Detection of ballastless track interlayer gap based on vehicle's multivariate dynamic response and deep learning

To adapt to the rapid detection of interlayer damage in ballastless track structures of high-speed railways, a cement asphalt mortar (CAM) gap localization and damage degree classification scheme based on multivariate data fusion and deep learning is proposed. Based on vertical axle box acceleration (VABA) and vertical wheel-rail force (VWRF) data, the variation patterns of multi-dynamic response data of vehicles under the edge type and internal type interlayer gaps are analyzed. An improved ensemble local mean decomposition algorithm (IELMD) is designed to achieve data denoising and preliminary enhancement preprocessing of weak damage features for VABA and VWRF under interlayer gaps. The measured and simulated VABA and VWRF constituted multiple dynamic response data sets for interlayer gap detection. An interlayer gap detection model: DTA-Tcnformer is constructed, integrating a dual temporal convolutional network with an attention mechanism, transformer architecture, and gate control mechanism. Multiple dynamic response data are input into the model to locate and classify 24 cases of the two interlayer gap types. The influence of damage cases and vehicle speeds on the location effect is analyzed. The learning ability of the proposed model on the gap features is visualized. The misidentification and omission of the gap cases are calculated, and the comprehensive recognition accuracy is 92.48 %. The recognition performance of the proposed model is compared and verified.