Heterogeneous graph attention network for rail fastener looseness detection using distributed acoustic sensing and accelerometer data fusion

Abstract

Ensuring rail fasteners' integrity is crucial for railway safety. Traditional methods for detecting loosened fasteners are laborious and economically inefficient. This paper introduces FusionHGAT, an attention-enhanced heterogeneous Graph Neural Network (GNN), designed for precise, automated detection of rail fastener looseness by fusing data from Distributed Acoustic Sensing (DAS) and accelerometers. The method collects sensor data during rail track excitations, constructs a graph based on spatial relationships, and implements FusionHGAT through a three-step procedure: feature extraction with 1D-Convolution Neural Networks, feature embedding via a Transformer module, and feature fusion using Graph Attention Network layers. Experimental results demonstrate FusionHGAT's outstanding performance, achieving 100 % accuracy and validating the model's superiority. Building on the results presented in this work, our graph-based methodology enhances the detection of fastener looseness through spatial-temporal data fusion, highlighting its potential for future real-time railway infrastructure monitoring.