Identification and rectification of safety hazards are crucial for the safety management of a construction site. However, the absence of structured knowledge necessitates a heavy reliance on experienced managers or experts for construction safety management. This dependence on manual formulation of rectification measures results in inefficient safety management and underutilization of data value. This study proposes a knowledge graph based approach for the safety hazards inspections and recommendation of rectification measures, which consists of four steps: (1) constructing the semantic expression framework of safety hazard ontology for information extraction; (2) extracting entities from textual and semi-structured data using a large language model (ChatGLM-6B); (3) knowledge fusion and inference; (4) storing structured knowledge and developing semantic retrieval and safety hazard rectification measures recommendation. The proposed method was validated in a hydropower project, where a knowledge graph of safety hazards was constructed, consisting of 108,000 entities and 121 relationships. The results demonstrate that this method can automatically provide targeted rectification measures for safety hazards such as fire, falls from heights, scaffold collapse and electric shock. This work can provide a practical reference for improving the effectiveness of the construction project safety management.