Abstract: In the coal mine emergency rescue auxiliary decision system, there are problems such as insufficient application of emergency plans, low application efficiency, and poor execution of rescue plans generated by the system. In order to solve the above problems, a design method for a coal mine emergency rescue auxiliary decision system based on emergency plans is proposed. This method uses information extraction technology based on large language models to extract key task elements from emergency plans, such as task names, triggering conditions, executing departments, and task content. This method forms meta tasks, and constructs a meta task library that classifies and stores meta tasks based on accident types and levels. When a coal mine safety accident occurs, this method uses semantic matching technology based on the SBERT model to classify and grade the accident based on the information collected on site. The method selects the meta task set that matches the current emergency needs from the meta task library. To improve the feasibility of tasks, this method combines meta tasks with real-time collected on-site data, constructs specific action instructions through instruction templates. The method uses task planning techniques to optimize and adjust the priority of instructions, and generate practical and feasible on-site rescue plans. The coal mine emergency rescue auxiliary decision system based on emergency plans fully utilizes the standardized content of emergency plans, forming a rescue plan closely integrated with on-site information and resource optimization. The system further improves the accuracy, scientificity, and intelligence level of rescue decision-making.