Robust Multiple-Fault Diagnosis of PMSMs in Dynamic Operations Under Imbalanced Datasets

Data-driven models for multiple-fault classifiers of permanent magnet synchronous machines (PMSMs) in electric power trains require historical data at faulty cases in a wide range of operations. However, data in healthy or steady-state operations are much more than the data in the faulty cases, resulting in imbalanced datasets and rendering challenges of expanding the robustness of the data-driven models in dynamic operations. To address these issues, this article proposes a robust two-stage learning scheme for detecting and classifying multiple incipient faults in PMSMs, namely, interturn short circuit (ITSC) local demagnetization (LD) and mixed faults (MFs), under dynamic operations. The first stage focuses on enhancing the effectiveness of anomaly detector in dealing with imbalanced datasets at low-severity faults by integrating prior domain knowledge into anomaly detectors and reducing false alarms using an extreme gradient boosting machine (XGBoost). In the second stage, a training algorithm based on a convolutional neural network (CNN) is introduced to classify multiple faults by studying the appearance of low-severity faults. The performance of the proposed anomaly detector is compared to conventional one-class models, including random forest, local outlier factor, and support vector machine. The accuracy and robustness of the suggested scheme are further evaluated by benchmarking pretrained classifiers such as MobileNetv2, ResNet50, and VGG16, using an imbalanced dataset at different operating conditions from an in-house test setup. A comparative study with a traditional CNN is conducted to explain how the model provides the prediction.