Abstract: Abnormal receiving and dispatching operations at railway stations are a key link in ensuring train safety. This paper designed and developed a question-answering system for operations of abnormal train receiving and dispatching at railway stations based on large language models to address the problems of fixed assessment modes, weak targeting, and difficulty in adapting to fragmented learning needs of employees in current training. It deployed a large language model and professional knowledge base in a local environment, used document vectorization processing, retrieval enhancement generation, and knowledge portrait construction techniques to implement dynamic intelligent testing paper generation, automatic evaluation and deep parsing, personalized training, and other functions, and constructed a closed-loop learning path of "learning assessment diagnosis improvement". The application results show that the system effectively improves the practical ability of operators in emergency response processes for abnormal train arrivals and departures, providing a reference for exploring new intelligent training models for railway workers.