Abstract: The artificial intelligence technologies currently used in nuclear power plant accident diagnosis, such as neural networks, are difficult to have both robustness and interpretability. The nuclear power plant accident diagnosis method based on Bayesian classifier was proposed in the paper, and the Bayesian classifiers were further refined into discrete naive Bayes classifiers, Gaussian naive Bayes classifiers and Bayesian networks. The performances of three Bayesian classifiers used in nuclear power plants accident diagnosis were compared. The analysis results show that the diagnosis method based on Bayesian classifier which combines knowledgedriven and datadriven has strong robustness and interpretability. Among the three classifiers, the Gaussian naive Bayesian method diagnosis has significant advantages in diagnosis accuracy, diagnosis efficiency, diagnosis accuracy of the size of the accident break, and types of accidents that can be diagnosed.