Answers are grounded in this paper and cite their sources.
Compares CNN performance in classifying CFRP damage, suggesting automated methods can enhance detection accuracy.
Carbon fiber-reinforced polymers (CFRP) are particularly vulnerable to low-velocity impact loads that can lead to material degradation like cracks and delaminations. Due to the difficult visual detectability of these damages, ultrasonic testing has been established for many decades as the standard test method. Usually, the interpretation of the ultrasonic C-Scans is performed manually which requires a high level of operator qualification and experience. Therefore, this study compares different convolutional neural networks (CNN) for the classification of damaged and undamaged ultrasonic C-Scans of CFRP-specimen. For the required dataset, in-house ultrasonic C-Scans were compiled and categorized into three different classes: low-velocity impact damage, “non-impact” damage and undamaged CFRP-material. Before use in a CNN, the dataset was edited with an automated image processing to reduce inhomogeneities. The final dataset was first attached to various pretrained networks. Further it was used for the training of a self-generated model. The adapted and pretrained models outperformed the self-generated one with an accuracy of up to 92,5%. Due to the fact, that the self-generated model offers more possibilities for fitting the architecture and the extracted features exactly to the given classification task it is still preferred.
No takes yet. Share an insight, caveat, or question.
Wolfrum et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: