Interpretable Deep Learning for Railroad Health Monitoring Using Accelerometers and Distributed Acoustic Sensing

Abstract

The health monitoring of long-distance infrastructures, such as railways, poses significant challenges due to extensive spatial coverage, operational complexity, and varying environmental conditions. Traditional methods are often limited by high costs and insufficient spatial resolution, making them unsuitable for large-scale and dynamic scenarios. Distributed sensing systems offer a promising alternative, enabling real-time, cost-effective monitoring over long distances. This study proposes an integrated system that combines accelerometers with Distributed Acoustic Sensing (DAS) to address these challenges. The system has been deployed on a subway line in Singapore for continuous monitoring under operational conditions. To overcome issues with DAS signal accuracy and physical interpretability, an interpretable deep learning framework is developed, fusing DAS and accelerometer data. This framework incorporates state-of-the-art interpretability techniques to enhance prediction accuracy and provide insight into model decision-making. As a case study, the proposed method is applied to detect rail fastener failures—common defects in railway systems. Field experiments using hammer impact excitation were conducted, and the resulting vibration responses were recorded by the integrated sensing system. These data were used to train and validate the proposed model. Results demonstrate that the framework outperforms traditional DAS-only approaches in both accuracy and generalization, leveraging interpretable features that align with known rail dynamic behavior. These findings underscore the method’s potential for scalable, reliable, and transparent structural health monitoring in railway infrastructure.