基于MEC-BP神经网络的基坑水平位移反演分析

Horizontal displacement back-analysis for deep excavation using MEC-BP neural network

随着城市地铁建设快速发展,地铁车站基坑的变形要求越来越严格,基坑土层设计参数的选择面临着极大挑战。利用思维进化算法(MEC)优化BP神经网络的初始权值和阈值,结合有限元数值模拟,提出基于基坑水平位移的土层参数反演分析方法。采用文献算例对该方法进行验证,并与不同反演方法进行对比。研究结果表明:1)MEC-BP神经网络对多工况水平位移的反演分析结果与文献结果基本一致,验证了该方法的有效性和实用性;2)MEC-BP神经网络的收敛速度快于遗传神经网络(GA-BP),其反演结果优于常规BP神经网络、GA-BP方法和修正高斯-牛顿法(G-N);3)采用标量误差函数F_(err)进行寻优,可以提高MEC-BP法水平位移反演分析结果的稳定性和准确性。

With the rapid development of urban subway construction,the deformation requirements of subway station foundation pit become more and more strict,so the selection of soil parameters for deep excavation is faced with great challenges.A new method of horizontal displacement back-analysis for excavation using the MEC-BP neural network was proposed.The initial weights and threshold values of the BP neural network were optimized using Mind Evolutionary Computation(MEC).The MEC-BP neural network was combined with the finite element simulation.Then,the method was validated by an example from literature and compared with different inversion methods.The following conclusions of this paper include.(1)It is showed that the backanalysis results of the MEC-BP neural network for multi-step excavation are basically consistent with the results of the literature,which verifies the effectiveness and practicability of the method.(2)The MEC-BP neural network can guarantee not only a better convergence speed than the Genetic neural network(GA-BP),but also a better back-analysis result than other inversion methods,such as the GA-BP neural network,BP neural network,and the Gauss-Newton method(G-N)proposed in the literature.(3)Using the scalar error function(Ferr)for optimization can improve the stability and accuracy of horizontal displacement back-analysis for deep excavation.