A Transfer Learning Strategy for Rotation Machinery Fault Diagnosis based on Cycle-Consistent Generative Adversarial Networks

Abstract

Transfer learning has been an important aspect in recent years for labeled data are pretty rare in real application under different conditions. In modern industry systems, collected sample signals are usually not equally distributed, meaning the quantity of data from different working conditions are barely the same. Researchers have proposed a number of methods tackling the issue, most of which try to extract the features of the original data to unify them. Basic and valid algorithm is distribution adaption which include transfer component analysis (TCA), joint distribution adaptation (JDA), correlation alignment (CORAL) and other varieties. These methods have shown great effectiveness in practice. Generative adversarial networks (GANs) are newly developed generative models which can generate new sample data similar to original data through a special designed competitive training procedure. While distribution adaption unify the signal under all the conditions, GAN models are able to achieve approximate distribution functions and generate fake samples for different working conditions. In this paper, a new fault diagnosis transfer learning approach is proposed with a cycle-consistent GAN model. The designed GAN tries to generate new sample for unknown conditions based on known conditions and makes it possible to pre-train a classifier for fault diagnosis. Different experiments were carried out to demonstrate the performance of our proposed model, and other distribution adaption methods are compared. Experiments show that our strategy is superior to existing methods for fault diagnosis transfer learning.