003 www.ijisrt.com 400

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

Fig 2: Model of Three-Phase Asynchronous Motor

 Data Processing: During training, input data (images) are fed into the
Before inputting data into the Transfer Learning model, CNN, where essential features are identified and extracted,
which would be RGB images, distinguishing healthy and ensuring consistent feature selection across samples, and
stator fault conditions, it is essential to clean and pre-process class predictions are generated. The model’s predictions are
the data to ensure accurate and effective training of the then compared to actual labels, with errors measured through
model. This action is executed with libraries that include a loss function. The model’s parameters are updated
TensorFlow, Scikit-Learn, and NumPy in Google Colab. accordingly to minimize this loss. Validation is performed
Some common steps we take in cleaning data for a CNN after each training cycle to monitor the model’s performance
model include: and prevent overfitting. In this phase, hyperparameters, such
as learning rate, batch size, filter count, epochs, and padding,
 Data Exploration: This involves understanding the are fine-tuned to optimize accuracy on the validation set,
dataset, visualizing and analyzing the data, identifying enhancing the model’s generalization to new, unseen data.
missing values, and removing irrelevant or redundant
data.  The CNN Base Layer:
 Data Pre-processing: This step includes converting data A specific type of CNN deep learning algorithm can
into a machine-readable format, scaling the data, and process an input image, apply weights and biases to
removing noise or outliers. recognize significant elements, and distinguish between
 Data Augmentation: Used to improve the diversity of the different objects in the image. Unlike traditional methods
training dataset by using transformations such as flipping, that require manual engineering of filters, ConvNets can
rotating, and shifting the images. learn these distinguishing characteristics with sufficient
 Data Labelling: In the dataset, each image is assigned a training, greatly reducing pre-processing needs compared to
label that corresponds to its class, such as a cat or dog. other classification algorithms.

Convolutional Neural Networks (CNNs) are trained One key benefit of CNNs over other neural networks is
using the backpropagation algorithm, which adjusts network their ability to identify critical features autonomously,
weights based on error rates calculated during training. This without human intervention. CNNs are also highly
optimization process aims to reduce the discrepancy between computationally efficient due to their use of convolution and
predicted and actual outputs by updating the network’s pooling operations, along with parameter sharing. This
weights. After data cleaning, the dataset is divided into a efficiency makes CNNs adaptable to various devices,
training set (83%) and a validation set (17%). The training enhancing their universal appeal. Furthermore, CNNs reduce
set is used to fit the model, while the validation set is used the need for pre-processing while learning distinctive filters
for tuning hyperparameters. and features on their own. CNNs also offer computational

IJISRT24NOV003 www.ijisrt.com 401

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

advantages over traditional neural networks, with weight  Transfer Learning Model:
sharing being a major asset. To enhance the training process for a new model on a
related task, we introduce a pre-trained model as the
CNN architecture is inspired by the human brain’s foundational layer. Transfer learning leverages the pre-
connectivity patterns, specifically the organization of the existing knowledge of this model, which has been trained on
visual cortex. Each neuron in a CNN responds to a small the extensive ImageNet dataset, to capture general features
region in the visual field, known as its receptive field, and and patterns. This approach enables efficient training on a
these fields collectively span the entire visual area. In our smaller dataset by reusing learned representations,
model, several CNN layers are used, reflecting these minimizing the computational demands and data
concepts in their structure. requirements. The pre-trained CNN model is fine-tuned and
adapted with our new dataset, functioning as the initial layer
The Flatten Layer in CNNs serves to reshape the of the transfer model for feature extraction. Integrating
multidimensional tensors generated by preceding transfer learning into our experimental model offers several
convolutional and pooling layers in the TL model into a one- advantages:
dimensional vector, preparing them for input into fully
connected layers. This layer acts as a bridge between the  The use of pre-trained knowledge significantly reduces
spatial feature maps extracted by TL layers and the linear the need for extensive data and computational power for
structure required by fully connected layers for classification training from the ground up.
or regression. Mathematically, if the input tensor has  It enhances the accuracy of the new model.
dimensions (batch size, height, width, channels), the flatten  The resulting model is better equipped to handle data
layer converts it into a vector of shape (batch size, height * variability, noise, and outliers.
width * channels), effectively rearranging the data for
seamless integration with subsequent dense layers. This Transfer models like InceptionV3, VGG 19 and
transformation facilitates the learning of higher-level ResNet152 are imported from the libraries of the Keras
relationships in the data. library. All would be trained with the base layer’s small
dataset, and the model that gives the best performance would
In a Convolutional Neural Network (CNN), the fully be chosen for our experimental model. Not forgetting, the
connected (or dense) layer serves to translate the high-level layers and dimensions of the base layer and transfer model
features identified by the convolutional layers into the final are same, preventing overfitting.
model output, whether for classification or regression tasks.
This dense layer is composed of neurons, each connected to
every neuron in the preceding layer. It computes a weighted
sum of inputs from the prior layer and typically applies an
activation function, such as softmax, to generate the final
predictions. For multi-class classification, the softmax
function (S) transforms the last layer’s outputs into a
probability distribution across mutually exclusive classes. In
contrast, for binary classification, a sigmoid function is
applied, categorizing the outcome as 0 or 1.

