Application of machine learning in caisson inclination prediction: model performance comparison and interpretability analysis

Abstract

This study combines data denoising techniques with TabPFN (Tabular Prior-data Fitted Network) model to address tilt prediction challenges in ultra-deep caissons. Using the Ligang Water Plant project as a case study, Savitzky-Golay filtering was applied for data denoising, and 611 samples were obtained through stratified sampling. Comparing nine machine learning algorithms, TabPFN demonstrated significant advantages, achieving R² values of 0.994 and 0.992 for east–west and north–south predictions with RMSE values of 10.34 mm and 9.51 mm respectively. Small-sample analysis revealed that TabPFN maintains superior performance with only 10 % training data, significantly outperforming traditional algorithms under data-scarce conditions. Feature dependency analysis identified key factors: sinking depth showed a critical turning point at 30–40 m stratum transition; soil elastic modulus exhibited larger SHAP (SHapley Additive exPlanations) values at higher values; and sinking rate remained stable at lower rates while high-speed sinking led to unpredictable tilt risks. This method avoids complex parameter tuning while demonstrating excellent small-sample learning capability, providing practical technical support for ultra-deep underground structure construction safety.