A Novel Hybrid Deep Learning for Attitude Prediction in Sustainable Application of Shield Machine

Accurate prediction of the shield attitude is critical for controlling the excavation direction, ensuring construction safety, and advancing the sustainability of shield tunneling by reducing energy and environmental disturbance. Traditional prediction methods for the shield attitude have a certain lag and low prediction accuracy, and existing machine learning methods lack research on the varying importance of different parameters affecting the shield attitude, while also ignoring the global information characteristics of the data. To accurately predict the shield attitude and support sustainability-oriented operations, this study proposes a novel prediction model based on a project in Shenyang, China. The model utilizes a channel domain attention mechanism to learn the importance of various influencing parameters and extracts spatial features via a convolutional neural network. Additionally, it captures long-range dependency and local temporal features using a transformer augmented with a bidirectional long short-term memory network. Experimental results show that the proposed model achieves lower MAE and RMSE and higher R 2 compared with baseline and sub-models. Its generalization and reliability are further validated using data from another shield tunnel section. From a sustainability perspective, timely and high-fidelity predictions enable proactive steering that reduces unnecessary corrective actions and extreme operating states (e.g., thrust/torque spikes), which are associated with higher energy use, accelerated consumable wear, over-grouting, and potential surface disturbance. Finally, integrating the model’s predictions with onsite adjustment measures effectively mitigates alignment deviations, contributing to more energy-efficient, resource-conscious, and low-disturbance trajectory control.