Research on Spatial Prediction of Rock Mass Strength Based on Advanced Boring Data

Sparse tunnel advance borehole data constrain the precise spatial prediction of rock mass strength, posing challenges to engineering safety. To address this, this paper proposes a machine learning model integrating random forests with adaptive regularization to enhance predictive performance under small sample conditions. This method characterizes spatial variability in rock mass by introducing local correction coefficients and employs random forests for spatial interpolation. It employs adaptive regularization to dynamically balance the model's goodness-of-fit against the spatio-temporal smoothness of its predictions. Comparisons with support vector machines, gradient-boosted trees, and Kriging demonstrate that the random forest model achieves higher prediction accuracy and better generalization across training, validation, and test datasets. Its mean absolute error and root mean square error are significantly lower than those of the comparison methods. Residual analysis further reveals that the random forest's error distribution most closely approximates a normal distribution, exhibiting minimal variance and robust statistical reliability. Using the derived calibration parameters, a three-dimensional rock mass strength field was generated and risk zoning was performed, achieving intuitive visualization of potential hazardous areas. This method provides an effective and reliable new approach for rock mass strength analysis and risk assessment based on limited geological forecast data.