土为非均质材料，物性参数具有很大的离散型，直接根据试样的参数难于准确预测地下结构施工引起的地面沉降。为更准确地预测盾构施工引起的地面沉降，提出利用进化神经网络进行多层土的关键物性参数反演分析方法。该方法首先利用参数敏感性分析选择对地表沉降影响显著的参数，然后，以试样的实测参数为基础生成样本，随后，建立基于人工蜂群算法优化人工神经网络参数的进化神经网络（artificial-bee-colony-based artificial neural network, ABC-ANN），并与基于遗传算法优化神经网络（genetic-algorithm-based artificial neural network, GA-ANN）和误差反向传播（back-propagation, BP）神经网络对比，用于盾构场地地表沉降的预测。以苏州地铁6号线盾构隧道下穿既有二号线平河路站为算例进行研究。计算结果表明：压缩模量和内摩擦角是地表沉降的关键影响参数。以实测地表位移为输入，利用训练好的神经网络预测土层参数。ABC-ANN预测参数作为有限元模型计算参数所得地表位移的误差最小。

Soil is a non-homogeneous material with physical parameters exhibiting significant variability, which could be challenged to accurately predict ground surface settlement caused by underground construction based solely on sample parameters. To enhance the accuracy of predicting ground surface settlement induced by shield tunneling, a method was proposed for inverse analysis of key soil physical parameters in multi-layered soils using evolutionary neural networks. Parameter sensitivity analysis was firstly employed in the method to identify parameters that significantly influence surface settlement. Subsequently, samples were generated based on measured parameters from soil specimens. An artificial bee colony algorithm was then utilized to optimize artificial neural network parameters(artificial-bee-colony-based artificial neural network, ABC-ANN), creating an evolutionary neural network. This approach was compared with a genetic algorithm-optimized neural network (genetic-algorithm-based artificial neural network, GA-ANN) and a standard back-propagation neural network (back-propagation artificial neural network, BP-ANN) for predicting ground surface settlement in shield tunneling sites. The shield tunnel undercrossing of the existing Pinghe Road Station on Line 2 of the Suzhou Metro Line 6 was used as a case study. The computational results show that the compression modulus and internal friction angle are the key parameters influencing ground surface settlement. Measured surface displacements is used as input, the soil layer parameters are predicted by the trained neural networks. It is shown that ABC-ANN predicted parameters, when used as input for finite element model calculations, yield the smallest error in surface displacement predictions.