Because of the complexity of the geological features, when the NATM method is used in Japan, the rock mass is evaluated in nine categories (A. condition of tunnel face, B. condition of excavation face, C. compressive rock strength, D. weathering and alteration, E. spacing of discontinuities, F. condition of discontinuities, G. direction of discontinuities, H. presence of water inflow, I. deterioration due to water) The evaluation is graded on four levels. The objective of this study is to use deep learning to quantitatively evaluate the frequency, condition, and morphology of fractures, as well as weathering and alteration of the tunnel faces; CNN was used to grade the three criteria regarding fractures. Furthermore, ratio of weathering area was detected by HSV color space for categories regarding weathering and alteration. We also applied Grad-CAM to verify whether the CNN model could actually evaluate rock fractures as a decision criterion.
In Japan, the New Austrian Tunneling Method (NATM) is a very popular tunnel construction method because it can adapt to the complex geological formation of Japan. This construction method relies on the surrounding rock mass to ensure the stability of the structure. Furthermore, to maximize safety and minimize costs, support structure (determined through support patterns) could be change from the original design based on the observed rock mass as stated in the "Index for Road Tunnels Observation and Measurement (2009)". Rock mass on tunnel face are evaluated and graded based on a set of criteria, and labelled with a support pattern. Rock mass evaluation are typically done by onsite engineers, but since these decisions are based on their individual experiences, there is a discrepancy in judgement resulting in differing evaluations. This study will apply a type of deep learning method, the Convolutional Neural Network (CNN), to the process of rock mass evaluation. This study will focus on evaluating the visually observable features of rock mass, namely the rock fractures. Finally, Gradient-weighted Class Activation Map (Grad-CAM) is implemented to visualize CNN. In addition to the evaluation of rock fractures, the weathering area of rocks is quantitatively extracted by using HSV color space.
