Intelligent Prediction of Bearing Capacity Degradation in Tunnel Linings via Pseudo-Crack Simulation and Deep Learning

Abstract

Cracks in tunnel linings induce bearing capacity degradation, making the rapid prediction of the damage ratio critical for structural safety assessment. However, quantifying the intrinsic mechanical performance solely from apparent surface cracks remains a challenge. To bridge this gap, this study proposes a physics-informed deep learning framework that maps visual crack features directly to the bearing capacity damage ratio. First, a high-fidelity numerical simulation system is established using the Pseudo-Crack Method implemented on the multi-scale thermodynamic platform. This approach avoids the mesh dependency of traditional fracture mechanics and is rigorously validated against existing physical model tests of lining structures in terms of crack morphology and load-displacement responses. Subsequently, a standardized synthetic dataset is constructed by distilling topological features from in-service hydraulic tunnels and applying a color-coded width visualization strategy. By conducting comparative training across eight state-of-the-art deep learning architectures, the ConvNeXtV1 model is identified as the optimal regressor, achieving a coefficient of determination of 0.89 on the test set. The proposed method effectively acts as a real-time "digital surrogate" for time-consuming non-linear finite element analysis, providing a mechanism-based, efficient solution for structural health monitoring of tunnel infrastructure.