基于深度学习的空间变异土体中隧道水平收敛安全系数计算

Calculation of Horizontal Convergence Safety Factor for Tunnels in Spatially Variable Soil Based on Deep Learning

为提升随机有限元法用于计算空间变异土体中隧道水平收敛安全系数的效率,提出一种空间注意力-卷积神经网络作为随机有限元的代理模型。该代理模型以具有空间变异性的土体参数为输入,以隧道水平收敛安全系数为输出,从随机有限元生成的少量样本中学习土体参数随机场与隧道水平收敛安全系数间的关系,进而在更多样本上代替随机有限元方法进行安全系数的计算。以上海某地铁隧道水平收敛安全系数计算问题为例测试该代理模型,结果表明:代理模型与随机有限元计算的隧道水平收敛平均安全系数相对误差小于2%;代理模型与随机有限元计算的水平收敛安全系数MAPE、RMSE、MAE分别小于10%、0.12、0.10,R~2大于0.8,可以满足工程准确度需求;同时,相比于随机有限元算法,代理模型的计算效率提升约880倍。

To improve the efficiency of using the random finite element method(RFEM)for calculating the safety factor of tunnel horizontal convergence in spatially variable soil,a spatial attention-convolutional neural network(SA-CNN)is proposed as a surrogate model for RFEM.This surrogate model takes spatially variable soil parameters as input and the tunnel horizontal convergence safety factor as output,learning the relationship between soil parameter random fields and the tunnel safety factor from a limited number of RFEM samples.It then replaces the RFEM method for calculating safety factors on larger samples.Tested on a Shanghai metro tunnel,the model shows a relative error of less than 2%compared to RFEM,with MAPE,RMSE,and MAE values below 10%,0.12,and 0.10 respectively,and R²above 0.8,meeting engineering accuracy requirements.Additionally,the calculation efficiency of the surrogate model is approximately 880 times higher than RFEM.