Torque Prediction Model of TBM Based on Modal Decomposition and Residual Learning

Cutterhead torque is a critical parameter for characterizing the mechanical state of a shield machine, the degree of ground disturbance, and the associated excavation risks. Accurate prediction of the torque is essential for enhancing intelligent tunneling control and ensuring construction safety. However, torque signals typically exhibit pronounced non-stationarity, multi-scale oscillations, and time-varying behavior driven by geological changes, making it difficult for conventional deep learning models to capture their complex dynamic characteristics effectively. To address these challenges, this study proposes a torque prediction method that integrates Variational Mode Decomposition (VMD) with residual learning. The method first utilizes VMD to decompose the raw torque sequence into intrinsic mode functions (IMFs) across different frequency bands, yielding more stable and structured multi-scale subsequences. Subsequently, an LSTM-based residual prediction structure is constructed to achieve high-accuracy reconstruction of the original torque signal. Engineering case studies demonstrate that the proposed method reduces the error in both RMSE and MAE metrics by approximately 20%–30% compared with baseline models, and exhibits higher accuracy and stability in both trend tracking and local oscillation prediction. The results indicate that the approach effectively handles the noise and non-stationary characteristics of TBM torque signals, offering a technically sound and practically valuable solution for intelligent monitoring and decision-making in shield tunneling operations.