Experimental study on CNN-based automatic rock classification through hammering sounds
Y. Maruyama, K. Kuribayashi, Shota Nakashima, H. Ueda · 2025
ABSTRACT: The characteristics of rock masses at excavation surfaces or tunnel faces are crucial for determining designs in civil engineering projects such as dams and tunnels. Rock masses consist of both rock and discontinuities, and hammering tests have been widely used for rock mass classification due to their cost-effectiveness and ease of handling. However, quantifying and systematizing the relationship between hammering sound data and rock deterioration is challenging, particularly for less experienced engineers. The aim of this study was to automate the assessment of rock hardness at excavation sites by utilizing a convolutional neural network (CNN) trained on data obtained from hammering tests, ultimately developing an automatic hammering evaluation system. The process involved converting hammering test data into spectrograms, which serve as an input feature for CNN-based rock grade classification. By analyzing the spectrograms, the CNN was found to effectively extract patterns associated with different rock grades, enabling an objective and systematic evaluation. As a result, the hammering waveforms exhibited distinct patterns, depending on the rock grade, demonstrating the potential of the automatic hammering evaluation system. These findings suggest that CNN can contribute to improving rock mass classification, making excavation site assessments more efficient and reliable.