Automatic prediction of railway ballast layer thickness using graph neural network based on GPR reflection signals

Abstract

Ground-Penetrating Radar (GPR) is widely employed for detecting the thickness of railway ballast layer. However, the complexity of the GPR data often requires manual interpretation by experts, which limits the efficiency of large-scale inspections. To address this challenge, this paper proposes a graph neural network-based method for automatic ballast layer thickness prediction. This method leverages Temporal Convolutional Networks (TCNs) to extract temporal patterns from the GPR A-scans and employs Graph Convolutional Networks (GCNs) with a self-adaptive adjacency matrix to dynamically learn and refine the spatial correlations across multiple A-scans. The proposed method was validated using a combined dataset of simulated and field data, and further tested through on-site applications. Experimental results show that the method outperforms four baseline models in prediction accuracy while maintaining high inference efficiency. In on-site tests, the average absolute prediction errors of Two-Way Travel Time (TWTT) and thickness were 0.25 % and 3.06 %, respectively. These findings demonstrate the effectiveness, efficiency, and potential scalability of the proposed method for railway ballast layer thickness detection.