Hierarchical fault diagnosis of train communication networks based on cross-dimensional information fusion and mixture-of-head attention mechanism

The Ethernet Train Communication Network (ETCN) is a vital platform for information exchange in train operations, where anomaly detection is essential for ensuring safety and efficiency. However, challenges such as labor-intensive data collection, limited annotations, and the shortcomings of existing fault diagnosis methods, including insufficient information utilization, limited flexibility, and poor generalization, hinder progress. To address these issues, we propose a hierarchical data collection method for the ETCN upper network (data link layer and above) and a cross-dimensional information fusion fault diagnosis model based on a temporal convolutional network and a mixture-of-head mechanism (TCN-MoH). The proposed method improves data collection efficiency through hierarchical indicators that comprehensively reflect network health. Meanwhile, TCN-MoH captures both long- and short-term dependencies as well as deep feature relationships in temporal and feature dimensions, enhancing its fault classification performance. Validation on two simulation datasets demonstrated classification accuracies of 99.82 % and 99.59 % for direct and indirect fault datasets, respectively, outperforming existing approaches. These results confirm the effectiveness of the proposed methods in ETCN fault diagnosis.