Advanced Deep Learning Approaches for Fault Detection and Diagnosis in Inverter-Driven PMSM Systems

This paper presents a comprehensive approach to fault detection and diagnosis (FDD) in inverter-driven Permanent Magnet Synchronous Motor (PMSM) systems through the innovative integration of transformer-based architectures with physics-informed neural networks (PINNs). The methodology addresses critical challenges in power electronics reliability by incorporating domain-specific physical constraints into the learning process, enabling both high accuracy and physically consistent predictions. The proposed system combines advanced sensor fusion techniques with real-time monitoring capabilities, processing multiple input streams including phase currents, temperatures, and voltage measurements. The architecture’s dual-objective optimization approach balances traditional classification metrics with physics-based constraints, ensuring predictions align with fundamental electromagnetic and thermal principles. Experimental validation using a comprehensive dataset of 10,892 samples across nine distinct fault scenarios demonstrates the system’s exceptional performance, achieving 98.57% classification accuracy while maintaining physical consistency scores above 0.98. The model ex-hibits robust performance across varying operational conditions, including speed variations (97.45-98.57% accuracy range) and load fluctuations (97.91-98.12% accuracy range). Notable achievements include perfect detection rates for certain critical faults, such as high-side short circuits and thermal anomalies, with area under ROC curve (AUC) scores of 1.0. This research establishes new benchmarks in condition monitoring and fault diagnosis for power electronic systems, offering practical implications for predictive maintenance and system reliability enhancement.