Back analysis of mechanical parameters of rock masses based on multi-point time-dependent monitoring data

Abstract

The physical and mechanical parameters of rock masses are important factors that affect the numerical simulation and safety evaluation of geotechnical engineering. Due to the discontinuity and heterogeneity of rock mass, the parameters obtained by laboratory test and field test can not represent the actual situation. With the development of computer technology, the parameter identification method based on machine learning provides a new way to solve the above problems. Aiming at the rheological model based on thermodynamics with internal state variables, a back analysis method of mechanical parameters based on IAGA-BP algorithm is proposed. The sensitivity of deformation parameters, damage effect parameters and viscoplastic parameters involved in the model are analyzed using range analysis method. And the parameters which have great influence on the deformation of surrounding rock are selected to reduce the dimension of output layer. The displacement corresponding to the representative time point are selected as the input layer neuron. A numerical model is established to finely simulate the driving process of TBM, which is used to generate training samples and test samples. BP, GA-BP, PSO-BP, APSO-BP and SVM are also used to inverse the parameters of surrounding rock. Two evaluation metrics are used to evaluate the performance of above algorithms, including prediction accuracy and stability. The method is applied to inverse parameters of surrounding rock of a tunnel. The results show that the time-dependent deformation curve of surrounding rock obtained by BP neural network optimized by improved adaptive genetic algorithm (IAGA-BP) fits well with the monitoring curve (the average prediction error is 3.8%, and the prediction stability is 16.3%), which is much better than other algorithms.