Complex domain extension network with multi-channels information fusion for remaining useful life prediction of rotating machinery. (1st June 2023)

• Record Type: Journal Article
• Authors: Cao, Yudong; Jia, Minping; Ding, Yifei; Zhao, Xiaoli; Ding, Peng; Gu, Liudong
• Is Part Of: Mechanical systems and signal processing. Volume 192(2023)
• Publication Date: 2023-06-01
• Subjects: Deep learning -- Complex domain extension network -- Multi-channels information fusion -- Remaining useful life prediction -- Rotating machinery; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2023.110190
• Languages: English
• ISSNs: 0888-3270

Abstract

Abstract: In the context of industrial big data, deep neural networks have been widely used in fault classification and remaining useful life (RUL) prediction of mechanical equipment due to their powerful nonlinear feature extraction capabilities. However, traditional deep learning models only stay at the level of real domain for feature mining, regardless of the importance of time–frequency domain analysis for rotating machinery. Furthermore, single-channel information from single source limits the ability of networks to learn mechanical degradation trajectories. To solve the above problems, complex domain extension network with multi-channels information fusion is proposed to realize RUL prediction of rotating machinery under different operating conditions. Specifically, the real domain network architecture is first extended to the complex domain, including complex domain convolution, complex domain batch normalization, complex domain weight initialization, complex domain parameter propagation, and complex domain activation functions. On this basis, complex domain extension network with multi-channels information fusion is proposed to extract degenerate features and finally establish an end-to-end mapping between feature information layers and prediction layer. The effectiveness of proposed RUL prediction framework is verified by two case studies with two sets of run-to-failure datasets. The comparison results with current state-of-the-art methods show that the proposed method is a promising method for remaining useful life prediction as its advantages in terms of prediction accuracy, interpretability, and generalization.