Fig 3: Architecture of the Flatten (in blue) and the Fully

Connected Layer (in Black)

IJISRT24NOV003 www.ijisrt.com 402

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

Fig 4: Transfer Learning Model Architecture

RESULTS Table 1: Split Dataset

A. Simulation
After the Simulink simulation, 276 current time
signatures where obtained at various time range and load.
Signatures between the training and validation dataset was
82 and 18 percent respectively.

B. Performance of Transfer Learning Models

In existing research, deep learning (DL) models have
been shown to effectively classify faults using raw current
signature data and frequency-based features extracted from
it. These models can identify motor health states (e.g.,
Fig 5: Healthy Waveform at no Load Healthy, Single phasing, Phase to Phase) and determine
which specific phase is affected. To evaluate the
classification capabilities of DL models, we created a dataset
by recording current signature data in the time domain from
a simulated motor model in various operating states within
MATLAB/Simulink. The collected data was then saved in
CSV format. After appropriate labeling, the dataset was used
for training and validating the DL models. Detailed
information about the dataset is outlined in Table 4, which
includes approximately 300 samples for each condition of
the motor, both healthy and faulty.

Table 2: Classes and their Labels

Condition Class Label
Healthy Healthy 0
Single-Phasing Single-Phasing 1
Phase-Phase Phase-Phase 2

Fig 6: Single Phasing Fault Waveform at 25% Load

IJISRT24NOV003 www.ijisrt.com 403

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

For this research, transfer learning models such as to their varied layer structures. The choice of model
InceptionV3, ResNet50, and VGG19 were utilized due to architecture, including the number of layers and units, is
their strong capabilities in analyzing time-series and influenced by the dataset’s nature and complexity. In cases
sequential data. These models are known for their advanced where the input data is complex and nonlinear, deeper
learning abilities, even when working with raw input data, models may be necessary to achieve optimal performance.
and can accurately predict outcomes as multiclass labels due

Table 3: Architecture of Modified Transfer Learning Model

Layers Units
InceptionV3 ResNet152 VGG19
Pre-trained Model 21,802,784 58,370,944 20,024,384
Flatten 1 1 1
1 x Dense 128 128 128
1 x Dense 64 0 64
Output 3 3 3

Table 4: Hyper Parameters of TL Models

Hyper Parameters InceptionV3 ResNet152 VGG19
Learning Rate 0.0001 0.001 0.0001
Batch Size 16 28 32
Loss function Categorical cross entropy Categorical cross entropy Categorical cross entropy
Epochs 20 15 20

We have trained and tested the 3 TL models for fault predicts different classes in a multi-class classification
detection and classification in Google Colab using the problem. The confusion matrix is constructed based on the
dataset elaborated in Table 1. The Confusion matrix comparison of predicted class labels and the actual ground
evaluates the performance of classification models. It truth labels of the dataset. The matrices of the three TL
provides a comprehensive overview of how well a model models are illustrated in Figure 7, 8, and 9.

