为充分利用现场监测数据所反馈的围岩变形信息，对岩体力学参数进行反演，以贵州省剑河至黎平高速公路TJ-1标段牛练塘隧道为工程背景，选择围岩弹性模量、黏聚力、泊松比及内摩擦角为影响因素，通过设计正交试验及有限元模拟，获取25组围岩物理力学参数组合及其对应的拱顶沉降值和拱腰收敛模拟值。基于随机梯度下降算法(stochastic gradient descent algorithm，简称SGD算法)对传统BP神经网络模型进行改进，建立以拱顶沉降值和拱腰收敛值为输入参数，以围岩弹性模量、黏聚力、泊松比及内摩擦角为输出值的基于SGD算法优化的BP神经网络模型，实现围岩参数的反演分析。将反演所得的围岩参数代入有限元模型，验证优化BP神经网络模型的可行性和准确性。最后，分析围岩变形及初期支护受力特性并给出施工建议。结果表明： 1）基于SGD算法优化的BP神经网络模型计算得出的拱顶沉降值、拱腰收敛值、拱肩收敛值与现场实测值的相对误差率在2.50%~24.01%，均低于传统BP神经网络模型计算得出的误差率（11.51%~93.71%），验证优化BP神经网络模型的可行性和优越性； 2）上、下台阶拱脚处的喷层和锚杆有应力集中现象，有破坏风险，建议施工中加强拱脚支护，防止发生工程事故。

In this study, onsite monitoring data are used to analyze the deformation information of surrounding rocks and conduct back analysis of the mechanical parameters of rock mass. A case study is conducted on the Niuliantang tunnel in the TJ1 bid section of the JianheLiping expressway in Guizhou, China. The influencing factors, such as the elastic modulus, cohesion force, Poisson′s ratio, and internal friction angle of the surrounding rock are selected. Orthogonal tests and finite element simulations are designed to obtain 25 sets of physicomechanical parameters of the surrounding rock and their corresponding simulation results for crown settlement and arch waist convergence. Furthermore, traditional back propagation(BP) neural networks are improved using a stochastic gradient descent(SGD) algorithm. Crown settlement and arch waist convergence are used as input parameters, and the elastic modulus, cohesion force, Poisson′s ratio, and internal friction angle of the surrounding rock are used as output parameters. This approach enables the inversion analysis of the surrounding rock parameters. The feasibility and accuracy of the SGDoptimized BP neural network model are validated by substituting the inversionobtained surrounding rock parameters into the finite element model. The deformation and initial load characteristics of the surrounding rock are analyzed, and construction suggestions are provided. The results reveal that the relative difference between the crown settlement, arch waist, and arch shoulder convergence calculated by the SGDoptimized BP neural network model and the monitoring results range from 2.50% to 24.01%, which is smaller than those calculated by the traditional BP neural network model(11.51%~93.71%). These findings validate the feasibility and superiority of the SGDoptimized BP neural network model. Because of the stress concentration of the spray layer and anchor bolt at the arch foot of the upper and lower benches, the arch foot support must be enhanced to prevent engineering accidents.