| The spatial variability of geotechnical parameters has a significant impact on the reliability analysis of geotechnical engineering.And random field theory is a powerful tool to reasonably consider the impact of spatial variability on the reliability analysis.The sampling samples obtained from random field theory simulation require combining with deterministic models established by traditional numerical analysis software,such as MADIS GTS/NX,FLAC3 D and ABAQUS.The mechanical response of geotechnical structures under these samples are obtained to realize the geotechnical reliability analysis.In this paper,we combine the K-L expansion method for random field simulation and propose a stochastic finite difference method using MATLAB and the finite difference software FLAC3 D.To improve the computational efficiency of reliability analysis,the high-dimensional feature parameters of random field are analogous to image pixels,and the specific application of random field theory in geotechnical engineering reliability analysis is realized by applying convolutional neural network(CNN)as a surrogate model of stochastic finite difference method.The reliability analysis method of geotechnical engineering based on stochastic finite difference method and CNN proposed is applied to tunnel engineering and slope engineering.The research results show that the reliability analysis method can effectively deal with the stability analysis of geotechnical engineering considering spatial variability.In the results of tunnel random field dispersion achieved by stochastic finite difference method,it is found that the tunnel surface settlement is more uniformly distributed as the coefficient of variation of elastic modulus increases.At the same time,it is found that the change in auto-correlation distance has a more significant effect on tunnel surface settlement when the auto-correlation distance is less than two or three times the distance of the tunnel diameter.When CNN is trained on the random field simulation results,the prediction performance of the CNN model gradually increases as the number of training samples increases,but when the training samples exceed a certain value,the prediction performance changes less.The prediction performance of the CNN model gradually increases with the increase of auto-correlation distance.The reliability analysis of tunnel surface settlement based on stochastic finite difference and CNN has high computational efficiency.In the case of slopes considering spatial variability characteristics,the proposed method has higher computational accuracy compared with artificial neural network(ANN)、response surface method(RSM),and is compared with the literature to prove the feasibility of the method. |
