Rock joint roughness determination method based on deep learning of time–frequency​ spectrogram

Abstract

Aiming at the problem that it is hard to fully reflect the surface morphology of rock joints with a certain feature parameter, a joint roughness determination method based on deep learning of time–frequency spectrogram was proposed. Firstly, regarding the rock joint profile as a time series signal, the time–frequency spectrogram was drawn through the Short-Time Fourier Transform (STFT) to comprehensively characterize the undulation location and degree of joint, and describe the roughness feature information of the joint profile more clearly and accurately. Then, the deep convolutional neural network was combined to extract and learn the features of the time–frequency spectrogram, and the roughness coefficient of the rock joint profile was identified, which effectively recovered the deficiency of the traditional artificial formulation feature parameters. The experimental results showed that compared with the conventional empirical regression method and machine learning (ML) models, the identified joint roughness coefficient based on deep learning of the time–frequency spectrogram was more consistent with the true value, and the calculation results were more reasonable and reliable, with higher recognition accuracy and generalization ability. It verified the feasibility and effectiveness of deep learning technology in the field of rock joint roughness identification. Finally, the influence of the sampling interval in this method was analyzed, and it suggested that the sampling interval of this method should not be longer than 0.4 mm when evaluating the joint roughness with about 10 cm in length.