Fault diagnosis of smart grid protection system is a key link to ensure the safe and stable operation of power grid. In this paper, a new method for fault diagnosis of smart grid protection system based on the fusion of improved deep Q-learning and knowledge graph is proposed, which effectively fuses the adaptive learning capability of DQN and the structured information of knowledge graph, and significantly improves the performance of fault diagnosis. The improved DQN adopts a dual network structure and a prioritized experience playback mechanism, which effectively improves the learning efficiency and stability. After that, the knowledge graph is introduced to provide rich domain expert knowledge for the model, which enhances the interpretability and generalization ability of the model. Finally, an attention-assisted multi-task learning framework is introduced to achieve the co-optimization of fault type identification, localization and severity assessment. The test results show that compared with traditional DQN and other methods, the method proposed in this paper achieves significant improvement in diagnostic accuracy, efficiency and interpretability, and provides new ideas and methods for the development of intelligent diagnostic technology for smart grid protection systems.