Rapid Detection and Classification of Subway Tunnel Cavities Using Moving Train Vibration and Advanced Network Training

Detecting cavity‐induced damage in subway tunnel structures remains a significant challenge due to the difficulty of full‐time and full‐area monitoring using traditional methods. This paper proposes a rapid approach for identifying and assessing the severity of cavity damage based on vibrations from moving trains. The methodology involves an advanced network training algorithm combining empirical mode decomposition (EMD), Hilbert transform, bidirectional long short−term memory (Bi‐LSTM), and self‐attention mechanisms. This novel combination enables the extraction of meaningful features from complex vibration data, facilitating precise damage detection. To validate the proposed method, a series of model tests with varying cavity sizes and locations was conducted. The three‐axis acceleration responses of a model train, recorded using a SmartRock sensor, were analyzed to evaluate the performance of the approach. Comparative analysis demonstrates the efficiency of the method, with the network achieving over 95% accuracy and a kappa coefficient above 0.95 for all classifiers in the validation set. The results confirm the potential of the proposed algorithm for rapid and accurate detection of cavity location and damage severity.