A physics-aware time-frequency autoencoder for anomaly detection in heavy-haul railway communications

Reliable anomaly detection in railway communication monitoring is an important engineering requirement for the synchronized control of long heavy-haul trains. Although existing deep time-series anomaly detection methods have demonstrated promising performance in reconstruction and representation learning, three challenges remain: (1) static multiscale fusion cannot adaptively balance local morphology and global dependencies; (2) phase topology is underused, while powerful reconstructors may overgeneralize to abnormal patterns; and (3) independently predicted amplitude and phase may be mismatched, whereas an analytical inverse Fourier transform propagates spectral errors to the time domain without learnable correction. To address these challenges, we propose a physics-aware phase diffusion multiscale time-frequency autoencoder (Phy-TFMAE) with a joint contrastive-adversarial and reconstruction architecture. The model first uses a memory-gated feature fusion mechanism to balance local high frequency distortions and global semantic representations. It then introduces an amplitude-guided phase diffusion branch to reconstruct phase topology on a circular manifold, capturing phase-sensitive structural variations overlooked by amplitude-based reconstruction. Finally, a neural-enhanced inverse fast Fourier transform module performs physics-consistent reconstruction in the complex domain. Phy-TFMAE mitigates overgeneralization through mutual regularization with the contrastive-adversarial objective. Experiments on five public datasets from diverse application domains and a real-world Long-Term Evolution for Railway operational dataset from the Shuohuang Railway show that Phy-TFMAE achieves the highest F1-score among the compared methods on all six datasets, reaching an average F1-score of 96.73%, with an average improvement of 0.79 percentage points over the strongest baseline.