Evaluating railway noise sources using distributed microphone array and graph neural networks

Identifying and evaluating railway noise sources is beneficial for establishing more targeted suppression and mitigation strategies. This paper proposes a novel approach to evaluate dominant railway noise sources for all bogie regions, by using an array of microphones and a physics-informed graph neural networks (GNN) model. The GNN model encodes not only the acoustic data collected by the microphone array but also the spatial relationship between noise sources and microphones, and it also considers the physical background of sound in terms of Doppler effect and acoustic attenuation. In-situ measurements of railway noises have been taken on an operating metro line to train and validate the GNN model which exhibits satisfactory performance in evaluating the noise sources of interest. The proposed approach enables a more flexible and cost-effective microphone array implementation strategy in comparison to the commonly adopted commercialized acoustic beamforming system.