Dr. R. SAMSON DANIEL1, samson.rapheal@gmail.com
Associate Professor, Department of
ECE, K. Ramakrishnan college
of engineering, Tamilnadu
ABISHEK J2, AKSHAYARAJA E3,
BHARANI KUMAR P4, BOOMISHWAR S5
2abishek21062007@gmail.com, 3akshayarajae@gmail.com, 4bharanikumarb071@gmail.com,
5senthilbhoomis@gmail.com
Student’s, Department of ECE, K.
Ramakrishnan college of engineering, Tamilnadu
Induction
motors are widely useful in commercial and industrial systems due to its
efficiency, low cost and reliability. Faults such as shaft misalignment, rotor
damage, bearing wear can degrade the performance of the motor and lead to
various failures. Detecting these faults at early stage can reduce the downtime
and maintain operational maintenance. The vibration signal analysis is one of
the most trustable as mechanical faults generates the vibration patterns. These
vibrations suffer from noise and other faults limiting the accuracy of
detection methods based on the amplitude and threshold analysis. To avoid this
machine learning techniques and signal processing are applied. The system
involves a Butterworth filter and wavelet-based denoising for preprocessing,
continued by extraction of time-domain and frequency-domain features. Similarly,
reduction using Principal Component Analysis (PCA) enhances computational
efficiency, while Support Vector Machine (SVM) and Random Forest (RF)
algorithms classify motor conditions. Experimental results from simulated and
real-time vibration data’s shows that the proposed method achieves high
accuracy under different loads and noise conditions. The approach is designed to be suitable for predictive maintenance
in Industries.
Induction
motor fault detection has been a major research focus in the field of
monitoring and maintenance. To detect bearing wear, rotor bar breakage, and
misalignment using vibration signals various techniques are involved. Using
signal processing and machine learning methods we can enhance the detection
accuracy and overall performance under operating conditions.
Cusido et al proposed a wavelet-based
method for vibration signal analysis, Jimenez et al proposed a monitoring
system with the Hilbert and envelope for detection of fault at manufacturing. Widodo
and Yang combined the method of Support Vector Machine and wavelet
decomposition to improve fault discrimination.
The
studies incorporated some key concepts of Wavelet Transform, that represents
the faults patterns more accurately. For bearing diagnostics Hilbert envelope
demodulation technique was used. Similarly, Artificial Neural Networks, Support
Vector Machine, Random Forest Algorithms were used in order to improve the
reliability, complexity nature of the computational cost. From the survey it
shows the recent trend that integrates the fault detection and classifications
build the real time fault detection. There were still some obstacles while dealing with this system
such as high computational demand, limited adaptability especially for embedded
designs.
To
resolve these issues, these solutions in this paper involves Butterworth and
wavelet denoising with feature extraction, Principal Component Analysis
(PCA) for dimensionality reduction, and classification using SVM and
Random Forest. This structure helps
to achieve high accuracy, low
operation cost, under changing load and noise conditions, making it suitable
for maintenance applications in Industrial environments.
This
system is useful to develop a fault detection model for induction motors using
vibrational signals. This involvement of signal processing, classification, and
features are used to improve the accuracy.
The system
involves the following
steps:
[1] Signal Acquisition
[2] Preprocessing and Noise Filtering
[3] Feature Extraction
[4] Dimensionality Reduction
[5] Fault Classification
The
data vibrations are collected from motors using various conditions. These raw
signals are used to transformation techniques before being analyses by machine
learning.
High-frequency
vibrations are captured from healthy and faulty induction motors. Data are
collected by accelerometers connected to a Data Acquisition module. Frequencies
between 10 kHz and 20 kHz are used to ensure that frequency components are used
with high resolution.
To ensure
environmental noise and unwanted frequency components, the vibration signal is
passed through a Butterworth band-pass
filter.
where,
fc is the cutoff frequency and n is filtering order. Wavelet Denoising is applied using the Wavelet Transform to
compress high-frequency noise is gaining essential fault. The denoised signal
supervisor the clarity of periodic fault impulses.
This can be extracted from both time domain
and frequency domain:
·
Mean, RMS, Variance,
Kurtosis, and Skewness.
·
Spectral Energy, Peak Frequency, and Band Energy.
Addition of, Wavelet Packet Transform (WPT) is
delated too sub-band energy
coefficients, providing a localized of vibration patterns.
To
avoid this used to overfitting and reduce computation, Principal Component Analysis (PCA) is used. PCA projects are
high-dimensional feature into a reduced number of components that can be
preserve using the maximum variance. This step can be enhances using this model
efficiency and ensures faster classification in real-time applications.
It consists
of classifiers Support Vector Machine
(SVM) and Random Forest (RF).
·
The SVM separates the different fault classes with maximum margin.
