Detecting Anomalies Inside Rolling Stocks using Audio Streams and Deep Learning

This paper addresses the enhancement of passenger safety and comfort in public transport by automatically processing the audio streams from an anomaly detection system.The importance of anomaly detection has recently captured the attention of numerous researchers.Consequently, automated techniques, primarily based on artificial neural networks, are increasingly being adopted.This expansion is largely driven by the availability of large datasets and the use of graphics processing units, which facilitate the training of such models.Thus, these technologies have become the foundation for models that meet the railway industry's needs in ensuring the safety of passengers in train coaches.Although these models are promising and deliver high performance, they do so at the expense of significant system complexity and high computational costs.The results obtained confirm the benefits of using audio signals to detect unusual events and highlight some challenges in defining the appropriate set of model hyperparameters.