Motor Fault Diagnosis Across Variable Power Using Deep Learning

Induction motors are crucial machines in industrial operations. Their unexpected failures can lead to downtime and operational stop. This research presents a general fault diagnosis system based on deep learning, specifically designed to classify motor faults across a wide range of power levels. The objective is to create a robust and generalizable model that can accurately identify faults under different fault scenarios. Time-series data including RPM, torque, power, vibration, rotor and stator currents, flux linkages, and terminal voltages were generated using MATLAB Simscape simulations of induction motors with power ratings of 4 kW, 8 kW, 15 kW, 37 kW, 75 kW, 110 kW, and 150 kW. Each motor configuration was simulated under five operating conditions: normal, broken bearing, broken rotor bar, stator winding short circuit, and voltage imbalance, resulting in 35 different motor-fault combinations. A transformer deep learning model was trained with this high-resolution dataset. The model consists of a linear embedding layer, stacked encoder blocks, 8 multi-head self-attention, and a global average pooling layer. The model was trained on a GPU-accelerated HPC platform. Results demonstrate excellent classification performance, achieving 100% accuracy on both validation, training and test sets. The proposed framework effectively captures complex fault patterns across power levels and offers reliable solutions for industrial motor health monitoring systems.