Prediction of shield tunneling attitude: a hybrid deep learning approach considering feature temporal attention

Abstract

Abstract Accurate prediction of shield attitude deviation is essential to ensure safe and efficient shield tunneling. However, previous studies have predominantly emphasized temporal correlation, which has limitations in engineering guidance and prediction accuracy. This research proposes a hybrid deep learning approach considering feature temporal attention (FTA-N-GRU) for shield attitude prediction. To elucidate the contributions of each parameter, the Integrated Gradients algorithm is leveraged for parameter sensitivity analysis. The results from the Bangladesh Karnaphuli River Tunnel Project indicate that: the proposed model outperforms other models in prediction accuracy. The integration of feature attention can adaptively allocate attention weights to input parameters, facilitating inexperienced operators in discerning crucial parameter variations and decision-making. By incorporating temporal attention, the model effectively explores the connection among different output time steps, improving overall prediction accuracy and reliability. Consequently, operators are empowered with timely information to proactively adjust operations before deviations occur, underscoring the significance of this approach in promoting safe and efficient shield tunneling practices.