Manifold learning and segmentation for ultrasonic inspection of defects in polymer composites

Abstract

Non-destructive ultrasonic testing is beneficial for monitoring the structural health of polymer composites. However, owing to scattering and other factors, ultrasonic data often appear as noisy signals or images containing artifacts. The analysis of ultrasound signals highly depends on the expertise of trained human inspectors. Hence, the development of ultrasonic data analysis methods, particularly unsupervised methods, is necessitated. In this study, a novel unsupervised method is developed for the ultrasonic inspection of defects in polymer composites, named manifold learning and segmentation. In a uniform manifold approximation and projection model, nonlinear dimensionality reduction is first performed on high-dimensional ultrasound data for extracting and visualizing defect features. Subsequently, semantic segmentation is performed to predict/discriminate between defects and backgrounds. Consequently, subsurface defects in the composites can be effectively detected. Experimental results and comparisons on two carbon fiber reinforced polymer specimens demonstrate the effectiveness of the proposed method.