Multi-scale wavelet feature fusion with transformer for fault diagnosis of 25 Hz phase-sensitive track circuit

Aiming at the problems of long-range dependency modeling difficulty, weak fault feature extraction challenge, and severe class imbalance with limited minority fault samples in long time-series data of 25 Hz phase-sensitive track circuits, this paper proposes a fault diagnosis model called WT-Transformer by integrating discrete wavelet transform and Transformer encoder. Firstly, the original track circuit signal is decomposed by multi-scale discrete wavelet transform to extract global long-term trend features and local abrupt change features caused by faults. A soft-threshold denoising method based on the Minimax rule is adopted to suppress noise interference while retaining critical fault information. Secondly, the wavelet-enhanced signal and the original signal are combined by sample concatenation to enrich feature diversity and improve the identifiability of minority-class faults. Sine-cosine position encoding is integrated to provide temporal structure of the track circuit signal. Finally, a Transformer encoder with multi-head self-attention, feedforward network, and residual connection is constructed to capture long-range temporal dependencies and enhance feature representation ability. Experimental results on a simulated fault dataset show that the proposed model achieves superior performance in track circuit fault diagnosis, especially in the classification of rare faults with few samples. The effectiveness of wavelet transformation and time-frequency feature enhancement is verified, which provides a feasible and effective solution for intelligent fault diagnosis of long sequence data in track circuits.