Thielen N, Braun J, Rachinger B, Franke J, Risch F, Reinhardt A (2025)
Publication Language: English
Publication Type: Conference contribution
Publication year: 2025
Publisher: Institute of Electrical and Electronics Engineers Inc.
Conference Proceedings Title: 2025 Pan Pacific Strategic Electronics Symposium (Pan Pacific)
ISBN: 979-8-3315-1198-2
DOI: 10.23919/PanPacific65826.2025.10908927
Quality control of through-hole technology (THT) and surface mount technology (SMT) solder joints suffers from a high number of false calls, even when automated optical inspection (AOI) is applied. These false calls lead to extensive manual re-inspection require optimization of the test routine. Besides the reduction of false calls, minimizing error slip is essential. In both, research and the industrial contexts, artificial intelligence (AI) is investigated in order to address this problem. While AI-models, especially deep learning-based models such as convolutional neural networks (CNN) can already perform solder joint classification adequately, performance strongly depends on the model architecture and parameters and especially the dataset for training the model itself. To address the latter, this works investigates the utilization of two different approaches to enhance dataset quality, especially if defect data is scarce. First, the systematic inclusion of synthetic data is investigated. Although sufficient synthetic data can be generated by deep generative models such as generative adversarial networks (GANs), robustly training these models and determining the optimal amount of data remains challenging. Utilization of multiple consecutive trainings of pretrained GAN increases robustness a, with a 100% success rate of convergence in the final GAN training, even if there is only a small amount of defective images. Second, methods of active learning (AL) are applied to perform a systematic data augmentation on the most valuable images. Thus, conventional data augmentation methods can be used without complex preparation of synthetic data. Furthermore, well established model architectures can be trained without small modifications in the training process itself. Directed augmentation of images improves recall by 16.5%, although precision slightly decreases.
APA:
Thielen, N., Braun, J., Rachinger, B., Franke, J., Risch, F., & Reinhardt, A. (2025). Improving Data Augmentation in Deep Learning-Based THT Solder Joint Classification with Synthetic Data and Active Learning. In 2025 Pan Pacific Strategic Electronics Symposium (Pan Pacific). Maui, HI, US: Institute of Electrical and Electronics Engineers Inc..
MLA:
Thielen, Nils, et al. "Improving Data Augmentation in Deep Learning-Based THT Solder Joint Classification with Synthetic Data and Active Learning." Proceedings of the 2025 Pan Pacific Strategic Electronics Symposium, Pan Pacific 2025, Maui, HI Institute of Electrical and Electronics Engineers Inc., 2025.
BibTeX: Download