Comparative study on machine learning prediction of cutter chamber pressure in small-sample scenarios

Accurate regression prediction of chamber pressure during shield tunneling in water-rich sandy gravel layers is a significant challenge due to strong non-linearity. Furthermore, small sample sizes often lead to arbitrary model configuration and insufficient accuracy. To address these bottlenecks, this study develops a comprehensive prediction scheme based on 602 sets of construction data. The scheme encompasses optimization, algorithm selection, and performance enhancement. Specifically, a multi-algorithm weighted fusion strategy, named FC7, was proposed by integrating six conventional feature extraction algorithms. The study also constructed a library of 12 algorithm categories, including linear models, tree-based models, and ensemble algorithms. This approach helped to identify the most suitable model for small datasets. The results demonstrate that the algorithm-driven FC7 feature combination offers optimal performance. It significantly enhances feature capture efficiency and accuracy. Tree-based ensemble learning algorithms (CatBoost, LightGBM, and Random Forest) proved most effective for this task. Among these, CatBoost achieved the highest performance. After optimization, the test set Coefficient of Determination reached 0.953, and the Root Mean Square Error was reduced to 0.043. Moreover, total thrust force and articulation jack stroke difference were identified as key control features, exhibiting clear non-linear interaction patterns. Based on these findings, a closed-loop prediction framework of “Data-Model-Decision-Update” is proposed. This strategy effectively overcomes the limitations of small-sample prediction for chamber pressure and provides a technical reference for similar engineering tasks.