Ensemble-Based Anomaly Detection with Comprehensive Feature Distances for Induction Motors

This study proposes a model for detecting anomalies in induction motors using audio signals. Compared to existing anomaly detection methods based on deep generative models, the model is presented in two main aspects. First, the proposed model consists of two generative adversarial network (GAN) modules, each processing the two types of mel-spectrograms with distinct time and frequency resolutions. These modules are trained to reconstruct input data effectively, enabling each module to extract the high-resolution time and frequency domain features. Secondly, the ensemble approach is proposed for calculating the anomaly score and loss function. This approach ensembles all the reconstruction losses and Comprehensive Feature Distances (CFDs) obtained from the two GAN modules. The CFD is calculated as the Maximum Mean Discrepancy (MMD) between the outputs of the encoder and decoder layers of the generator. The effectiveness of the proposed scheme is validated through experiments conducted with 0.75 kW induction motors.