A fast sparsity-free compressive sensing approach for vibration data reconstruction using deep convolutional GAN. (1st April 2023)

• Record Type: Journal Article
• Authors: Dong, Guan-Sen; Wan, Hua-Ping; Luo, Yaozhi; Todd, Michael D.
• Is Part Of: Mechanical systems and signal processing. Volume 188(2023)
• Publication Date: 2023-04-01
• Subjects: Compressive sensing -- Generative adversarial networks -- Deep learning -- Vibration data; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2022.109937
• Languages: English
• ISSNs: 0888-3270

Abstract

Abstract: Vibration data from physical systems, such as civil structures and machinery, often carries important information about the dynamic characteristics, but streaming acquisition of higher-frequency vibration often accrue large volumes of data, resulting in data transmission and storage challenges. Compressive sensing (CS) is a relatively newly-developed technique for efficient data representation, capable of reconstructing the target signal using only a few random measurements through sparse optimization. However, the real-world application of CS is hindered by the strong assumption of signal sparsity and a costly reconstruction process. In this work, we propose a novel deep learning method for vibration data reconstruction by using deep convolutional generative adversarial networks (DCGAN), which is composed of a generator G and a discriminator D . A modified 1D symmetric U-net architecture with shortcuts is presented for G to flexibly deal with different inputs, while a typical 1D classifier is used as D . A composite adversarial loss function is proposed considering errors in both time and frequency domains. The proposed DCGAN approach has several appealing properties. First, it directly learns the end-to-end mapping between the compressed and original signals without employing the sparsity assumption or random sampling, which fundamentally differs from existing sparsity-based CS methods. Second, the reconstruction process is highly computationally efficient as the network is fully feed-forward and no optimization is needed during data reconstruction. The proposed DCGAN approach is evaluated using the simulation data from a numerical 9-floor frame as well as experimental data collected from a large test steel grandstand. The results demonstrate the superiority of the proposed DCGAN in computational accuracy and efficiency compared to the tested sparsity-based algorithms. Furthermore, the influences of network configurations (network depth, down-sampling strategy, and shortcuts) are comprehensively explored. Highlights: A novel CS approach is proposed for vibration data reconstruction using DCGAN. The DCGAN can reconstruct vibration data without employing sparsity assumption. The proposed feed-forward network is highly computationally accurate and efficient. The DCGAN is evaluated using simulated and experimental data, respectively. The influences of network configurations are comprehensively explored.