Diagnosis of Demagnetization Faults in PMSGs Based on Performance Evaluation of Machine Learning Algorithms Through Flux and Current Signals

This study proposed an effective machine learning (ML)-based fault diagnosis method for demagnetization faults, including “healthy, 30% unipolar demagnetization, 50% multimagnet demagnetization, 100% adjacent pole demagnetization, and 40% uniform demagnetization,” in a 25-kW permanent magnet synchronous generator (PMSG) that considers algorithm accuracy and signal processing technique efficiency, top informative features, and training time of the ML algorithm. Our research involves developing a 3-D finite element PMSG model in ANSYS Maxwell and conducting an industrial experiment to capture stator current and flux data. We use advanced feature extraction techniques such as matching pursuit (MP) and discrete wavelet transform (DWT), assessing their accuracy using ensemble, SVM, and KNN algorithms. In addition, we optimize and rank features using ReliefF, mRMR, Chi2, ANOVA, and Kruskal-Wallis filters, evaluating their impact on ML algorithms. The investigation extends to the evaluation of eight ML classifiers—ensemble, SVM, KNN, neural network, kernel, discriminant analysis, and naive Bayes—along with 31 subclassifiers to identify the most effective output signal, signal processing technique, informative features, suitable filters, and the optimal ML algorithm for diagnosing demagnetization faults in PMSGs. Our findings reveal that using flux signals in conjunction with the DWT, alongside the ANOVA feature ranking filter and fine Gaussian SVM algorithm, achieves 100% fault classification accuracy with a training time of 0.961 s. This approach proves highly effective for the early detection of demagnetization faults in PMSGs.