Mining Causal Chains for Tower Crane Accidents Using an Improved Transformer and Complex Network Model

Tower crane structural failures remain a major safety concern on construction sites. To improve accident prevention, this study proposes an intelligent framework that combines an improved Transformer model with a Directional Interest Score (DIS) Apriori algorithm and complex-network analysis. A corpus of 535 tower crane accident reports (2002–2024) was compiled and annotated with causal and accident entities according to system–safety theory. Segment embeddings were introduced to the Transformer to reinforce boundary detection, enabling accurate extraction of causative factors and relation triples. The DIS-Apriori algorithm was then used to mine both positive and negative association rules while aggressively pruning irrelevant item sets. Eventually, causative factors were mapped into a weighted, directed complex network where edge weights reflect the absolute frequency difference between positive and negative rules, and edge directions correspond to their signs. Experiments show that the Transformer achieves higher precision and recall than baseline models, and DIS-Apriori substantially reduces unnecessary item-set complexity while preserving critical rules. Network analysis revealed five critical causal links and a closed-loop causal link that warrant priority intervention. The proposed method delivers a data-driven, explainable tool for pinpointing key risk sources and designing targeted mitigation strategies, offering practical value for intelligent safety management of tower cranes.