Traditional methods for analyzing superconduction radio-frequency (SRF) cavity faults rely on a priori knowledge, featuring high labor and time costs, poor accuracy and low consistency. They are less efficient when dealing with complex devices and large amounts of data. In this paper, a method for classifying SRF cavity faults based on machine learning algorithms is investigated. Using the SRF cavity fault data generated during the operation of BEPCII, the classification of SRF cavity faults is achieved through image information extraction, feature selection and optimization, machine learning algorithm training, and analyzing the accuracy and consistency of the model via K-fold cross validation. The results of the study show that Random Forest, Support Vector Machine and Bagging algorithms have better classification results when dealing with faulty pictures. The accuracy and consistency of supervised learning methods are significantly higher than those of unsupervised learning. The fault classification realized in this research achieves high accuracy and consistency. It enables to quickly and efficiently distinguish the faults occurring on the SRF cavity of BEPCII. It also provides a reference for the diagnosis of SRF cavity faults in other particle accelerators.