Knowledge graph and mitigation measures recommendation for safety hazards in large-scale hydropower projects using diverse heterogeneous inspection data

Large-scale hydropower project (LHP) sites are fraught with numerous construction safety hazards (CSHs). When the efficiency of addressing these CSHs falls to meet safety management requirements, accidents may occur. Existing inspection records of CSHs contain a wealth of useful information, yet these unstructured texts hinder their efficient utilization. Furthermore, current mitigation measures for CSHs largely depend on human experience, leading to low efficiency. To address these issues, this paper proposes a BERT-Att-BiLSTM-CRF model based on massive daily inspection data, achieving precise extraction of CSHs entities (F1 > 95 %); construction a multi-dimensional knowledge graph with nine entity types and eight relationships; a mitigation measures recommendation based on Sentence-BERT (SBERT) model demonstrates superior performance (Pearson = 0.92, Spearman = 0.85) through semantic similarity; for novel CSHs, a safety management standards-based semantic model recommends compliant solutions. Validation confirms the research results capability to automate safety knowledge extraction from unstructured texts, establishing a replicable paradigm for infrastructure risk management. • Develops a BERT-Att-BiLSTM-CRF model achieving >95 % F1-score for precise identification of CSHs entity. • Constructs a domain-specific knowledge network integrating 9 entity types and 8 relationships • A safety management standards-based semantic model addressing unseen hazards. • The knowledge graph can achieve update.