The present deep learning (DL)-based detection of crack damage of shield tunnels through images can only provide part of geometric information of the crack damage, which prevents the comprehensive understanding of the mechanical behavior of the cracked tunnel linings. This study bridges the gap between the geometric information identified by DL models and the transformation law of mechanical behavior of tunnel linings revealed by numerical methods. A mask-region-based hybrid attention convolutional neural network (Mask R-HACNN) and mesoscale cohesive numerical model are developed to obtain the geometric information of crack damage and investigate the structural response of segments and joints with crack damage, respectively. The numerical simulation results demonstrate that the relationship between in-depth crack length and crack opening displacement and the relationship between the degradation stiffness and in-depth crack length are independent of the loading path. Based on these findings, prediction formulas are established to interpret the relationships between the stiffness degradation coefficients, flexural stiffness, in-depth crack length, joint openings, and crack opening displacement. Then, the geometric information of crack damage obtained by the Mask R-HACNN is provided to the mesoscale cohesive numerical model to estimate the in-depth crack length, flexural stiffness, and stiffness degradation coefficient of cracked segments. Results show that the deep learning informed-mesoscale cohesive numerical model is satisfying in understanding the mechanical behavior of segments and joints during crack damage evolution.
Read full abstract