Visualization of fracture structures in tunnel surrounding rocks: Deep learning of borehole images and heterogeneous zonal reconstruction

Accurate three-dimensional (3D) characterization of complex fracture structures in deep rock masses is critical for stability evaluation in tunnel engineering. To address the inefficiency and subjectivity of conventional manual core logging, as well as the neglect of spatial heterogeneity in conventional 3D discrete fracture network (DFN) reconstruction, this study proposes a novel method that integrates deep learning-based intelligent borehole fracture image identification with heterogeneous zonal reconstruction. First, a three-stage cascaded image enhancement framework is established to tackle defects in raw borehole images, such as uneven illumination, low contrast, and blurred edges. Next, a novel Frac-RAUNet model is introduced, achieving 86.4% recall, 85.1% precision, and 85.7% F1-score on an independent test set of 2822 images acquired under complex working conditions. Subsequently, a feature parameter extraction algorithm based on sinusoidal curve consistency is developed to automatically quantify the fracture count, apparent trace length, and fracture occurrence. During reconstruction, multi-scale sliding windows and Kolmogorov-Smirnov (K-S) tests based on fracture spacing are introduced to objectively delineate structurally homogeneous zones. Furthermore, using the method of moments and the coefficient of variation (CV) criterion, the log-normal distribution is demonstrated to be the optimal probability model for characterizing the sizes of deep fractures. Finally, to mitigate the spatial geometric distortion inherent in 3D reconstruction, an innovative targeted calibration approach is proposed. This approach involves rotation angle ( ) correction based on apparent trace length probability density functions, alongside volumetric fracture intensity calibration via forward modeling. Engineering validation conducted on a 60.5 m-long exploration borehole demonstrates that the reconstructed DFN model statistically aligns with field data and authentically captures the spatial nonlinear evolution characteristics of the surrounding rocks. These findings provide a more reliable foundation for evaluating rock mass quality and stability.