Fault Diagnosis for Electric Drives Using Averagely Pooled and Downsampled Data Fusion on Embedded Systems

This study introduces a data-driven approach for diagnosing faults in electric drive systems on embedded systems. The method is designed to diagnose current Hall sensor offset, scale errors, and power switch open faults without the need for additional sensors or external computing resources. The proposed approach addresses three key points. First, it identifies and suggests data types in electric drive systems that are rich in fault-related information. Additionally, it uses statistical analysis to propose an effective fusion method for the input data types to create a diagnostic model among the various candidates. Second, a downsampling method utilizing average pooling (AP) is introduced to effectively sample data from the large volume of raw data processed by the digital signal processor (DSP), preserving essential information while reducing data size. Finally, a diagnostic scheme employing a neural network (NN)-based classifier is proposed to accurately diagnose faults using the downsampled data. The proposed method’s effectiveness, from sampling to diagnosis, is demonstrated through both analytical and statistical validation. Furthermore, real-time diagnostic results are presented using the TMS320F28379S DSP.