Fig 7: Confusion Matrix of InceptionV3

IJISRT24NOV003 www.ijisrt.com 404

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

Fig 8: Confusion Matrix of ResNet153

Fig 9: Confusion Matrix of VGG19

From the matrices, various performance metrics are From 10, the performance metrics of ResNet152 stands
derived for insight. One is the accuracy of the individual out with precision, recall and f-1 score of 97%, 96% and
conditions, shown in Table 3. It shows that ResNet152 97% respectively, due to its complexity and better feature
exhibits a better model or algorithms for each condition. representation. Also in 11, this model leads in the accuracy
Other metrics include: Precision: It represents the proportion domain with a training accuracy and validation accuracy of
of correctly predicted positive instances among all instances 96.49% and 97.92% respectively. The least losses displayed
predicted as positive. Formula: P = TP / (TP + FP) are by the ResNet152 model in 12, with values main loss
being 12.59% and validation losses, 14.1%. It can be said
Recall: It represents the proportion of correctly the ResNet152 model has the best suitable architecture that
predicted positive instances among all actual positive captures the underlying patterns in the data effectively. Also
instances. Formula: R = TP / (TP + FN) the hyperparameters were tuned optimally. It also has the
best weight initialization to help the optimization process
F1-Score: The harmonic mean of precision and recall. find a more optimal set of parameters during training.
It balances precision and recall and is useful when dealing
with imbalanced classes. Formula: 2 * (Precision * Recall) /
(Precision + Recall)

IJISRT24NOV003 www.ijisrt.com 405

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

Table 5: Condition Accuracy Comparison of TL Models

Class InceptionV3(%) ResNet152(%) VGG19(%)
Healthy 75 100 75
Single Phase 81 94 88
Phase to Phase 100 100 100

