A sparse domain adaption network for remaining useful life prediction of rolling bearings under different working conditions. (March 2022)

• Record Type: Journal Article
• Authors: Miao, Mengqi; Yu, Jianbo; Zhao, Zhihong
• Is Part Of: Reliability engineering & system safety. Volume 219(2022)
• Publication Date: 2022-03
• Subjects: Bearings -- Remaining useful life prediction -- Transfer learning -- Sparse domain-adversarial learning -- Convolutional neural network -- Selective kernel width; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2021.108259
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: Sparse domain adaption network (SDAN) is proposed for RUL prediction of bearings. A sparse domain-adversarial learning is developed to learn domain invariant features. Adaptively convolutional neural network is proposed to improve feature learning. The experimental results prove effectiveness of SDAN for RUL prediction. Abstract: As a key component in the machinery, the health of bearings directly affects working performance of machinery. Recently, many data-driven methods have been proposed to predict remaining useful life (RUL) of rolling bearings. However, most methods neglected the problem of data distribution difference caused by different operation conditions, which will lead to prediction performance deteriorating greatly on other bearings. To solve the domain shift problem in bearing RUL prediction, a sparse domain adaption network (SDAN) is proposed in this study. Firstly, an adaptive selection mechanism is proposed to select important input features in SDAN. Besides, a novel feature extractor, adaptively convolutional neural network (ACNN) is proposed to capture essential information from the selected features by adjusting receptive fields adaptively. The sparse feature selection layer is developed to suppress noise and remove ineffective features based on the noise filtering of sparse representation. Besides, the sparse domain adaption is used in SDAN by integrating domain-adversarial leaning and unsupervised sparse domain alignment to solve the problem of data distribution shift. Finally, the effectiveness of SDAN is verified on the PRONOSTIA rolling bearing dataset. The results demonstrate that SDAN can extract essential features and provide transferable RUL prediction performance under different working conditions.