Fault Detection in Electric Drives Using Neural Networks and Physics-Based Data Generation

With the global trend toward electrification, ensuring the reliability and effective fault prevention of electric drive systems has become a crucial aspect of industrial and automotive design. This thesis presents a robust, data-driven framework for the detection and classification of faults in Permanent Magnet Synchronous Motors (PMSMs), with a specific focus on Interior Permanent Magnet Synchronous Motors (IPMSMs).The proposed methodology integrates physics-based modeling with advanced deep learning techniques. Using Matlab Simscape, a high-fidelity model of the IPMSM and its control system was developed to simulate a wide range of operating conditions and critical failure modes, including stator short-circuits, rotor demagnetization, and resolver sensor faults. To overcome the computational limitations of time-domain simulations, an efficient data generation strategy based on high-resolution Look-Up Tables (LUTs) was implemented. This approach enabled the creation of large-scale synthetic datasets emulating dynamic, real-world automotive driving cycles.The diagnostic performance of different neural network architectures was investigated, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and hybrid CNN-LSTM models. Results demonstrate that the hybrid CNN-LSTM architecture provides the highest accuracy by effectively combining local feature extraction from electrical signals with the modeling of long-term temporal dependencies. Through coordinated parameter optimization— including temporal windowing, feature normalization, and class balancing—the optimized model achieved high validation performance, significantly improving upon baseline diagnostic methods.This work bridges the gap between accurate physics-based simulation and the data requirements of modern deep learning, providing a scalable and reliable solution for predictive maintenance in high-performance electric drives.