·
The Random Forest classifier builds multiple decision
trees and averages their
outputs to achieve high performance
The
system is measured with the help of various parameters such as accuracy, precision, recall, and F1-score. Confusion matrix analysis is
used to visualize classification performance across all fault types. The system
depicts strong capability for early detection, supporting strategies.
The
proposed fault detection framework was implemented using MATLAB and Python to
check its performance on vibration data collected from induction motors under different
conditions—healthy, bearing fault, rotor bar defect, and misalignment. The
vibration signals were sampled at 12 kHz and processed through Butterworth
filtering and wavelet denoising before feature extraction.
Figure 1.
Figure1.
Represents that the raw vibration signal versus Butterworth-filtered signal
demonstrating effective noise suppression while preserving fault-relevant
vibration characteristics.
Classification performance: Principal Component Analysis and
are differentiated using Support Vector Machine and Random Forest Algorithm.
|
Fault Condition |
SVM Accuracy (%) |
RF Accuracy (%) |
|
Healthy |
98.3 |
97.9 |
|
Bearing Fault |
97.5 |
98.1 |
|
Rotor Fault |
96.8 |
97.3 |
|
Misalignment |
97.9 |
98.4 |
|
Average Accuracy |
97.6 |
97.9 |
The results
confirm that both classifiers
achieved high accuracy, with Random.
Figure 2.
Frequency spectrum of the
filtered vibration signal obtained using Fast
Fourier Transform (FFT)
illustrating dominant fault-related vibration components are mentioned in figure 2. The Butterworth filtering and wavelet denoising increase the signal
to noise ratio by 25% improving the performance and clarity. The denoised signal shows
periodic impulse to fault frequency. On analyzing it shows that RMS, Kurtosis, and Wavelet Packet
Energy form a framework for the accuracy
classification. PCA reduced the
reductant information and performance time by 40%. These classifiers maintained stable
Operation under discrete conditions.
Figure 3.
Confusion matrix representing fault classification accuracy
across motor conditions using trained ML model in figure
3. The incorporation of signal processing with this classification
performance provides a symmetry between the efficiency and the accuracy of the
model. This model approaches for a low-cost maintainability operating efficiently on limited data making it suitable for
embedded systems.
The
proposed model provides a smart solution for fault detection in induction
motors using vibration signal analysis. This model inculcates Butterworth
filtering, features extraction, wavelet based denoising, including random
forest algorithms. On analyzing the results of our experiment, we came to know
that our model ensures high accuracy delivering bearing detection, rotor and
other faults under wide operating conditions.
This
model not only shows the strength of signal processing but also the predictive
strength of machine learning improving accuracy, to withstand noise suitable
for diverse industrial applications. This system mainly processes on limited
data suitable for low-cost maintenance.
The
future development of this model includes integration with adaptive learning
for integrating with sensor data, real time issue estimation and foe
eliminating the design on embedded for onsite motor monitoring.
[1] G. A.
Jiménez, et al., “Online Motor Fault Detection Using Hilbert and Envelope
Analysis,” IEEE Transactions on
Industrial Applications, vol. 56, no. 4, pp. 4123–4132, 2020.
[2]
J. Cusidó, et al., “Wavelet-Based
Vibration Analysis for Induction Motor Fault
Detection,” Mechanical
Systems and Signal Processing, vol. 158,
pp. 107761, 2021.
[3] A. Widodo and
B. Yang, “Support Vector Machine-Based Condition Monitoring for Induction
Machines Using Wavelet Features,” Expert
Systems with Applications, vol. 41, no. 6, pp. 2913–2921, 2022.
[4] P. Kumar, et
al., “Wavelet Packet Transform for Bearing Fault Classification in Induction
Motors,” Measurement, vol. 200, pp.
111054, 2022.
[5] H. Park, et
al., “Hilbert Envelope Demodulation for Bearing Fault Identification,” IEEE Sensors Journal, vol. 21, no. 18,
pp. 20534–20542, 2021.
[6] M.-C. Kim,
J.-H. Lee, D.-H. Wang, and I.-S. Lee, “Induction Motor Fault Diagnosis Using Support Vector Machine, Neural Networks,
and Boosting Methods,” Sensors, vol. 23, no. 5, p. 2585, 2023.
[7] R. N. Toma
and J.-M. Kim, “Bearing Fault Classification of Induction Motors Using Discrete
Wavelet Transform and Ensemble Machine Learning Algorithms,” Applied Sciences,
vol. 10, no. 15, p. 5251, 2020.
[8] C.-Y. Lee and
Y.-H. Cheng, “Motor Fault Detection Using Wavelet Transform and Improved PSO-BP
Neural Network,” Processes, vol. 8, no.
10, p. 1322, 2020.