Objective Accurate prediction of shield attitude is crucial for the tunnel axis deviation control and guarantee of finished tunnel quality. Existing purely data-driven prediction methods fail to consider multivariate spatial dependencies, showing deficiencies in feature extraction capacity. Therefore, further research on shield attitude prediction methods is necessary.

Methods A PIFE-LSTM (physics-informed feature enhancement long short-term memory) model is proposed. Starting from the mechanism by which the shield thrust system influences attitude changes, the model extracts thrust vectors (including resultant thrust and thrust application point) in the feature engineering stage to characterize the spatial distribution of thrust forces. These thrust vectors are used as enhanced features to reinforce the physical orientation of the feature extraction process, and the LSTM model is used to capture the temporal dependencies of shield attitude. Measured data from the tunnel section between Longdong Avenue Station and Pudong Football Stadium Station of Shanghai Metro Line 21 are applied to the proposed model training and comparative verification.

Result & Conclusion  The introduction of thrust vectors effectively improves the model's capability to capture the nonlinear mapping relationships. Compared with the model without introduced thrust vectors, the determination coefficient increases by 0.06 and the mean absolute error decreases by 0.40 mm. Moreover, in comparison with other classical time-series prediction models, the PIFE-LSTM model, benefiting from its gating mechanism, effectively alleviates the vanishing gradient problem in long-sequence data processing and exhibits superior nonlinear fitting capability.