Data-Driven Onboard Inter-Turn Short Circuit Fault Diagnosis for Electric Vehicles by Using Real-Time Simulation Environment

Various fault detection methods, particularly focused on onboard Condition-Based Monitoring (CBM) in Electrical Machines and Drives (EMDs), face limitations such as sensitivity to load variations, slow fault detection, and the absence of fully automated solutions. AI and Data-Driven methods offer flexible alternatives, utilizing historical data for pattern and anomaly identification. Among Electrical Signature Analysis techniques for electrical motor diagnostics, the Space Vector Theory (SVT) is extensively used, while Park’s Vector based diagnostic solutions lack real-time Inter-Turn Short Circuit (ITSC) fault severity assessment, with available techniques often limited to binary classifiers. Implementing AI with SVT for real-time Electric Vehicle (EV) use is underdeveloped, hindered by data scarcity and diverse dataset collection challenges. Real-time simulation, accurate fault modeling, and hardware limitations pose challenges, especially for embedding AI models into processors. To achieve intelligent onboard diagnosis for ITSC fault severity in this paper, a multi-modal approach model is proposed, employing MobileNetV2 to classify Park’s Vector trajectories based on the fault features related to the number of shorted turns. Performance assessments encompass both the standard MobileNetV2 and the proposed multi-modal approach model across various fault severity levels. Furthermore, to address the challenge of limited data availability, an accelerated real-time AI development environment is designed using an FPGA to generate synthetic fault pattern datasets, aligning with the standards of the Electric Vehicle industry. For modeling PMSM with ITSC faults, a fault circuit model is employed. The dataset of 900 Park’s Vector trajectory images is automatically generated by varying the torque request from 10 to 100 Nm with a 10 Nm resolution. At each torque operating point, the motor currents are recorded by adjusting the number of shorted turns. Simulation results confirm the outstanding performance of MobileNetV2 in binary classification, achieving an accuracy of 99.26 %. In case of 5-class ITSC fault severity classification, the prediction accuracy reaches only 72.55 %. The here proposed multi-modal MobileNetV2 model excels, achieving a remarkable accuracy of 99.163 % in the 3-class fault severity classification and 84.907 % in the 5-class classification. These results support the superiority of the proposed multi-modal MobileNetV2 model, which is trained on the generated rich dataset. It outperforms existing Park’s Vector Analysis based ITSC fault detection methods, particularly in early ITSC fault detection as it can detect faults from 6 shorted turns. Additionally, it allows for online fault severity assessment during transient operation and meets stringent requirements for onboard applications. Altogether, the results of investigations prove the presence and extractability of fine detail information in Park’s Vector trajectories, for assessing ITSC fault severity. This contributes to a deeper understanding and analysis of faults in electrical motors through the use of Park’s Vector trajectories.