期刊
INTERNATIONAL JOURNAL OF FATIGUE
卷 163, 期 -, 页码 -出版社
ELSEVIER SCI LTD
DOI: 10.1016/j.ijfatigue.2022.107028
关键词
Ultrasonic Lamb waves; Fatigue crack; Deep learning; Long short-term memory; Spectral analysis; Noise reduction
资金
- National Research Foundation of Korea (NRF) - Korean Government (MSIP) [2017R1A5A1014883]
In this study, a hybrid method combining a deep learning model with high-erorder spectral analysis is proposed for reliable fatigue crack detection in noisy environments. The experimental results validate the improved performance of the proposed method.
Recently, the technique of nonlinear Lamb wave mixing has been developed for the detection of fatigue crack in engineering structures. In this technique, two or three Lamb waves with distinct frequencies are applied to a structure. The cross-mixing between these waves results in nonlinear mixed components depending on the sum and difference of the incident frequencies. However, the amplitude of the mixed components generated by the fatigue crack becomes weak in a noisy environment. Thus, noise elimination is critical for reliable crack detection. To address this gap, a novel hybrid method that incorporates a deep learning (DL) model with high-erorder spectral analysis is proposed in this study. First, a nonlinear Lamb wave mixing technique is developed to capture ultrasonic data from the aluminum plates during fatigue testing. Subsequently, the DL model based on long short-term memory (LSTM) accepts an original ultrasonic time signal as input and yields an output of a reconstructed ultrasonic signal after noise reduction. Finally, the random noise in the reconstructed signal is eliminated and the mixed components are extracted by trispectrum (TS)-based higher-order spectral analysis. Furthermore, the proposed and existing methods (e.g., power spectrum) are applied to the ultrasonic data collected from fatigue experiments. The results validated the improved performance of the proposed LSTM-TS method for reliable fatigue crack detection in noisy environments.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据