Abstract: In current earth pressure balance shield tunneling projects, the effectiveness of soil conditioning is commonly evaluated using slump tests or empirical judgment. A uniform standard is applied to assess the conditioning effectiveness for different soils, resulting in evaluation criteria that lack specificity and adaptability. Due to the lack of clear evaluation metrics and assessment method for soil conditioning effectiveness, it is difficult to quickly assess the conditioning results and identify deficiencies. This often leads to reduced construction efficiency and increasing engineering issues such as cutterhead clogging, surface settlement, and screw conveyor spraying. To address the aforementioned technical challenges, a self-developed indoor experimental apparatus is employed to investigate the properties of conditioned soil in sandy soil layers with varying clay content, including compressibility ratio, consistency coefficient, and yield shear strength. Based on indoor experimental results, a method combining cluster analysis with machine learning is proposed to evaluate the soil conditioning effectiveness by applying yield shear strength, slump, and compression ratio (200 kPa) as evaluation indicators. The proposed evaluation criteria classify conditioned soils into three categories, defining the characteristic distribution ranges for each type. The XGBoost algorithm is adopted to establish a predictive model for soil conditioning effectiveness. The selected input variables for the predictive model include bentonite slurry injection ratio, bentonite slurry concentration, particle size of unconditioned sandy soil, clay content, and foam injection ratio, yielding a prediction accuracy rate of 80%.

Key words: earth pressure balance shield, soil conditioning, machine learning, cluster analysis, classified evaluation