Abstract:  A dynamic classification method is developed based on the actual revealed conditions of the surrounding rock of tunnels during tunnel construction. This method can reduce economic losses, safety accidents, and other problems caused by differences between the actual classification of surrounding rock and geological surveying results. First, a case study is conducted on the Muzhailing tunnel on the Weiyuan-Wudu expressway in Gansu, China; tunnel face images are collected using a DSLR camera, and a T-ResNet model is designed based on the deep-learning ResNet18 network to identify the features of the tunnel face rock images qualitatively. Then, by leveraging digital image processing technology to quantitatively extract joint and fracture characteristics, the integrity index of the tunnel face is determined. Finally, a rock quality assessment index system is established based on indices of rock hardness, rock mass integrity, main structural plane occurrence, groundwater development condition, initial ground stress, and joint expansibility. Thus, an intelligent dynamic classification of tunnel rock is achieved using the feature-weighted K-nearest neighbors(KNN) algorithm. The results demonstrate the following: (1) The accuracies of the T-ResNet model in the joint fracture test set and the groundwater test set reach 83.23% and 92.86%, respectively, showing its feasibility and high precision in surrounding rock recognition. (2) The machine vision method used in processing the on-site surrounding rock images can rapidly extract the rock mass integrity coefficient and groundwater development condition, realizing intelligent and efficient analysis. (3) The KNN intelligent dynamic classification method has good applicability and high accuracy in the case study.

Key words: highway tunnel, deep learning, digital image processing technology, K-nearest neighbor algorithm, surrounding rock classification