Verification of the accuracy of rock material determination by hammer strike sound using deep learning

In geological fieldwork and concrete inspection, the sound of hammer blows is used to determine the quality of the material. In the case of rock, the sound of hammer blows is greatly influenced by the density of rock fractures and the heterogeneity of the rock composition. Therefore, even experienced geotechnical engineers may judge the goodness or badness of rock differently from person to person. Furthermore, the social issue of the lack of human resources of geotechnical engineers requires the application of new technologies, such as sensing technology and deep learning, to solve the problem.

In the field of civil engineering, deep learning technology is used to determine the state of deterioration from the sound of concrete being struck. However, rocks are more heterogeneous than concrete, and their applicability needs to be verified. In addition, there are no examples of matching technical decisions made by geotechnical engineers with deep learning models or verifying their accuracy. Therefore, in this study, a deep learning model was constructed based on spectral analysis of impact sound frequencies and learning audio information in order to quantitatively determine the material quality of rocks by impact sound. The validation target was rocks at a dam construction site in Japan.

The deep learning model employed a CNN, which has been reported to be used in a number of general audio classification problems, such as environmental sounds. Specifically, we constructed (1) YAMNET, which was transfer-trained on the impact sound of rock materials, and (2) 2D-CNN, which was trained by converting the impact sound of rock materials into log-mel spectrogram images, and conducted comparative verification. In order to construct the model, stratified 5-folds cross-validation was performed using data excluding test data, and optimal hyperparameters were searched.Also, the percentage of test data is 20% for all data.

As a result, we were able to construct a model with an F-score of approximately 90% with respect to the geotechnical engineer's judgement results. In the comparison between YAMNET and 2D-CNN, the F-score of 2D-CNN was superior by a few per cent. This difference can be attributed to the length of time of the input audio signal.

Finally, the model can estimate whether rock materials are good or bad with almost the same accuracy as a geotechnical engineer in the field. In addition, the deep learning model can make a decision in a few seconds. In the future we plan to make effective use of smartphones equipped with the model to improve the efficiency of field work and save manpower. In addition, validation will be carried out to estimate not only whether the rock material is good or bad, but also the detailed material classification. For the deep learning learning algorithm, we intend to compare and study state-of-the-art technologies, such as transformers, and to carry out verification to improve the accuracy of the system.