Self-correcting prediction of rock mass discontinuity ahead of tunnel face: A model-driven and observation-corrected loop

Accurate, long-range prediction of rock mass structure is essential for safe and efficient tunnel construction. For tunnels excavated by the drill-and-blast method, deep learning-based spatiotemporal forecasting offers a non-contact approach for predicting discontinuity distributions in the rock mass, yet these data-driven models are susceptible to the noise, incompleteness, and sparsity of real-world observations. This paper proposes a framework for multi-source data fusion that combines advanced drilling data with a deep learning-based spatiotemporal model to predict the distribution of traces in the rock mass. This framework overcomes the inherent limitations of standalone predictors by synergistically integrating a spatiotemporal forecasting model with sequential data assimilation. The SAM-ZRTM-ConvLSTM spatiotemporal model serves as the core predictive engine for trace map prediction. Monte Carlo (MC) dropout is employed to quantify predictive uncertainty and, critically, to generate a diverse forecast ensemble. Subsequently, the Ensemble Kalman Filter (EnKF) is integrated to assimilate sparse borehole data, continuously correcting the model's state to mitigate compounding errors and suppress observational noise. The results demonstrate that the framework effectively filters observational artifacts and corrects model predictions, producing a coherent and reliable long-range geological forecast, particularly under conditions involving construction-induced cracks simulating real-world excavation. This work demonstrates the potential of combining uncertainty-aware deep learning with data assimilation to maximize the value of sparse and noisy data for reliable geological prediction in challenging tunnelling environments.