Evaluation of rock joint on construction tunnel face using convolutional neural network

In Japan, evaluation of rock mass is crucial in mountain tunnels that are constructed using the NATM (New Austrian Tunneling Method) to determine the support pattern that in turns determine the final support structure of the tunnel. The current evaluation method is highly dependent on the observations of on-site expert engineers that grades the rock mass in four levels across nine criteria: A. Condition of tunnel face. B. Condition of excavated surface. C. Compressive rock strength. D. Weathering. E. Spacing of discontinuity. F. Condition of discontinuity. G. Orientation of discontinuity. H. Spring water. I. Water degradation. However, in recent years, the number of said experts are decreasing rapidly due to labour shortage caused mainly by an aging population and declining birth rates. This is one of the biggest challenges that the construction sector currently faces, and this may affect the efficiency of tunnel construction projects in the future. Thus, this study proposes the implementation of a deep learning tool called the convolutional neural network (CNN) to quantitatively evaluate rock fractures on tunnel face. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) is applied to visualise CNN, as well as verify the applicability of CNN to the evaluation of rock fractures. In other words, by creating heat maps of tunnel face images, CNN is utilised to evaluate rock mass in categories that are associated with discontinuities. The CNN model is most accurate when six convolution layers, four pooling layers, and three dense layers are used. This yields an accuracy of 75.0~88.0 per cent. Thus, the CNN model is verified to be feasible method to evaluate rock fractures.