A Semi-Supervised Variational Autoencoder for Fault Detection of Low-Severity Inter-Turn Short-Circuit in PMSMs

Maintenance has been critical in energy and marine systems, ensuring safety, reliability and efficiency. Fault detection, a cornerstone of prognostics and health management systems, faces challenges due to limited faulty data. This study proposes a semi-supervised Variational Autoencoder (VAE) with Long Short-Term Memory Networks for fault detection. The method is demonstrated via an example of low-severity inter-turn short-circuit faults in permanent magnet synchronous machines. Trained only on healthy data, the model can generate anomaly scores based on the VAE's reconstruction error. The model is tested on time-series current data under varying conditions, achieving high detection accuracy. Comparative analysis shows its superiority over traditional methods such as One-Class Support Vector Machine, Isolation Forest and Autoencoder.