An intelligent framework for identifying unstable fracture states in layered sandstone using optical monitoring and machine learning

Accurate identification of unstable fracture states in rocks is critical for geological hazard monitoring and early warning. In this study, uniaxial compression tests are performed on layered sandstone, and digital image correlation (DIC) is used to quantify the full-field strain evolution. Based on the tangential sliding displacement along the failure bedding plane, fracture states are classified into binary labels: stable (0) and unstable (1). Guided by statistical theory, seven global statistical features are extracted from the strain fields to construct a structured dataset for fracture state recognition. On this basis, four representative machine learning models, including backpropagation neural network (BPNN), support vector machine (SVM), random forest (RF), and adaptive boosting (AdaBoost), are developed to identify unstable fracture states, with 10-fold cross-validation used for hyper-parameter optimization. The results show that all models achieve high accuracy in identifying unstable fracture states. Among them, BPNN ( Accuracy = 0.983, F 1 - Score = 0.889) and RF ( Accuracy = 0.971, F 1 - Score = 0.808) demonstrate strong overall performance, while AdaBoost achieves the highest Recall (0.972), making it particularly suitable for early warning applications. Although SVM exhibits the highest Precision (1.000), it tends to underpredict unstable fracture states. SHapley Additive exPlanations (SHAP) interpretability analysis indicates that the average strain (Ave) is consistently the most influential feature across all models. Further analysis reveals that increasing the training set ratio and applying appropriate normalization (particularly Z-score) significantly enhances model performance. This study provides a theoretical and technical foundation for smart monitoring and early warning in rock engineering.