关键词: 盾构渣土改良, 机器学习, 参数预测, 逆向设计, SVR-BPNN混合模型

Abstract: To achieve intelligent design and precise prediction of soil conditioning, this study proposes a hybrid modeling approach integrating machine learning and inverse reasoning. Sandy soils collected from the Shenyang Metro Line 3 project were used as test materials, and a dataset containing 120 groups of data was established through laboratory tests, covering key parameters such as effective particle size, foam injection ratio, slurry concentration, and slurry injection ratio. The performance of six machine learning algorithms in predicting soil slump and permeability coefficient was systematically compared, and the characteristics importance was solved using Shapley additive explanations (SHAP) method. For overfitting and non-unique solution in parameter inverse design, a support vector regression (SVR)-back propagation neural network (BPNN) hybrid model integrated modeling strategy was proposed, parameter combinations were generated through BPNN, and the model was validated and iteratively optimized using SVR model. The results show the following: (1) The SVR model performed best in both tasks, demonstrating high prediction accuracy and strong robustness. (2) Interpretability analysis based on the SHAP method indicates that effective particle size is the most critical factor influencing conditioning effectiveness, and different output variables exhibit significant differences in sensitivity to input features. (3) The SVR-BPNN model effectively addressed the verfitting problem in inverse parameter design, improving the effectiveness and feasibility of the parameter combinations. (4) The model was successfully applied in the Shenyang Metro Line 3 project. With target slump of 200 mm and permeability coefficient of 2.0×10^－5 cm/s as inputs, the model inversely derived conditioner parameter combinations that met error requirements. Field measurements aligned well with target values, and shield tunneling proceeded smoothly, validating the model’s effectiveness and practical value in real-world engineering.

Key words: shield muck conditioning, machine learning, parameter prediction, inverse design, support vector regression-back propagation neural network hybrid model