Reply to Comment on ‘A novel transfer learning bearing fault diagnosis method based on multiple-source domain adaptation’

Abstract Fault diagnosis methods based on a data-driven approach have achieved considerable attention in recent years. It is assumed that there are enough supervised data to set a responsible model during training, which is contrary to engineering. The machine normally runs most of the time and rarely runs in a faulty state. It is challenging to collect data from different working conditions. To address this, a transfer learning fault diagnosis method based on multiple-source domain adaptation is proposed in this article. A multiple-domain-adaptation learning strategy is adopted to decrease the distribution discrepancy between source and target domains. The maximum mean discrepancy and joint maximum mean discrepancy are applied in the fully connected layer to align the distribution and reduce the feature and label space discrepancy, and the discrepancy loss function is used to decrease the difference between diverse classifiers. Also, the learning process of the convolutional neural networks model and the effects of different loss functions are presented. Finally, two different bearing experiment datasets are introduced to show the performance of the proposed approach.