Prediction of joint roughness coefficient via hybrid machine learning model combined with principal components analysis

Joint roughness coefficient (JRC) is the most commonly used parameter for quantifying surface roughness of rock discontinuities in practice. The system composed of multiple roughness statistical parameters to measure JRC is a nonlinear system with a lot of overlapping information. In this paper, a dataset of eight roughness statistical parameters covering 112 digital joints is established. Then, the principal component analysis method is introduced to extract the significant information, which solves the information overlap problem of roughness characterization. Based on the two principal components of extracted features, the white shark optimizer algorithm was introduced to optimize the extreme gradient boosting model, and a new machine learning (ML) prediction model was established. The prediction accuracy of the new model and the other 17 models was measured using statistical metrics. The results show that the prediction result of the new model is more consistent with the real JRC value, with higher recognition accuracy and generalization ability. • A JRC prediction model based on machine learning techniques was proposed. • Two principal components were extracted through principal component analysis. • The robustness of the proposed ML model was evaluated by datasets. • The superiority of the proposed ML model was demonstrated with the existing models.