Intelligent identification of classification features of tunnel surrounding rock and visualization
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 20 000 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,which provides some explanations for the intelligent classification of surrounding rock in the heading face,and also proves that the prediction effect of the classification model is ideal.The research results provide a new idea for the feature extraction of surrounding rock by deep learning.
characteristics of surrounding rockconvolution neural networkmodel visualizationdeep learningtunnel face
万方数据
维普
