A renewable fusion fault diagnosis network for the variable speed conditions under unbalanced samples

Deep learning technology has been gradually applied to solve a variety of fault diagnosis problems because of its outstanding feature learning and nonlinear classification abilities. However, few deep learning network models can be applied to both variable speed conditions and unbalanced samples scenarios in fault diagnosis, especially in extreme cases where fault samples are missing. And most of the fault diagnosis models do not have the ability to update automatically as the collected fault data increases. To deal with the above problems, a deep learning model named renewable fusion fault diagnosis network (RFFDN) is proposed. The network has three main parts: improved feature classification network; second order statistics fusion network and unbalanced feature comparison network. Moreover, these three networks are simultaneously organized on a two-branch convolutional neural network (CNN) architecture with fused data input, so as to facilitate the network to learn the depth nonlinear domain invariant features. Finally, the RFFDN model and other mainstream fault diagnosis models are tested on two different datasets. The results show that the RFFDN model not simply achieves high diagnostic accuracy in diagnosis results, but also extracts the domain invariant features at variable speed conditions under unbalanced samples, and accurately classifies the new faults. These prove that the model can not only be applied to a variety of operating modes, but can be updated as more data are collected as well, which is of great significance to the field of fault diagnosis. (C) 2019 Elsevier B.V. All rights reserved.