Integrated transformer approach for shield attitude prediction in marine soft soils considering small curvature effects

Accurately predicting shield attitude in marine soft soils remains challenging, particularly in small-curvature tunnel sections where excavation disturbs soil structure and causes nonlinear responses. Conventional data-driven methods often neglect soil properties and curvature effects, reducing reliability in complex ground conditions. This paper develops an integrated empirical mode decomposition (EMD)–principal component analysis (PCA)–Transformer framework that combines EMD for denoising, PCA for feature compression, and a Transformer encoder for sequence learning. Using field data from a shield tunnel excavated in marine soft soils, 32 parameters were analysed to quantify cross-dependencies. Incorporating marine soft soil parameters improved the prediction accuracy, reducing the RMSE of tail horizontal deviation by 46.15% and increasing R 2 from 0.85 to 0.97. Compared with five benchmarks, the proposed framework demonstrated superior accuracy and stability. The results highlight the importance of integrating soil behaviour and curvature constraints to achieve robust and transferable shield-attitude prediction for intelligent tunnelling in complex geological environments. • An EMD-PCA-Transformer framework is proposed for shield-attitude prediction. • Adding soil parameters cuts deviation error by 46% and raises accuracy to R 2 = 0.97. • The model outperforms five advances models with field settlement under 6.3 mm. • It enables robust prediction in complex geology by integrating soil and curvature effects.