Cutting-edge approaches to specific energy prediction in TBM disc cutters: Integrating COSSA-RF model with three interpretative techniques

Abstract

Specific energy (SE) is an important index to measure crushing efficiency in mechanized tunnel excavation. Accurate prediction of the SE of tunnel boring machine disc cutters is important for optimizing the crushing process, reducing energy consumption, and minimizing machine wear. Therefore, in this paper, the sparrow search algorithm (SSA), combined with six chaotic mapping strategies, is utilized to optimize the random forest (RF) model for predicting SE, referred to as the COSSA-RF prediction models. For this purpose, an SE prediction database was established for training and validating model performance, encompassing 160 sets of experimental data, each with six input parameters: uniaxial compressive strength (UCS), Brazilian tensile strength (BTS), disc cutter diameter (D), cutter tip width (T), cutter spacing (S), and cutter penetration depth (P), along with a target parameter, SE. The evaluation results indicate that the COSSA-RF models demonstrate superior performance compared to other four machine learning models. In particular, the Chebyshev map-SSA-RF (CHSSA-RF) model achieves the most satisfactory prediction accuracy among all models, resulting in the highest coefficient of determination R² and dynamic variance-weighted global performance indicator values (0.9756 and 0.0814) and the lowest values of root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) (6.4742, 4.0003, and 20.41%). Lastly, the results of interpretability analysis of the best model through SHapley Additive exPlanations, local interpretable model-agnostic explanations, and Vivid methods show that the importance of input parameters ranked as follows: UCS, BTS, P, S, T, and D. Moreover, interactions between parameters (UCS and BTS, BTS and P, and BTS and S) significantly influence the model predictions.