@article{WANG2025, 
author = {Jinhua WANG and Rui LIU and Jie CAO},
title = {Unlabeled data fault diagnosis method based on multi-domain adaptation},
year = {2025},
journal = {Journal of Beijing University of Aeronautics and Astronautics},
volume = {51},
number = {4},
pages = {1185-1194},
keywords = {rolling bearing, fault diagnosis, transfer learning, multi-domain adaptation, pseudo-label strategy},
url = {https://www.sciopen.com/article/10.13700/j.bh.1001-5965.2023.0166},
doi = {10.13700/j.bh.1001-5965.2023.0166},
abstract = {in industrial production, due to the difference in the distribution of source domain data and target domain data and the small amount of labeled fault data, the accuracy of domain adaptation-based bearing fault diagnosis algorithms proposed in the past is generally not high. In view of this, the multi-domain adaptation neural network (MDANN) fault diagnosis method was proposed in this paper, which was used for rolling bearing fault diagnosis without labeled data. Firstly, the original vibration signal was processed by using wavelet packet transformation (WPT) to reduce signal redundancy and avoid the loss of key signal features. Secondly, the multi-kernel maximum mean discrepancy (MK-MMD) algorithm was used to calculate the difference of input eigenvalues, and the network parameters of MDANN were updated by backpropagation so that the network can extract domain invariant features. Finally, in order to ensure that unlabeled target domain data can participate in network training normally, the maximum probability label was used as a pseudo-label strategy of the real label to solve the problem that unlabeled target domain data cannot be trained and enhance the acquisition of reliable diagnosis knowledge of the model. Two publicly available datasets, CWRU and PU, were used for validation. The experimental results show that the proposed method has higher diagnosis accuracy compared with common domain adaptation methods, which further shows that the method can effectively learn the transferable features and fit the discrepancy in data distribution between the two datasets.}
}