Monitoring and Prediction of Deformation and Failure of Roadway Surrounding Rock Based on Binocular Vision and Random Forest

The deformation and failure of surrounding rock in underground roadways are governed by complex mechanical interactions and environmental factors, yet the fundamental scientific patterns behind these processes remain unclear. This lack of real-time, data-driven understanding limits the development of intelligent monitoring and prediction systems in mining engineering. To address this challenge, this study aims to establish an intelligent system for the dynamic monitoring and prediction of roadway surrounding rock deformation based on binocular vision and machine learning. An improved Semi-Global Block Matching (SGBM) algorithm is developed for real-time 3D deformation measurement, while a physical similarity model is constructed to visualize the deformation–failure evolution. The Random Forest (RF) algorithm is employed for deep deformation prediction, and its optimal parameters are determined by minimizing the mean square error. Experimental results show that the average measurement errors of the binocular vision method are 1.22 mm and 0.92 mm, outperforming total station monitoring. The gradient-enhanced Random Forest (GERF) model achieves RMSE values of 0.0164 and 0.0113, with R2 values of 0.8856 and 0.8356, respectively. Compared with AdaBoost, XGBoost, and Vision Transformer models, GERF improves predictive accuracy by 7.82%, 8.68%, and 3.87%, respectively. These findings demonstrate the scientific feasibility and technical advantage of the proposed intelligent system, offering a new approach to understanding and predicting roadway deformation and failure in intelligent mining.