改进一维卷积神经网络的隧道围岩收敛变形分级预测  

Prediction of Tunnel Squeezing Classification Based on Improved One-Dimensional Convolutional Neural Network

作　　者：吴浩[1] 陈运涛[2] 朱赵辉[1] 李秀文[1] 岳强[3] WU Hao;CHEN Yuntao;ZHU Zhaohui;LI Xiuwen;YUE Qiang(China Institute of Water Resources and Hydropower Research,Beijing 100048,China;CCCC Tianjin Port Engineering Institute Co.,Ltd.,Tianjin 300222,China;Shandong Agricultural University,Taian 271018,China)

机构地区：[1]中国水利水电科学研究院,北京100048 [2]中交天津港湾工程研究院有限公司,天津300222 [3]山东农业大学水利土木工程学院,山东泰安271018

出　　处：《应用基础与工程科学学报》2024年第1期145-159,共15页Journal of Basic Science and Engineering

基　　金：国家自然科学基金面上项目(51879284,52179121);中国水利水电科学研究院流域水循环模拟与调控国家重点实验室开放研究基金项目(IWHR-SKL-KF202112)。

摘　　要：隧道围岩收敛变化是认识围岩和支护结构动态作用及其时空演变机理的前提,变形分级准确预测是对围岩稳定性和支护结构有效性评估的重要基础.本文提出了改进一维卷积和支持向量机融合深度网络的隧道收敛变形分级预报模型.根据围岩变形的主要影响因素和特征类型,选取强度应力比、隧道埋深、隧道等效直径、支护刚度和岩体质量指标,建立了变形等级的识别框架.收集了159组国内外经典的隧道收敛变形实例数据,采用全局均值池化层和支持向量机改进传统卷积神经网络中全连接层和Softmax层进行等级分类,利用改进的一维卷积神经网络自动提取隧道变形的隐含典型特征.运用衰减学习率和Dropout正则化深度学习训练技巧,防止模型出现过拟合.与其他方法的结果对比,证明了该方法有着更好的准确率和鲁棒性,模型完全利用数据驱动实现有限数据集的深层复杂且微妙关系学习.应用于多雄拉公路隧道的围岩收敛变形分级预测,预测结果与现场实际一致,进一步验证了方法的准确性和适用性.研究结果有利于提高隧道收敛变形预测的理论水平和可靠性,为类似工程提供参考.The change of tunnel squeezing is the premise to understand the dynamic role of surrounding rock and support structure and its spatial and temporal evolution mechanism.The accurate prediction of tunnel squeezing classification is an important basis for assessing the stability of rock and the effectiveness of support structure.A prediction model of tunnel squeezing classification with IDCNN and SVW fused deep network is proposed.According to the main influencing factors and characteristic types of tunnel squeezing,the framework for identifying the tunnel squeezing intensity is established by selecting five evaluation indices,i.e.,strength stress ratio,tunnel burial depth,rock quality index,tunnel equivalent diameter and support stiffness.159 groups of typical tunnel engineering case data are collected and adopted as the sample data.The global mean pooling layer and SVW are used to improve the fully connected and Softmax layers in traditional CNN for rank classification.The implicit typical features of tunnel squeezing are automatically extracted using the improved IDCNN.Deep learning training techniques are applied to prevent model overftting,such as Dropout regularization and decaying learning rate.The comparison with the results of other methods proves that this method has better accuracy and robustness.The model is completely data-driven to achieve deep,complex and subtle relationship learning with limited data sets.The model is applied to predict the convergence deformation classification of the surrounding rock of the Doxiongla highway tunnel,and the prediction results are consistent with the actual conditions in the field,which further verified the accuracy and applicability of the model in this paper.The results can be used to improve the theoretical level and reliability of tunnel squeezing prediction and provide a good reference for similar projects.

关 键 词：隧道工程 收敛变形 预测 深度学习 改进一维卷积神经网络 

分 类 号：TV697.1[水利工程—水利水电工程]