Abstract: In shield tunneling, traditional data-driven models struggle to achieve accurate predictions due to the sparsity of transverse surface settlement monitoring data. To address this challenge, a transverse surface settlement prediction method based on a multi-fidelity residual neural network (MRNN) is proposed. This approach integrates multi-source data, including theoretical analytical solutions, finite element simulations, and engineering measurements, to construct a multi-level nested residual learning framework. First, the feasibility of the MRNN architecture is validated using low-fidelity (analytical solution) and mediumfidelity (finite element simulation) data. High-fidelity (measured) data are then introduced to train the residual model, progressively refining the prediction results. Experimental results reveal the following: (1) The low- and medium-fidelity models achieve a coefficient of determination (R²) of 0.99 on the finite element data test set, confirming the model’s feasibility; (2) After measured data are incorporated, the MRNN model achieves an R² of 0.91 and mean absolute error of 0.76 mm on engineering measured data, significantly outperforming the single neural network model (R²=0.34), thus highlighting the importance of varying fidelity data; (3) Compared with other commonly used models, such as random forest, support vector regression, and extreme gradient boosting(R² of 0.70, 0.48, and 0.33, respectively), the MRNN model yields superior results. The MRNN model shows distinct advantages in scenarios with small sample sizes. These results confirm that the MRNN framework can effectively integrate data sources of varying accuracy and cost, leveraging large amounts of low-cost theoretical data to capture global trends while employing limited high-precision measured data for local refinement. This approach offers a new technical pathway for achieving high-precision predictions of surface settlement in shield tunneling, even under conditions of limited sample data. 

Key words: shield tunneling, transverse surface settlement, multi-fidelity modeling, residual neural network, multi-source data fusion