Abstract: This paper introduces Bayesian techniques from the machine learning field into the application of power plant accident diagnosis in nuclear emergency. A new approach for plant accident diagnosis based on Naive Bayes Classifier is proposed. The PWR nuclear power plant simulator is used to obtain accident case data, and the naive Bayes classification model is trained to realize the diagnosis of multiple types of accidents (LOCA, SGTR, MSLB) in nuclear power plants. The test results show that the accident diagnosis methods based on naive Bayesian classifier have significant advantages in diagnostic accuracy, diagnostic efficiency, expandability of accident types and program autonomy diagnosis. It is found the prior distribution of different accident types in model training has little influence on the training performance, indicating the good applicability of the new approach.

Key words: nuclear emergency, accident diagnosis, Naive Bayes, machine learning