Dual-Parallel FCN With Attention for Fault Diagnosis in Urban Rail Transit Energy Router

Energy routers (ER) significantly reduce traction energy consumption when integrated with traditional rail power systems. However, the complex multiport structure and dynamic fluctuations in ER systems pose substantial challenges to traditional model-driven fault diagnosis methods, which may fail under such conditions due to modeling inaccuracies or limited adaptability. To address this, a data-driven fault detection and diagnosis method is proposed. By modeling fault mechanisms and combining time-frequency domain features, a deep learning framework based on a dual-channel fully convolutional network with a squeeze and excitation mechanism is developed. Time and frequency domain features, derived from raw signals and their fast fourier transform (FFT) transforms, are processed in parallel using two convolutional branches: one captures temporal variations via a deep convolutional neural network (CNN), and the other extracts spectral patterns using a lightweight CNN. The method enables accurate early detection and fault classification. Experiments show it surpasses traditional methods in accuracy and robustness, supporting the safe operation of ER in urban rail transit.