Probability analysis on tunnels in heterogeneous strata based on borehole data-driven conditional random fields and convolutional neural network

Abstract

Tunnels in heterogeneous strata always encounter spatially varied geological formations, causing asymmetric responses and localized failure in the supporting structure. The homogeneity assumption for surrounding strata, commonly adopted in tunnel design and construction, will neglect the inherent spatial uncertainty of rock mass and lead to the overestimation in tunnel bearing capacity. The conventional stochastic calculations for analyzing tunnel performance in heterogeneous strata also fail to reflect the statistical asymmetry in mechanical behaviors of supporting structure. With the application of mechanized equipment with built-in sensors in drilling and blasting construction, rock parameters at borehole locations can be promptly derived through the drilling data. This systematic on-site monitoring necessitates a rational and stationary extrapolation using rock parameters from the excavation face to the surrounding strata, as the inversion results provide a more precise depiction of the properties of surrounding strata and enable the dynamic design for supporting structure during construction. Therefore, an innovative approach was proposed in this research to conduct probability analysis on the mechanical behaviors of tunnels in heterogeneous strata based on conditional random field models. The statistical characteristics of random variables in these fields were constrained by the derived rock parameters on the excavation face using Hoffman method. The probability distributions of mechanical behaviors were analyzed for tunnels with both symmetric and asymmetric anchor cable systems. In addition, a trained convolutional neural network (CNN) model was implemented to reduce the computational resources required in massive numerical simulations. The tunnel deformation at different circumferential locations can be predicted with an acceptable accuracy and minimal time consumption that significantly facilitated the probabilistic assessments.