Spatiotemporal deep learning for multi-attribute prediction of excavation-induced risk

This paper presents a hybrid deep learning model named the Online Learning-based Multi-Attribute Spatial-Temporal Transformer Network (OMSTTN) to predict excavation-induced risks during foundation pit excavation. OMSTTN integrates a hybrid Transformer offline model with a parallel embedding layer to process diverse monitoring attributes and employs a Spatial-Temporal Transformer block to capture complex spatiotemporal correlations. An online learning mechanism enables dynamic adaptation to evolving conditions, enhancing prediction accuracy. Validated on a real-world XuZhou Rail Transit project, OMSTTN achieves strong prediction performance (MAE: 0.0461, RMSE: 0.0699, R2: 0.9441). Comparative experiments demonstrate its effectiveness in handling multi-attribute data, dynamic changes, and spatiotemporal patterns. In short, OMSTTN narrows the research gap by providing a spatiotemporal framework for accurate risk prediction, offering significant potential for early risk detection and proactive management in excavation engineering.