Fault Detection in Induction Motor Using Time Domain and Spectral Imaging-Based Transfer Learning Approach on...

Area of Science:

• Electrical Engineering
• Machine Learning
• Industrial Systems

Background:

• Induction motors are crucial for industrial drives but prone to rotor faults like broken bars.
• Early fault detection is essential to reduce maintenance costs and prevent failures.

Purpose of the Study:

• To develop an effective method for classifying induction motor rotor faults under varying load conditions.
• To compare the performance of different feature extraction domains and classification models.

Main Methods:

• Utilized an open-source dataset of induction motors with broken rotor bars.
• Performed feature extraction in time, frequency, and time-frequency domains.
• Employed Random-Forest (RF) and Convolutional Neural Network (CNN) models for classification.

Main Results:

• Time and frequency domain features with RF achieved up to 88.58% accuracy.
• Short Time Fourier Transform (STFT) spectrograms with a fine-tuned CNN achieved 97.67% accuracy.
• Time-frequency analysis proved superior for diagnosing rotor bar severity.

Conclusions:

• The proposed CNN-based transfer learning framework using STFT spectrograms offers a highly accurate solution for induction motor rotor fault diagnosis.
• Time-frequency domain analysis is critical for effective fault severity assessment in induction motors.