A physics-constrained dictionary learning approach for compression of vibration signals. (15th May 2021)

• Record Type: Journal Article
• Authors: Lu, Yanglong; Wang, Yan
• Is Part Of: Mechanical systems and signal processing. Volume 153(2021)
• Publication Date: 2021-05-15
• Subjects: Compressed sensing -- Dictionary learning -- Sparse coding -- Data compression -- Rotating machinery; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2020.107434
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: Measurement and basis matrices are trained separately. Physical constraints of sampling rates and time stamps are incorporated. New constrained FrameSense and adaptive K-SVD algorithms are developed. Reconstruction errors are significantly reduced compared to traditional dictionary learning. Abstract: Monitoring the health condition of rotating machinery in manufacturing systems usually requires vibration signals to be continuously collected, transmitted, and stored. The available bandwidth in communication channels for transmission of a large amount of data is limited in an industry setting. Therefore, reducing the amount of data in communication and storage without sacrificing the amount of information collection is necessary. Here, a new technique called physics-constrained dictionary learning is proposed to reduce the volume of data in storage and communication using compressed sensing. In compressed sensing, the original signals can be reconstructed with a much smaller amount of data determined by a measurement matrix, if the representation of signals in the reciprocal space is sparse. The proposed physics-constrained dictionary learning approach optimizes the measurement and basis matrices simultaneously to improve the accuracy of reconstruction, where physical constraints of time stamps of sampling and sampling intervals are considered. New training algorithms are developed. The proposed scheme is applied to compress the vibration signals of roller bearings. It is shown that the reconstruction performance of the proposed scheme is significantly improved from traditional dictionary learning.