Towards Rail Noise Identification and Localization Based on Deep Learning

Noise and vibration remain a major reason for sustainable railways. Furthermore, noise and vibration in railways are often related to defects or faults. It is essential to understand the source of noise and vibration to identify the issue behind it. However, the complexity of noise analysis is embodied with acoustic signals with overlaps, as is usually the case in rail systems with the urban environment, while seldom discussed in previous studies. Leveraging cutting-edge acoustic processing, this study proposes a novel method incorporating trendy contrastive learning and neural network structure to identify overlapping noises and capture their directions. The model is validated on both the public sound and the synthetic rail datasets. Model effects on various sound conditions are then discussed. Results show that the proposed method offers a new perspective to disentangle intertwined noises, shedding light on mitigation measures and thus a sustainable rail system.