A dual-branch time-domain network for rail defect signal separation with gradient-guided time–frequency attention and triplet constraints

With the advancement of transportation and economic development, rail transportation has become a critical mode of transit, and rail health monitoring is essential for ensuring its operational safety. However, traditional signal processing methods often fail in complex vibration and noise environments due to spectrum overlap and non-stationary characteristics. In this context, we propose a novel dual-branch deep network architecture, time-domain acoustic signal separation network (TDASS), which integrates a time–frequency-aware attention mechanism driven by local gradients to effectively distinguish between low-frequency periodic noise and high-frequency impulsive defect events for blind source separation of rail defect signals. The process by which neural networks learn to ‘listen’ and ‘separate’ is pivotal for both model optimization and the interpretability of the learning process. To further enhance the network’s discriminative ability and improve the transparency of feature extraction, a triplet loss strategy was introduced, enforcing discriminative representation learning by modeling the anchor-positive-negative relationship between predicted and ground-truth signals. Extensive experiments on real-world acoustic emission datasets demonstrate that our method achieves superior performance in terms of scale-invariant signal-to-noise ratio (SI-SNR) and waveform similarity coefficients, with an average similarity of 0.9719 and an average SI-SNR of 20.18 dB in the rail defect separation task. These results validate the efficacy of combining gradient-aware frequency attention with structural task constraints for robust and accurate defect signal extraction.