A multilevel track defects assessment framework based on vehicle body vibration

High-frequency detection of track defects is crucial for accurate track condition assessment and system safety. Onboard vibration data collection devices can significantly increase detection density without additional costs. However, defect assessment based on this is significantly challenging, including the spatial heterogeneity of track parameters, distribution mismatch between vibration data and defect labels, and variability in vibration responses across different defects. This study proposes a multilevel track defect assessment framework based on vehicle body vibration. The correlation intensity between vibrations and heterogeneity factors was analyzed, and a correlation-view spectral clustering algorithm was designed to achieve effective data set partitioning. A spectral-normalized neural Gaussian process-based adaptive-threshold self-training method (SNGP-ASM) was developed to generate high-quality pseudo-labels and generate a fully labeled data set. An attention-guided multitask cascaded convolutional neural network (CNN) was constructed to progressively assess track defects using channel-wise attentions and a cross-hierarchical attention guidance module. Validations on multiple Chinese metro lines demonstrated that the framework achieved a high performance in training and testing for most defect assessment tasks within lines, and the trained model can effectively adapt to new lines with only lightweight fine-tuning. Moreover, the framework maintained a high computational efficiency, enabling high-frequency track condition monitoring in practical deployment scenarios.