[Objective] To address the problems of large sample imbalance and low diagnostic efficiency in traditional hydropower station monitoring and alarm event diagnosis， as well as the lack of interpretability and difficulty in ensuring diagnostic accuracy in artificial intelligence-based diagnostic methods， a deep learning framework based on a large language model （LLM） for hydropower station alarm event sample enhancement and fault diagnosis is designed. Moreover， regular expressions based on expert knowledge are introduced to verify the alarm diagnosis results and improve their credibility. [Methods] First， considering the issue of imbalanced monitoring of event samples in a hydropower station， the SimBERT model is used to increase the amount of data for types with fewer samples. The enhanced data are then mixed with the original data to form the input samples of the deep learning network. Second， based on power grid monitoring alarm rules， regular expressions for typical faults are constructed to achieve accurate discrimination of typical faults. Finally， based on the ERNIE LLM， word vector embeddings of hydropower station monitoring events are constructed and input into the hierarchical attention network to learn the features of each event. The classification results are then produced. The final fault diagnosis results are then obtained by verifying the results produced using regular expressions. [Results] The test results show that， compared with traditional monitoring alarm event diagnosis methods for hydropower stations， this model can improve the diagnostic accuracy by more than 2%， reduce the model training time by 30%， and achieve diagnosis within 0.1 seconds. [Conclusions] The proposed deep learning framework based on an LLM for enhancing alarm event samples and fault diagnosis of hydropower stations can achieve fast and accurate results during system failures， which is conducive to ensuring long-term stable and safe operation of the power system.