Deep learning-based inversion of stress curves and crack quantification in NPR anchored rock masses

In geotechnical engineering applications, including mining and tunnel construction, the stability of fractured rock masses is paramount to ensuring structural safety. The spatial distribution and temporal evolution of internal fractures fundamentally govern the mechanical behavior and failure mechanisms of rock masses. Nevertheless, the inherent complexity and structural concealment of rock mass systems pose significant challenges for the direct measurement of critical internal mechanical parameters. This study explores the use of deep learning to invert mechanical responses of NPR (Negative Poisson's Ratio) anchored fractured rock masses. Discrete Element Method (DEM) simulations were conducted to generate datasets including stress-strain curves and crack numbers under various initial fracture distributions. Three models—GRU, CNN+GRU, and CNN+GRU+ATT—were developed to predict rock mechanical parameters from NPR cable force data. Results show that the CNN+GRU+ATT model achieves superior accuracy, with R² > 0.90 and RMSE < 5 on stress prediction tasks. It also accurately estimates initial crack quantity (n_p), with mean prediction error under 10% for high-fracture scenarios. The proposed model effectively captures stress fluctuations, offering early-warning potential for failure. The approach demonstrates strong generalization and robustness across varying crack configurations, providing a feasible framework for real-time health monitoring and mechanical parameter estimation in fractured rock engineering.