Multi-Fault Classification of PMSM with Frequency-Domain Feature Analysis using Supervised Machine Learning

Abstract

This study presents a comparative data-driven method to enhance the robustness of low-severity, multi-fault classification in noisy dynamic conditions for permanent magnet synchronous machines (PMSM) with supervised machine learning models Extra Trees (ET) Support Vector Machine (SVM). Utilizing optimization of hyperparameters, evaluating impact of frequency-domain (FD) features and feature importance on the accuracy, prediction confidence, training time, prediction time, energy requirements and model size. Where the training dataset includes Signal-to-Noise (SNR) ratio of both 120 dB and 10 dB with inter-turn short circuit fault of 2.2% in one phase, demagnetization fault of one north pole magnet by 30%, and a combination of these two faults. The PMSM is operated at both constant- and varying speeds, ranging from 750 RPM to 2100 RPM. Extra Trees algorithm achieved cross-validated average accuracy of 93.0%, training time 2.732 seconds, prediction time per sample 0.122 milliseconds, with energy consumption 0.017 Wh, and model size of 317 MB. Compared to Support Vector Machine with cross-validated average accuracy of 77.48%, training time 223.4 seconds, prediction time per sample 3.678 milliseconds, with energy consumption 8.541 Wh and model size 67 of MB. While the ET model’s average accuracy is higher than the presented SVM model, it is lower than the state-of-the-art solutions with more complex feature extraction. However, due to the noisy dataset and inclusion of FD features, its increased robustness in dynamic operation can be affirmed by the higher confidence in its predictions.