隧道围岩分级特征智能识别及可视化研究  

Intelligent identification of classification features of tunnel surrounding rock and visualization

作　　者：陈卫东 李天斌[2,3] 黄音昊 杨罡 王皓 肖华波 CHEN Weidong;LI Tianbin;HUANG Yinhao;YANG Gang;WANG Hao;XIAO Huabo(Chengdu Engineering Corporation Limited,Power China,Chengdu 610072,China;State Key Laboratory of Geohazard Prevention and Geoenvironment Protection,Chengdu 610059,China;College of Environment and Civil Engineering,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]中国电建集团成都勘测设计研究院有限公司,四川成都610072 [2]地质灾害防治与地质环境保护国家重点实验室,四川成都610059 [3]成都理工大学环境与土木工程学院,四川成都610059

出　　处：《铁道科学与工程学报》2024年第1期406-421,共16页Journal of Railway Science and Engineering

基　　金：国家自然科学基金联合基金资助项目(U19A20111);国家自然科学基金资助项目(42130719);四川省科技计划重点研发项目(2021YFS0317);四川省科技计划项目应用基础研究项目(2021YJ0041)。

摘　　要：准确、快速获取隧道围岩级别对工程设计、施工及运营意义重大。结合多个四川高速公路隧道项目,共采集7000余张隧道掌子面图像,并利用数据增广方法将数据集扩充至20000余张。按节理裂隙特征、风化卸荷情况、地下水发育情况3种像特征对数据集进行分类标注,并按8:2的比例划分为训练集与验证集。结合深度学习方法,实现掌子面围岩分级特征参数提取识别。搭建了VGG系列、ResNet系列、DenseNet系列、GoogleNet、InceptionV3等卷积神经网络分类模型,并引入准确率、查准率、召回率及F1值等多种评价指标对比分析多种卷积神经网络分类模型的围岩特征(掌子面图像的节理裂隙特征、风化卸荷特征以及地下水发育特征)识别效果。研究结果显示基于DenseNet模型分类识别效果最好,分类准确率分别为:围岩节理裂隙特征87.5%,风化程度特征90%,地下水发育程度特征91.5%,且各特征的F1值均在0.789以上,最高为0.944,平均值为0.852。此外,对DenseNet系列分类模型进行可靠性验证,基于CAM以及Grad-CAM对模型进行分类决策可视化研究分析,分类决策热力图结果显示分类结果与标签特征的强度、位置及范围强相关,为掌子面围岩智能分级提供一定的可解释性,同时也证明了分类模型的预测效果理想。研究成果为利用深度学习进行围岩特征提取提供了新思路。It is of great significance for engineering design,construction,and operation to obtain the surrounding rock grade of tunnel accurately and quickly.Combined with several Sichuan expressway tunnel projects,more than 7,000 tunnel face images were collected,and the data set was expanded to more than 20000 by the data augmentation method.According to the characteristics of joints and fissures,weathering and unloading,and groundwater development,the data set was classified and labeled,and divided into a training set and verification set according to the ratio of 8:2.Combined with the deep learning method,the feature parameters of surrounding rock classification in the working face were extracted and identified.The classification models of convolutional neural networks,such as VGG series,ResNet series,DenseNet series,GoogleNet,and InceptionV3,were established.The identification effects of surrounding rock characteristics(joint and fissure characteristics,weathering and unloading characteristics,and groundwater development characteristics)of various convolutional neural network classification models were compared and analyzed by introducing various evaluation indexes such as accuracy,precision,recall,and F1 value.The research results are drawn as follows.The classification based on the DenseNet model has the best recognition effect.The classification accuracy is 87.5%for the characteristics of surrounding rock joints and fissures,90%for the characteristics of weathering degree,and 91.5%for the characteristics of groundwater development degree.The F1 values of all the characteristics are above 0.789,with the highest value of 0.944 and the average value of 0.852.In addition,this paper verifies the reliability of DenseNet series classification models.Based on CAM and Grad-CAM,the classification decision visualization of the model is studied and analyzed.The results of the classification decision thermogram show that the classification results are strongly related to the intensity,location,and range of the tag features,whic

关 键 词：围岩特征 卷积神经网络 模型可视化 深度学习 隧道掌子面 

分 类 号：O319.56[理学—一般力学与力学基础]