Comparative Study between Physics-Informed CNN and PCA in Induction Motor Broken Bars MCSA Detection.

Area of Science:

• Electrical Engineering
• Machine Condition Monitoring
• Fault Diagnosis

Background:

• Broken bars are a common fault in induction motors, leading to performance degradation and potential failure.
• Effective detection methods are crucial for predictive maintenance and operational reliability.
• Existing techniques often require specific operating conditions or complex signal processing.

Purpose of the Study:

• To evaluate and compare two distinct methods for broken bar fault detection in induction motors.
• To assess the performance of Motor Current Signature Analysis (MCSA) combined with Convolutional Neural Networks (CNN).
• To analyze the efficacy of Principal Component Analysis (PCA) in detecting induction motor anomalies.

Main Methods:

• Motor Current Signature Analysis (MCSA) utilizing Convolutional Neural Networks (CNN) with frequency-domain processing.
• A novel double-input CNN architecture for robust detection.
• Principal Component Analysis (PCA) applied to time-domain current signals to calculate the Q statistic for anomaly detection.

Main Results:

• The CNN-based MCSA approach achieved 100% detection accuracy for broken bars, independent of motor speed and load torque.
• PCA demonstrated effective anomaly detection in the time domain by analyzing the Q statistic.
• Both methods proved highly effective, though significant differences in their application and outcomes were observed.

Conclusions:

• Both MCSA with CNN and PCA are viable and effective techniques for broken bar detection in induction motors.
• The choice between methods may depend on specific application requirements, data availability, and desired processing domain (frequency vs. time).
• Further analysis is warranted to fully delineate the practical advantages and limitations of each approach in diverse industrial settings.