Classification of Rotor Imbalance in Trains Using Airborne Sound With Real-World Data

In this paper, we address the task of classifying rotor imbalances in induction motors of trains by analysing airborne sound data. We propose the use of a residual-based Convolutional Neural Network (CNN) architecture that stands out for its small number of trainable parameters. The proposed model is evaluated with real-world data that was gathered from state-of-the-art metro trains during regular operation. The experiments show that with the proposed approach rotor imbalances can be reliably classified even when the model is confronted with new damages that were not included in training.