Quantification of water inflow in rock tunnel faces via convolutional neural network approach

Quantifying water inflow information from rock tunnel faces is critical for field engineers to assess the rock mass rating and subsequently make appropriate construction management decisions. This paper proposes a novel convolutional neural network (CNN)-based water inflow evaluation method that emulates a typical field engineer's inspection process. It is integrated by a classification step and a semantic segmentation step: the first one is to classify the non-damaged regions and the damaged regions; and the second one is to segment the detailed water inflow damage to the rock tunnel faces. An image database of water inflow in rock tunnel faces was applied for comprehensive training, validation and testing. The experiments on the testing data demonstrate an ideal performance in terms of convergence speed and classification accuracy, as well as quantitative water inflow segmentation. The proposed automatic quantification approach significantly reduces the ergodic damage segmentation procedure through the early exclusion of undamaged samples during the classification process.

Automated summary

The paper presents a novel convolutional neural network-based method for evaluating water inflow in rock tunnel faces, which achieves accurate quantification through classification and segmentation steps. Experimental results demonstrate excellent performance in terms of speed, accuracy, and segmentation capability.