Abstract: In order to improve the accuracy of the fault diagnosis of the RV reducer, a residual network is used to diagnose the faults of the RV reducer. The vibration signals of four kinds of fault mode and normal mode of the RV reducer are measured through the vibration test bench, the training and test data sets are constructed with those singals, and the data of the training set was augmented. After that, with data preprocessing, one-dimensional signal samples are converted into two-dimensional signal samples and are imported in the residual network for training and 5-folds cross validation. Then, the classification accuracy of the residual network is compared with that of DNN、LeNet and 10 layers CNN. The results show that the residual network is better than the traditional methods, the classification accuracy for the RV reducer reaches 98.11%. Furthermore, the western reserve university's bearing data set is used to verify the universality of the new model. Finally, LDA(Linear Discriminant Analysis) is used to reduce the dimension of the output of the average pooled layer of the residual network, and the relationship between the scattergram and the fault type of the RV reducer was studied.

Key words: deep learning, fault diagnosis, residual network, RV reducer, vibration signal