Data-driven remaining useful life prediction based on domain adaptation

• Published in: PeerJ. Computer science Vol. 7; p. e690
• Main Authors: Wen, Bin cheng, Xiao, Ming qing, Wang, Xue qi, Zhao, Xin, Li, Jian feng, Chen, Xin
• Format: Journal Article
• Language: English
• Published: San Diego PeerJ. Ltd 01.09.2021; PeerJ, Inc; PeerJ Inc
• Subjects: Adaptation; Algorithms and Analysis of Algorithms; Artificial Intelligence; Bidirectional Gated Recurrent Unit; Cable television broadcasting industry; Data collection; Data Mining and Machine Learning; Datasets; Deep learning; Domain Adversarial Neural Network; Domains; Life prediction; Literature reviews; Machine learning; Methods; Neural networks; Prognostics and Health Management; Remaining Useful Life; Sensors; Training; Transfer Learning; Useful life; United States
• ISSN: 2376-5992; 2376-5992
• DOI: 10.7717/peerj-cs.690

As an important part of prognostics and health management, remaining useful life (RUL) prediction can provide users and managers with system life information and improve the reliability of maintenance systems. Data-driven methods are powerful tools for RUL prediction because of their great modeling abilities. However, most current data-driven studies require large amounts of labeled training data and assume that the training data and test data follow similar distributions. In fact, the collected data are often variable due to different equipment operating conditions, fault modes, and noise distributions. As a result, the assumption that the training data and the test data obey the same distribution may not be valid. In response to the above problems, this paper proposes a data-driven framework with domain adaptability using a bidirectional gated recurrent unit (BGRU). The framework uses a domain-adversarial neural network (DANN) to implement transfer learning (TL) from the source domain to the target domain, which contains only sensor information. To verify the effectiveness of the proposed method, we analyze the IEEE PHM 2012 Challenge datasets and use them for verification. The experimental results show that the generalization ability of the model is effectively improved through the domain adaptation approach.