Ultrasonic detection of porosity in composites based on wavelet packet transform and convolutional neural network

Abstract

Carbon fiber reinforced polymer composites (CFRPs) are widely used in many applications, while the pores have a significant influence on mechanical performance as a critical defect. In this work, wavelet packet transform (WPT) method is utilized as an effective feature extraction method to capture the information about pore defects in ultrasonic A-scan signals. Given the large amount of A-scan data and the spatial distribution of pores within CFRPs, A-scan signals are randomly extracted from multiple regions within the specimen to ensure a comprehensive representation of the material’s porosity. A multi-scale features obtained by this method not only compress the data volume but also reflect the details and variations of the pore’s distribution. These features are used as inputs to a convolutional neural network (CNN) for porosity classification. The experimental results showed that the method based on the combination of WPT and CNN can effectively distinguish the samples with different porosities with an accuracy as high as 98%. The results showed a promising application for determining the porosity of composites.