Abstract
Uncertainties in rock abrasivity often result in the failure of mechanical excavation and excessive cutter wear during TBM tunnelling. Cerchar abrasivity characteristic tests on weakly cemented sandstones (WCS) in western China revealed that their abrasivity changed with physical state and exhibited fuzzy randomness. In the dry state, the abrasivity index reaches its maximum, with the mud-saturated state being intermediate, and the water-saturated state showing the lowest value. In the dry state, the sandstone's Cerchar Abrasivity Index(CAI) increased (CAI) by 10.64% compared to the mud-saturated state, and by 19.23% compared to the water-saturated state. Studies on the microstructure of WCS in different states indicated that the higher CAI value in the dry state is attributable to its well-preserved internal structure, strong cementation, and high strength. In the water-saturated state, the presence of a slurry film on the sandstone surface led to a higher steel stylus wear than in the fully saturated state. The conventional RBF neural network model was improved by introducing neurone fuzzy splitting and stochastic adjustment of connection weights. Compared to MLR, SVR, and traditional RBF: the MSE of improved RBF is reduced by up to 65.7%, the MAE is reduced by up to 47.4%, and the occurrence rate of local optimality is decreased by 53.8%. Based on this, an improved fuzzy stochastic RBF neural network model was established to predict rock abrasivity using hardness, wave velocity, porosity, and equivalent quartz content as inputs, and the CAI and Cerchar Abrasivity Ratio (CAR) of WCS as outputs. Engineering examples show that the improved RBF fuzzy stochastic model's prediction of rock abrasivity rate has an error of less than 5% compared to the actual measured values, and the predicted Coefficient of Determination (R2) reaches 0.967. Therefore, the enhanced prediction model successfully addressed the uncertainty in abrasivity characteristics of WCS in western China.