Damage characterization using CNN and SAE of broadband Lamb waves

Fei Gao; Jiadong Hua

doi:10.1016/j.ultras.2021.106592

Damage characterization using CNN and SAE of broadband Lamb waves

Ultrasonics. 2022 Feb:119:106592. doi: 10.1016/j.ultras.2021.106592. Epub 2021 Sep 21.

Authors

Fei Gao¹, Jiadong Hua²

Affiliations

¹ School of Reliability and Systems Engineering, Beihang University, Xueyuan Road No. 37, Haidian District, Beijing, China; Advanced Manufacturing Center, Ningbo Institute of Technology, Beihang University, Ningbo 315100, China; Aero-engine Thermal Environment and Structure Key Laboratory of Ministry of Industry and Information Technology, Nanjing 210016, China. Electronic address: youfeigao@buaa.edu.cn.
² Aero-engine Thermal Environment and Structure Key Laboratory of Ministry of Industry and Information Technology, Nanjing 210016, China; Science & Technology on Reliability and Environmental Engineering Laboratory, Beihang University, Xueyuan Road No. 37, Haidian District, Beijing, China. Electronic address: huajiadong@buaa.edu.cn.

PMID: 34583152
DOI: 10.1016/j.ultras.2021.106592

Abstract

The method based on Lamb wave shows great potential for structural health monitoring (SHM) and nondestructive testing (NDT). Deep learning algorithms including convolutional neural networks (CNN) and stack autoencoder (SAE) are promising to extract features from Lamb wave signals that can be linked with damage for subsequent localization and quantification. Generally, narrowband Lamb wave with purified mode and suppressed dispersion is used because of clear relationship model between damage features and recorded signals. However, model performance is limited because contained damage information of narrowband Lamb wave is inadequate. To overcome this limitation, a broadband Lamb wave deep learning algorithm is proposed for damage localization and quantification. Compared with narrowband, broadband Lamb wave generated at a large frequency range contains richer information of structural damage. In this study, different mode selections, different signal processing methods and different deep learning algorithms are applied to extract damage features from different perspectives, and fusion of all extraction results facilitates the full utilization of rich broadband information. An experiment is given to demonstrate the effectiveness and high-accuracy of proposed method.

Keywords: Convolutional neural networks; Damage quantitativeevaluation; Lamb waves; Stack autoencoder.