An Adaptive Sparse Graph Learning Method Based on Digital Twin Dictionary for Remaining Useful Life Prediction of Rolling Element Bearings

• Published in: IEEE transactions on industrial informatics Vol. 20; no. 9; pp. 10892 - 10900
• Main Authors: Cui, Lingli, Wang, Xin, Liu, Dongdong, Wang, Huaqing
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.09.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; Adaptive sparse graph learning (ASGL); Data models; Degradation; Dictionaries; digital twin dictionary (DTD); Digital twins; Feature extraction; Learning; Life prediction; Parameters; Performance degradation; Predictive models; Regularization; Regularization methods; remaining useful life (RUL); Roller bearings; rolling element bearings; Topology; Useful life
• ISSN: 1551-3203; 1941-0050
• DOI: 10.1109/TII.2024.3399882

The remaining useful life (RUL) prediction of rolling element bearings is usually subject to the following limitations. First, it is difficult to obtain the massive performance degradation data, which resulting in the insufficient learning of the historical degradation law. Second, the parameters in most of existing models depend heavily on the manual selection, which leads to the poor generalization performance. To address these problems, a novel adaptive sparse graph learning (ASGL) method based on digital twin dictionary (DTD) is proposed in this article. To facilitate the prediction when the data are insufficient, the extended exponential models and the extended linear piecewise models are first established, then a DTD that covers the various degradation behaviors is constructed. Besides, a new objective function of graph learning is designed and the sparse regularization method is introduced to adaptively obtain the topology graph of data. Therefore, the method avoids the wrong adjacency relationship caused by inappropriate parameters. The simulation and experimental results show that the DTD has higher prediction accuracy than the experimental samples, and the ASGL method is easy to implement and has lower dependence on the parameter selections. In addition, compared with some state-of-the-art methods, it can obtain better RUL prediction results.