Scholarlink

<p>Monitoring and measurement, as an important component of New Austrian tunneling method, further guides the secondary lining construction by collecting surrounding rock deformation data from the tunnel construction site and analyzing and processing them. However, the monitoring data cannot identify the potential trend of deformation and predict unknown events. In this study, machine learning technologies, artificial neural network (ANN) and genetic programming (GP), are used to predict the variation of tunnel vault subsidence with days via on-site measurement data, and establish a prediction model of subsidence and tunnel depth, analyze the mutual influence between parameters, and compare the regression analysis to verify the prediction effect of machine learning; through the established prediction model, it is substituted into FLAC3D to simulate the tunnel excavation, and calculate the tunnel roof deformation after excavation. The results show that both GP and ANN demonstrated exceptional predictive performance, with GP exhibiting marginally superior accuracy (<i>R</i>²= 0.9990) and output stability compared to ANN’s higher terminal displacement errors. Regression analysis yielded lower fidelity. The GP model’s asymptotic prediction of 31.73&#xa0;mm long-term settlement provides a critical benchmark for evaluating surrounding rock stability over extended service periods. The burial depth-dependent settlement model also shows that the highest prediction accuracy and trend of the GP model. A numerical model of tunnel deformation was constructed and validated by comparing GP-predicted vault settlements with FLAC3D-simulated values, and the numerical simulated subsidence is also very close to the on-site situation. The prediction model of subsidence and burial depth can provide an effective reference for the design of the reserved deformation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of Tunnel Deformation Based on Machine Learning and Numerical Simulation

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Siqi Li,
  • Anchang Lan,
  • Lingxiao Yin,
  • Beichen Yu,
  • Yanbo Sun

摘要

Monitoring and measurement, as an important component of New Austrian tunneling method, further guides the secondary lining construction by collecting surrounding rock deformation data from the tunnel construction site and analyzing and processing them. However, the monitoring data cannot identify the potential trend of deformation and predict unknown events. In this study, machine learning technologies, artificial neural network (ANN) and genetic programming (GP), are used to predict the variation of tunnel vault subsidence with days via on-site measurement data, and establish a prediction model of subsidence and tunnel depth, analyze the mutual influence between parameters, and compare the regression analysis to verify the prediction effect of machine learning; through the established prediction model, it is substituted into FLAC3D to simulate the tunnel excavation, and calculate the tunnel roof deformation after excavation. The results show that both GP and ANN demonstrated exceptional predictive performance, with GP exhibiting marginally superior accuracy (R²= 0.9990) and output stability compared to ANN’s higher terminal displacement errors. Regression analysis yielded lower fidelity. The GP model’s asymptotic prediction of 31.73 mm long-term settlement provides a critical benchmark for evaluating surrounding rock stability over extended service periods. The burial depth-dependent settlement model also shows that the highest prediction accuracy and trend of the GP model. A numerical model of tunnel deformation was constructed and validated by comparing GP-predicted vault settlements with FLAC3D-simulated values, and the numerical simulated subsidence is also very close to the on-site situation. The prediction model of subsidence and burial depth can provide an effective reference for the design of the reserved deformation.