Hybrid physics-data-driven diagnosis approach for transverse mechanical state of segmental tunnels considering discontinuous joint movements

Forward prediction and inverse identification of the transverse mechanical state are essential for the structural health assessment of jointed segmental tunnels. Existing data-driven, physics-based, and hybrid approaches have limitations in addressing discontinuous joint movements and the inversion of external loads with unknown distributions. To this end, a hybrid physics-data-driven approach is proposed for intelligent diagnosis of the transverse mechanical state of jointed segmental tunnels, with discontinuous joint movements explicitly considered. In this method, characteristic functions representing different joint movement behaviors are encoded into the network inputs to generate discontinuous displacement mappings. Additionally, a system of first-order governing equations derived from Euler curved beam theory and tensionless Winkler foundation theory is employed to describe the mechanical behavior of tunnel lining segments and construct physics-based loss terms, thereby mitigating the accuracy degradation and training difficulties associated with higher-order formulations. A numerical example and an engineering case study are then employed to verify the accuracy and effectiveness of the proposed method by benchmarking its results against analytical solutions, existing approaches, and measured data. The results demonstrate that the proposed approach can effectively characterize joint rotation and relative displacement between segments. For forward problems, the predicted results show excellent agreement with analytical solutions. For inverse problems, when the joint stiffnesses and subgrade reaction coefficients are known, the coefficient of determination ( R 2 ) values exceed 0.93 for external load inversion and 0.98 for mechanical field reconstruction. Even with noise-free monitoring data from 13 points or monitoring data from 24 points with 20% Gaussian noise, the R 2 values for inverted external loads remain above 0.5 and those for reconstructed mechanical fields exceed 0.8, demonstrating robustness against sparse and noisy monitoring conditions. When joint stiffnesses and subgrade reaction coefficients are simultaneously unknown, the proposed approach can still achieve accurate reconstruction of key mechanical fields, with R 2 values approaching 0.99, although parameter coupling may lead to non-unique inversion solutions.