Dynamics-driven signal enhancement and local-global learning for track interlayer void diagnosis

Abstract

The accurate identification of interlayer gaps within the Cement Asphalt Mortar (CAM) layer of CRTS II ballastless tracks is consistently hampered by intense high-speed background vibrations and the inherent non-stationarity of fault-induced signals. Conventional data-driven approaches often struggle to reconcile these signal-masking issues with the sparsity of critical failure features, leading to suboptimal robustness in field applications. To address these challenges, this study introduces Physics-Driven Time-Frequency Modal-Synergy Network (PTFM-Net), a novel diagnostic framework integrating physics-based data generation, dynamics-guided signal enhancement, and local-global feature learning. A validated vehicle-track dynamics framework is established for multi-condition void data generation. Specifically, a dynamics-guided signal enhancement mechanism integrating IWOA-VMD and physics-enhanced continuous wavelet transform (PE-CWT) is developed to enhance void-sensitive transient features and suppress speed-dependent background vibration. To capture both fine-grained impact morphologies and non-local energy distributions, PTFM-Net employs a collaborative local-global fusion architecture. Within this framework, an SE-ResNet18 module serves as a local feature encoder to isolate the specific transient signatures of interlayer debonding, while a parallel Transformer branch models global dependencies to track fault feature evolution across long-range time-frequency domains. Experimental results show that PTFM-Net achieves a diagnostic accuracy of 95.2%, outperforming existing mainstream approaches by 4.1%–8.7%. These findings validate the effectiveness of PTFM-Net in challenging environments, offering a robust and practical solution for the intelligent operation and maintenance of high-speed railway systems.