Central moment discrepancy based domain adaptation for intelligent bearing fault diagnosis

In recent years, deep learning based bearing fault diagnosis is developing rapidly due to the increasing amount of industrial data. However, two major issues limit the application for deep learning: a) labeled data is difficult to obtain, and a lot of unlabeled data is more common in actual industrial production; b) the distribution of training and testing dataset will be different under different production environments or operations, which makes it difficult to generalize the trained model to another working condition. To solve these issues, we propose a domain adaptation convolutional neural network to diagnostic fault using Central Moment Discrepancy (CMD). In the proposed method, a convolutional neural network is applied to extract features from two differently distributed raw vibration signals, and the distribution discrepancy is reduced using CMD criterion. The proposed method can extract features with similar distribution from two different domains and make fault diagnosis for unlabeled data. The proposed method is proved to be effective in using CWRU dataset and Paderborn dataset under different working conditions. (C) 2020 Elsevier B.V. All rights reserved.