In order to make full use of the information contained in the massive data and effectively identify the bearing faults， the cloud theory method is used to map the bearing fault data to its corresponding fault type， and the cloud distribution model of each feature of the rolling bearing under different states is established. Based on this， a cloud judgment knowledge base of bearing faults is constructed. Meanwhile， the Relief-F method is introduced to determine the weight coefficients of each feature of the training set. Combined with the cloud distribution membership coefficient， the final membership calculation method for bearing faults is proposed. Through the comparison of classification accuracy between cloud classification knowledge bases established by different numbers of training samples， it is proved that the method has the ability to learn data. Moreover， the classification method and other traditional classification methods are tested by using noise-containing test samples， and the results show the superiority of the classification method in terms of noise resistance.