Attention-guided cascaded network for predicting tunnel surrounding rock properties using measurement-while-drilling data

The quality of rock mass is a critical factor in evaluating the stability of surrounding rock and determining excavation and support strategies during tunnel construction. This paper presents a physical mechanism-constrained cascade (PMC) model strategy for multi-task learning (MTL), effectively capturing both the correlations and distinctions among tasks. In addition, an attention-guided cascade progressive layer extraction (AGCPLE) model is proposed to predict key rock mass quality indicators, including the surface rock quality designation index (S-RQD), Schmidt rebound value (RV), and basic quality (BQ) of surrounding rock. The AGCPLE model utilizes measurement-while-drilling data from multiple blastholes, and tunnel construction data as input to predict these indicators. Performance and generalization capabilities are evaluated using data from the Yangjiawopu tunnel in China. Results show that the AGCPLE model outperforms conventional deep learning and machine learning approaches. Furthermore, the PMC model strategy shows improved predictive performance compared to other MTL strategies.