The accurate prediction of roof stress in mined-out areas is crucial for ensuring mine safety. However, existing study methods often overlook the increasingly available image data and fail to balance the model predictive capability with interpretability. To address these issues, this study innovatively integrates 3D laser scanning image features into the prediction of roof stress in mined-out areas. Image features are extracted using pre-trained deep-learning models and combined with traditional geological parameters to construct multiple machine-learning models for prediction. The experimental results demonstrate that models incorporating image features significantly outperform traditional models that rely solely on geological parameters in terms of prediction accuracy, interpretability, and complexity, revealing the critical role of image features in stress prediction. Furthermore, the use of SHapley Additive exPlanations (SHAP) to interpret the random forest model uncovers new domain knowledge, such as the relationship between spatial patterns and stress concentration. This study theoretically validates the effectiveness of image data and effectively balances the predictive capability and interpretability of the model, facilitating knowledge discovery in the field. On a practical level, it also provides guidance for mine safety management.
Read full abstract