Abstract: Most current knowledge graph-based analyses of coal mine accidents remain at a relatively static modeling level, and insufficient consideration is given to the process by which accident causation changes over time. Introducing temporal factors into the modeling process and organizing accident knowledge in chronological order can help reveal changes in the structure of accident causation at different stages and allow the transmission process of key risks in the system to be observed more intuitively. A total of 82 coal mine accident reports from 2017 to 2025 were used as the main data source to construct a semantic model of coal mine accident knowledge elements integrating dynamic features. The BERT-BiLSTM-CRF model was used to automatically extract accident causation knowledge, and a dynamic knowledge graph of coal mine accidents was constructed. Ucinet software was used to construct the co-occurrence matrix of accident causation and calculate complex network analysis indicators, and information on the co-occurrence networks of accident causation in various periods was obtained. The evolutionary characteristics of accident causation knowledge were analyzed from the perspectives of network structure, nodes, risk-cause mapping, and paths. The results showed that the structure of the accident causation network evolved, with a dispersed structure gradually becoming more tightly coupled. Accident causation evolved from the behavioral operation level to the institutional and system level. Management, technical, and behavioral risks showed significant differentiation in exposure degree and transformation capacity. Shorter key accident causation paths in the early stage gradually evolved into cross-level and multi-link transmission chains.