A Novel Model of Bearing Fault Diagnosis Based on Convolutional Auto-encoder and Capsule Network

Abstract

A convolutional auto-encoder and capsule network based fault identification method (CaCAE) for rotating machinery is proposed. Firstly, the convolutional auto-encoder is constructed, to quickly and simply extract shallow-level features from the raw fault data without a priori knowledge. Then, the improved capsule network accepts the feature information extracted by the convolutional auto-encoder, and extracts higher-level feature capsules to represent the health status by a dynamic routing algorithm. The feature capsules can enhance the information mining capability of the model and improve the classification accuracy. Finally, the lengths of the capsules are used to classify the fault category. To verify the effectiveness of the proposed method, an example with 100% classification accuracy and 38s running time was obtained on a rolling bearing test bench. In addition, the performances of artificial neural network, classical convolutional neural network and the original capsule network are evaluated for contrast with the proposed method. In conclusion, the CaCAE obtains higher classification accuracy and faster convergence speed.