The rock mass boreability assessment for tunnel boring machine (TBM) is of great significance for safe and efficient tunneling. This study presented an improved attribute-weighted deep belief network model (IAWDBN) to perform point and interval estimation of rock mass boreability. In the model, the Bayesian optimization algorithm was integrated to optimize the hyper-parameters automatically, and the early stopping strategy was merged to prevent overfitting; 219 sets of data in total were collected from three different tunnel projects to train the model. Each set of data was composed of four input variables (i.e., rock uniaxial compressive strength, rock quality designation, angle between weakness plane and TBM advancing direction, and tunnel diameter) and one corresponding output variable (i.e., field penetration index). In data preprocessing, the Kriging interpolation and CRITIC (criteria importance through intercriteria correlation) weighting algorithms were separately implemented to complement the missing data and determine the variable weight in the database. Then, the model was applied in Yinsong and LXB water conveyance tunnels (China). Results indicated that the root mean square error (RMSE), mean absolute percentage error (MAPE), determination coefficient (R-2), and interval coverage probability (ICP) of 37 sets of data in Yinsong water conveyance tunnel were 1.92, 7.88%, 0.9470, and 100%, respectively, and those of 49 sets of data in LXB water conveyance tunnel were 1.95, 5.85%, 0.9913, and 100%, respectively. Further, the impact of data weighting and confidence level on model performance was discussed, verifying the advantage of data weighting and suggesting that the confidence level should not be less than 93%. Finally, the comparison analysis was conducted with back-propagation neural network, extreme learning machine, support vector regression, random forest, one-dimensional convolutional neural network, and long short-term memory network in terms of prediction accuracy and running speed, demonstrating the superiority of the model built in this study.
