CNN-based assessment of the degrees of fracture development in bedrocks and classification of bedrock weathering zones

Background Bedrock weathering zones exhibit distinct vertical gradational characteristics. Accurately identifying the degrees of fracture development in the zones is critical to the safe production of mines in shallowly buried coal seams covered by thick water-bearing unconsolidated layers.

Methods Focusing on the Qinan coal mine in Anhui Province, this study proposed a hybrid convolutional neural network (CNN) model, which enabled feature extraction using a synergistic architecture consisting of conventional convolutional layers and inception modules. Specifically, based on conventional log data and calculated shale content and fracture development index, combined with core observations, this study constructed a dataset containing labels that indicated four degrees of fracture development: fractured, high, medium, and low degrees. The synergistic architecture was then designed, in which conventional convolutional layers and inception modules were employed to extract local subtle features and conduct multi-scale feature fusion, respectively. Subsequently, hyperparameters, including the sequence length and learning rate, were optimized through grid search. Finally, the CNN model was trained to determine the degrees of fracture development intelligently.

Results and Conclusions  Compared to the random forest (RF), XGBoost, and support vector machine (SVM) models, as well as other CNN models, the hybrid CNN model demonstrated superior performance, with an accuracy reaching 92.71% and a weighted average F1 score of up to 93% on the test set. Furthermore, the proposed model can accurately capture the log response characteristics of fracture development in rocks, significantly enhancing the efficiency and continuity of log-based fracture assessment. The hybrid CNN model was applied to 25 boreholes in a certain mining area of the study area. Accordingly, a 3D geological model of weathered zones in the mining area was created. With an average root mean square error (RMSE) of 8.9 m in the prediction of the top boundary depths of various weathering zones, the geological model clearly exhibits the spatial distribution patterns of completely, strongly, and moderately weathered zones, together with slightly weathered to unweathered zones, thereby providing a quantitative geological basis for grouting engineering, along with the prevention and control of roof water hazards, in the coal mine.