Assessment of Machine Learning Techniques for Identifying Demagnetization Stress in PMSG Using Flux and Current Signals

This study focuses on the detection of faults in wind turbine operations to ensure reliability, safety, and cost efficiency. It introduces a machine learning-based diagnostic approach for identifying various faults, including eccentricity faults, stator short circuits, and both damaged and aging magnets in a 25 kW Permanent Magnet Synchronous Generator (PMSG). A 2D simulation model utilizing the Finite Element Method (FEM) to analyze the output monitoring signals and assess the performance of the machine learning algorithms. Moreover, the study uses stator current and leakage flux signals, applying the Matching Pursuit (MP) for feature extraction. Furthermore, the performance of 17 sub-classifiers across three core machine learning classifiers-Support Vector Machine (SVM), K-Nearest Neighbour (KNN), and Ensemble techniques-is evaluated for diagnosing demagnetization faults in PMSG. The results reveal that when leakage flux signals are used as the output, seven out of the seventeen classifiers achieved 100% accuracy in classifying all faults in PMSG. In contrast, only four classifiers reached the same level of accuracy using current signals. This indicates that leakage flux signals provide a more reliable basis for diagnosing multiple demagnetization faults in PMSG, making any of the seven successful machine learning sub-classifiers a viable choice for this task.