Leveraging Deep Transfer Learning and Time-Aware Convolutional Transformers for Stator Winding Fault Diagnosis in Electric Motors: A Digital Twin Approach

Electric motors are pivotal yet vulnerable components in electric vehicles (EVs), with stator winding faults, particularly inter-turn faults, being among the most frequent and hazardous. Early detection of these faults is crucial for predictive maintenance and ensuring the reliability and safety of EVs. This study introduces a Time-aware Convolutional Transformer (TaCT) model that integrates transformer architecture with convolutional layers and a novel time-series specific positional encoding to enhance fault diagnosis performance by capturing long-range dependencies in time-series data, which are critical for detecting subtle, developing faults. A novel deep transfer learning approach, integrated within a digital twin framework, which creates a virtual replica of the physical motor, is proposed to improve fault diagnosis efficiency and generalization by treating data from time-varying conditions as a continuous domain shift. Four distinct transfer learning methodologies were employed to update and refine digital twin models for fault diagnosis. The TaCT model demonstrated markedly superior performance, maintaining an accuracy above 0.95 across all update steps, outperforming other deep learning models. Notably, TaCT's architecture proved particularly effective for short-circuit fault detection, as confirmed by a Conover test showing it achieved the highest average rank. The digital twin transfer learning approach mitigated the issue of catastrophic forgetting, which occurs when a model loses previously acquired knowledge upon learning new information, and significantly improved model performance over multiple update steps. This research highlights the advantages of integrating advanced deep learning models with digital twin frameworks and transfer learning techniques, offering substantial improvements in EV motors' predictive maintenance and fault diagnosis.