Fault diagnosis method of rolling bearings under varying working conditions based on deep feature transfer

Abstract

Rolling bearing vibration data and their labels are difficult or impossible to obtain under varying working conditions. Thus, multistats identification of different fault positions and degradation degrees has relatively low accuracy. This paper proposes a fault diagnosis method based on deep feature transfer. Sparse denoising autoencoder extracts deep features of the frequency-domain amplitude sequences of rolling bearing vibration signals, and the features are used to compose feature sample sets of the source and target domains. Joint geometrical and statistical alignment adaptively processes feature samples of the source and target domains, and this way reduces the distribution divergence and subspace transform shift of the inter-domain feature samples. The classification is achieved using softmax. Experimental results show that the feature visualization effect by visualization algorithm t-SNE using the proposed method is better than those of other methods in this paper. A higher accuracy can also be achieved for rolling bearing fault diagnosis under varying working conditions.