Autoencoder-Based Unsupervised Anomaly Detection in Induction Motors Diagnosis

The autoencoder is widely used in anomaly detection. The anomaly detection-based approach assumes a reconstruction error for the anomalous samples higher than the normal ones and can identify abnormal data from normal data. This paper compares two deep autoencoders to earlier anomaly detection in electric motors. Our approach is based on reconstructing the stator current temporal signal for rotor broken bars detection. To evaluate the performance of the Dense and the conv1D autoencoder, we have used an experimental database acquired under varied load conditions and with different types of rotor faults. The results obtained show a promising induction momotornomaly detection approach.