基于深度学习的隧道工作面岩石结构自动化判别

Automatic Identification of Rock Structure at Tunnel Working Face Based on Deep Learning

从掌子面图像中快速准确获取建设阶段隧道工作面的表观岩体结构特征对于掌握待开挖围岩的稳定性及跟进阶段的施工决策意义重大。文章结合自研数字照相设备获取云南蒙屏高速公路13条隧道在不同工况、温湿度、照度、粉尘浓度环境下的150余个掌子面42100张图像样本,选取现场数据集出现的块体、层状、碎裂、散体、镶嵌等5种主要结构类型,以训练及测试损失率、准确率、召回率等为主要评价指标,建立基于TensorFlow-GPU的岩石隧道掌子面结构的卷积神经网络Inception-ResNet-v2模型,对模型进行训练并实现岩体结构类别的自动识别与分类。研究表明:(1)采用训练集和测试集中的掌子面图像对模型进行分类研究,训练、测试集的准确率分别达到98.21%和94.61%,召回率达到96.14%;(2)测试可视化结果显示Inception-ResNet-v2模型对复杂的现场条件具有较好的鲁棒性,而局部的错检现象需要通过进一步提高样本丰富性和纹理多样性来规避。

The rapid and accurate acquisition of the apparent rock structure characteristics of tunnel working faces during the construction stage from images of the tunnel face is of great significance for the understanding of the sta⁃bility of the surrounding rocks to be excavated and for the decision-making during the follow-up construction stage.This paper uses self-developed digital photographic equipment to acquire 42,100 image samples from more than 150 working faces of 13 tunnels on the Mengzi-Pingbian Highway in Yunnan under different working conditions,temperatures and humidity,illuminations and dust concentrations,selects 5 main structural categories,such as blocky,layered,fractured,granular and mosaic structures,that appear in the field data set,and develops the Tensor⁃flow-GPU-based convolutional neural network Inception-ResNet-v2 model for tunnel face rock structures with the loss rate,precision rate and recall rate of training and testing as the main evaluation indicators.Through model train⁃ing it achieves automatic identification and classification of rock structure categories.The study shows that:(1)us⁃ing the tunnel face images in the training and testing sets to classify the rock categories in the model,it could achieve a precision of 98.21%and 94.61%in the training and testing sets respectively,and a recall rate of 96.14%;(2)the visualization results of the testing show that the present framework has better robustness for complex site con⁃ditions,while the phenomenon of partial identification errors should be circumvented by further improving sample richness and texture diversity.