Machine Learning-Enriched Lamb Wave Approaches for Automated Damage Detection

• Published in: Sensors (Basel, Switzerland) Vol. 20; no. 6; p. 1790
• Main Authors: Zhang, Zi, Pan, Hong, Wang, Xingyu, Lin, Zhibin
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 24.03.2020; MDPI
• Subjects: Algorithms; Artificial intelligence; Automation; Civil engineering; Classification; damage identification; data-driven approach; Datasets; Deep learning; Defects; Knowledge discovery; lamb wave; Machine learning; Methods; Neural networks; Noise; Physical restraints; Physics; Simulation; structural health monitoring; Support vector machines; Wavelet transforms; Lamb wave; structural health monitoring; damage identification; machine learning; data-driven approach
• DOI: 10.3390/s20061790
• ISSN: 1424-8220

Lamb wave approaches have been accepted as efficiently non-destructive evaluations in structural health monitoring for identifying damage in different states. Despite significant efforts in signal process of Lamb waves, physics-based prediction is still a big challenge due to complexity nature of the Lamb wave when it propagates, scatters and disperses. Machine learning in recent years has created transformative opportunities for accelerating knowledge discovery and accurately disseminating information where conventional Lamb wave approaches cannot work. Therefore, the learning framework was proposed with a workflow from dataset generation, to sensitive feature extraction, to prediction model for lamb-wave-based damage detection. A total of 17 damage states in terms of different damage type, sizes and orientations were designed to train the feature extraction and sensitive feature selection. A machine learning method, support vector machine (SVM), was employed for the learning model. A grid searching (GS) technique was adopted to optimize the parameters of the SVM model. The results show that the machine learning-enriched Lamb wave-based damage detection method is an efficient and accuracy wave to identify the damage severity and orientation. Results demonstrated that different features generated from different domains had certain levels of sensitivity to damage, while the feature selection method revealed that time-frequency features and wavelet coefficients exhibited the highest damage-sensitivity. These features were also much more robust to noise. With increase of noise, the accuracy of the classification dramatically dropped.