Unsupervised Anomaly Detection for Electric Drives Based on Variational Auto-Encoder

Machine learning has been increasingly applied to electric drive systems especially in the field of fault diagnosis. Among various machine learning methods, previous studies mostly adopted supervised learning for fault diagnosis of electric drives. However, the supervised learning-based machine learning model has a limitation in that it can only diagnose failures that have been trained. Moreover, in the real world, a sufficient amount of abnormal data is difficult to obtain for training the model. Therefore, this paper proposes an unsupervised learning-based anomaly detection model for electric drive. The proposed model uses reconstruction errors of input data on a Variational Auto-Encoder (VAE) when detecting anomalies. Keras with TensorFlow backend is accompanied for training and statistical evaluation of the model. The data for training and testing is acquired through a 200W servo-motor experimental testbed. Statistical evaluations demonstrate the effectiveness of the proposed anomaly detection model. Furthermore, the anomaly detection algorithm is also verified on the TMS320F28379S digital signal processor (DSP) in real-time operation.