A Multisource Scale Subdomain Adaptation Network With Multihead Hybrid Self-Attention for Remaining Useful Life Prediction of Rolling Bearings
Hao Tian, Yan Han, Pingan Yang, Mi Zhu, Decheng Wu
Abstract
Recently, domain adaptation (DA) has gained widespread application in the prediction of the remaining useful life (RUL) of rolling bearings. However, existing network models of DA methods mostly use single-source domain data as input, which may cause the extracted degradation features to become unitary and the extracted degradation features are also not sufficient. Meanwhile, all of these issues may make the DA network with poor generalization ability and low prediction accuracy. Aiming at these problems, a multisource scale subdomain adaptation network (MSSAN) with multihead hybrid attention is proposed for RUL prediction of rolling bearings. First, a multihead hybrid attention mechanism is designed and embedded in the MSSAN, which weights different domains and helps the network to extract degradation features from different domains more adequately. Second, source domain features with different weights are fed into a multiscale extractor to achieve a comprehensive representation of degradation features. Then, through two-stage multisource DA, the domain distribution discrepancies are reduced between different source and target domains and also aligned the decision boundaries between regressors in multiple different domains. Finally, the experiment is carried out on the PHM2012 dataset, and contrast test results demonstrate that the proposed method has better cross-domain RUL prediction performance and prediction accuracy.