Fig 10: Performance Scores of TL Models

Fig 11: Accuracy Comparison of TL Models

Fig 12: Losses Comparison of TL Models

IJISRT24NOV003 www.ijisrt.com 406

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

IV. CONCLUSION [6]. Khanjani, M., Ezoji, M. (2021). Electrical fault

detection in three-phase induction motor using deep
This paper explores the detection and classification of network-based features of thermograms.
faults in three-phase induction motors by analyzing current Measurement, 173, 108622.
signatures in the time domain using deep learning [7]. Gundewar, S., Kane, P., Andhare, A. (2022).
techniques. Simulations were conducted in MATLAB to Detection of broken rotor bar fault in an induction
introduce faults such as single phasing and phase-to-phase motor using convolution neural network. Journal of
disruptions into healthy waveforms. Due to the challenge of Advanced Mechanical Design, Systems, and
distinguishing between different fault types, accurately Manufacturing, 16(2),
classifying the affected phase(s) is essential for condition JAMDSM0020JAMDSM0020.
monitoring systems. To address this, transfer learning [8]. Huang, Z., Wang, T., Liu, W., Valencia-Cabrera, L.,
models—InceptionV3, ResNet152, and VGG19—were fine- Perez-Jim´ enez, M. J., Li, P. (2021). A fault analysis
tuned, trained, and evaluated using a locally generated method´ for three-phase induction motors based on
dataset representing both healthy and faulty motor states. spiking neural P systems. Complexity, 2021(1),
Each model was tested for its ability to classify different fault 2087027.
conditions based on raw stator current data. The findings [9]. Almounajjed, A., Sahoo, A. K., Kumar, M. K. (2021).
indicated that these models achieved effective fault detection Diagnosis of stator fault severity in induction motor
and classification, with performance evaluated through based on discrete wavelet analysis. Measurement,
metrics including the confusion matrix, precision, recall, F1- 182, 109780.
score, and average score. Of the three models, ResNet152 [10]. Lee, J. H., Pack, J. H., Lee, I. S. (2019). Fault
outperformed the others and was therefore selected as the diagnosis of induction motor using convolutional
best option for meeting the project’s objectives. neural network. Applied Sciences, 9(15), 2950.
[11]. Barcelos, A. S., Cardoso, A. J. M. (2021). Current
RECOMMENDATION based bearing fault diagnosis using deep learning
algorithms. Energies, 14(9), 2509.
A larger dataset, probably in the millions, used for [12]. Hussein, A. M., Obed, A. A., Zubo, R. H., Al-Yasir,
learning, would significantly improve the performance of the Y. I., Saleh, A. L., Fadhel, H., ... Abd-Alhameed, R.
TL models. This project can also be extended to include rotor A. (2022). Detection and Diagnosis of Stator and
faults and eccentricity faults. Such robust models should be Rotor Electrical Faults for Three-Phase Induction
tested using real-time industry current signatures. Motor via Wavelet Energy Approach. Electronics,
11(8), 1253.
ACKNOWLEDGMENT [13]. Garcia-Calva, T. A., Morinigo-Sotelo, D., Fernandez
Cavero, V., Garcia-Perez, A., Romero-Troncoso, R.
We would like to express our gratitude first to Almighty D. J. (2021). Early detection of broken rotor bars in
for life. Also our abled supervisor and Dean of the Electrical inverter-fed induction motors using speed analysis of
Engineering Department, Professor Emmanuel Assuming startup transients. Energies, 14(5), 1469.
Frimpong, and Dr. Daniel Opoku for their guidance and [14]. Gyftakis, K. N., Marques-Cardoso, A. J. (2019,
technical expertise during conduction of this research. October). Reliable detection of very low severity
Finally, Mr. Ibrahim Musah of RAIL-KNUST, friends and level stator inter-turn faults in induction motors. In
family for all support garnered. IECON 2019-45th Annual Conference of the IEEE
Industrial Electronics Society (Vol. 1, pp. 1290-
REFERENCES 1295). IEEE.
[15]. Hussain, M., Memon, T. D., Hussain, I., Ahmed
[1]. T. G. Calva, D. M. Sotelo, V. F. Calero, R. R. Memon, Z., Kumar, D. (2022). Fault Detection and
Tronsoco “Early Detection of Faults in Induction Identification Using Deep Learning Algorithms in
Motors—A Review”, Energies, pp. 2 Induction Motors. CMES-Computer Modeling in
[2]. A. A. Qazi, J. Daudpoto, S. A. Shaikh “Comparison Engineering Sciences, 133(2).
of Fault Detection Techniques for Induction Motors”, [16]. Morales-Perez, C., Rangel-Magdaleno, J.,
International Journal of Computer Applications (0975 PeregrinaBarreto, H., Amezquita-Sanchez, J. P.,
– 8887), Volume 183, pp. 1 Valtierra-Rodriguez, M. (2018). Incipient broken
[3]. Sabir, H.; Ouassaid, M.; Ngote, N. An experimental rotor bar detection in induction motors using vibration
method for diagnostic of incipient broken rotor bar signals and the orthogonal matching pursuit
fault in induction machines. Heliyon 2022, 8, e09136. algorithm. IEEE Transactions on Instrumentation and
[4]. Siddique, A.; Yadava, G.; Singh, B. A review of stator Measurement, 67(9), 2058-2068.
fault monitoring techniques of induction motors.
IEEE Trans. Energy Convers. 2005, 20, 106–114.
[5]. Sabir, H., Ouassaid, M., Ngote, N. (2022). An
experimental method for diagnostic of incipient
broken rotor bar fault in induction machines. Heliyon,
8(3).

IJISRT24NOV003 www.ijisrt.com 407

Volume 9, Issue 11, November – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24NOV003

APPENDIX A

CODE AND ORIGINAL DATASETS USED FOR PROJECT

 VGG MODEL: VGGGitHub

 INCEPTIONV3: GitHub
 RESNET152: GitHub
 DATASET: Google Drive
 INDUCTION MOTOR MODEL: Google Drive

IJISRT24NOV003 www.ijisrt.com 408