Abstract

Porosity and voids are a type of defect that may appear during the manufacturing of composite structures. Its possible presence is checked in production parts by different non-destructive techniques based on ultrasonic inspection. It is important to perform this evaluation  considering that a high content of these defects decrease the mechanical properties compared to the original design. In recent years, optical microscopy of composite materials has become a very powerful tool, which is  available for the industry to inspect and look at the internal structure and help to reach conclusions in the field of quality, failure analysis and I+D+I. In addition, the use of automatic scanning techniques of wide cross sections has opened the possibility of quantifying the percentage of porosity/voids, as well as become an alternative to more laborious testing procedures such as acid digestion of the resins. This study contributes to existing test standards within the aeronautical sector, in the sample preparation strategy. For this purpose, samples of epoxy matrices with carbon fibers, both unidirectional and fabric tapes have been analyzed to evaluate the influence of the quality of sample preparation, the stacking sequence and the initial porosity levels, the repeatability of the measurements and the quality of the results. An attempt has been made to correlate the results with standar physical / chemical analysis.

Abstract

This work presents the development of an image analysis software based on artificial intelligence for the detection and quantification of pore and void defects in carbon fiber reinforced polymer (CFRP) laminates. The tool aims to improve the inspection process of these materials, which is traditionally carried out through optical microscopy and manual analysis, methods that require significant time and effort from technical personnel. Machine learning and computer vision techniques, specifically the YOLOv8 instance segmentation model, were employed to automate the analysis of micrographs of fiber-reinforced composites (FRPs), including carbon and glass fibers. The results show a significant reduction in analysis time – from 30 minutes to just a few seconds – although the model’s accuracy is not yet sufficient for full implementation in a laboratory setting. While the results are acceptable in some cases, validation by qualified technicians remains necessary. This work is part of the ASSISTER project, funded by the Secretariat of State for Digitalization and Artificial Intelligence, under the Ministry of Economic Affairs and Digital Transformation, through the C005/21-ED call by red.es, and by the Andalusian Technological Corporation (CTA).