A data-driven and knowledge graph-based research on safety risk-coupled evolution analysis and assessment in shield tunneling

Abstract

Shield tunneling encounters critical safety challenges stemming from spatiotemporal risk coupling and insufficient utilization of accident data. This study developed a hybrid entity–relationship extraction model called BBi-MA-DCGCN by integrating bidirectional encoder representations from Transformer embeddings, bidirectional long short-term memory architectures, multihead attention mechanisms, and densely connected graph convolutional networks. Subsequently, a risk-coupled evolution knowledge graph was constructed to support intelligent inference and probabilistic reasoning of accident scenarios. Furthermore, an interaction matrix framework was implemented within the graph structure to quantitatively evaluate risk interdependencies. Results showed that (1) the BBi-MA-DCGCN model achieved an F1 score of 90.33%, demonstrating robust capabilities in entity and relationship extraction. (2) The coupling evolution reasoning method enabled the rapid inference of the most probable risk evolution path and potential accident types from any given risk node. (3) Among the top 25 key risk factors obtained through quantitative risk assessment, 52% were attributed to organizational management factors, with management, supervision, and training identified as having the most substantial effects. This study provides a novel method for understanding the spatiotemporal coupling evolution mechanism of safety risks in shield tunneling and enhances the accuracy of risk identification, dynamic inference, and quantitative evaluation.