Physics-Informed Hybrid Deep Learning for Damage Detection in Metro Tunnel Structures Under Moving Train Loads

In this study, a novel physics-informed hybrid deep learning (PI-HDL) algorithm is proposed for intelligent condition assessment of metro tunnel structures. The algorithm integrates state-of-the-art deep learning models, including auto-encoder and convolutional neural network (CNN) models. Physics-inspired constraints derived from structural dynamics have been embedded into the loss functions of these models. Via an effective learning, merging, and training methodology, the designed model reconstructs high-level features of 1D time-domain vibration response signals in an unsupervised manner. Then, these physics-consistent high-level features are fed into a supervised 2D CNN trained with a composite loss function that incorporates both data fidelity and physics-based residuals derived from the stiffness–frequency relationship and moving load characteristics. The developed methodology is evaluated using vibration signals collected numerically from a concrete metro tunnel structure under various moving train loads and damage scenarios. The proposed PI-HDL framework embeds the underlying physics of wave propagation and structural dynamics directly into the learning process. This enables accurate identification of predefined damage conditions and stiffness reduction quantification within a closed set of ten structural scenarios, even with limited training data and noisy measurements. Comparisons prove that the PI-HDL algorithm achieves superior closed-set classification performance compared to purely data-driven HDL approaches.