Seismic resilience assessment of shield tunnels under contact loss defects using a hybrid neural network model driven by machine learning and numerical simulation

Abstract

Contact loss defects (CLD) have a significant impact on the seismic performance of shield tunnels with segmental joints. Accurately evaluating the seismic resilience of tunnels under the influence of CLD is crucial for ensuring the safety of urban rail transit systems. In this study, a comprehensive dataset considering the effects of CLD was constructed using actual ground penetrating radar (GPR) detection data, combined with Monte Carlo sampling and finite element simulations. A machine learning and numerical simulation-driven MoE-BP (Mixture of Experts–Backpropagation) neural network model was proposed to efficiently and accurately assess the seismic resilience of shield tunnels with segments joints. Based on the assessment results, targeted recommendations for defect repair strategies were developed. The results indicate that the shear wave velocity of the stratum and the circumferential angle of the CLD are the most critical factors influencing the seismic performance of tunnel; when the defects are located near segmental joints, which pose the most adverse impact on resilience. The proposed seismic resilience assessment indicators—combining damage variables and the inner contour area of the tunnel—effectively capture the coupling effects between CLD and segmental joints. Compared with a conventional BP neural network, the MoE-BP model achieves significantly higher accuracy and computational efficiency, with a 70% reduction in mean squared error and a 13% increase in correlation coefficient after 1000 training epochs.