Abstract:

In rock engineering scenarios, such as mining extraction, underground energy storage, and deep tunnel excavation, rock fracture behavior is often abrupt and potentially destructive. Therefore, accurately perceiving the different stages of rock fracture is essential for geological hazard early warning and engineering risk mitigation. However, how to achieve multi-source feature fusion and reliable state discrimination of rock fracture behavior based on existing monitoring techniques, especially acoustic emission, still requires further investigation. This study proposes an intelligent identification framework for rock fracture stages driven by the fusion of knowledge and data. Taking granite as a representative material, stage classification rule knowledge was first established based on rock mechanics theory, and the resulting rules were then used to automatically annotate acoustic emission data. On this basis, three machine learning algorithms, namely extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM), were employed in a data-driven framework to identify different stages of granite fracture. A comparative analysis of the three models shows that the XGBoost model achieves the best overall performance, with a weighted F1-score (the weighted average of per-class F1-scores by sample count) of 0.860 3, a macro-average recall of 80.16%, and a macro-average AUC of 0.97. Macro-average AUC refers to the arithmetic mean of AUC values across all classes. The receiver operating characteristic (ROC) curve plots false positive rate versus true positive rate to illustrate model discriminative performance at different thresholds. AUC denotes the area under the ROC curve, ranging from 0 to 1, where higher values correspond to better binary classification performance. The key interpretable features for stage recognition include cumulative energy, cumulative ring-down count, ring-down count, and hit count. Among these, cumulative energy and cumulative ring-down count showed the largest SHAP (quantify the contribution of each feature to model predictions) contributions. This study establishes a new paradigm for the automatic identification of rock fracture states and disaster warning in rock engineering.

Key words: knowledge-driven, data-driven, interpretability, rock fracture, acoustic emission, intelligent identification