This study addresses the challenge of developing an automated defect detection system for laser-brazed welds on car body surfaces, with a focus on overcoming the limitations of traditional manual inspection methods in automotive manufacturing. By integrating deep-learning techniques, including convolutional neural networks (CNNs), transfer learning, and image denoising, the research aimed to improve defect detection accuracy in noisy production environments. Three models were evaluated—VGG16, ResNet50, and a custom 2-layer CNN—based on their training progress, performance on unseen data, and statistical significance of their results. VGG16 consistently outperformed the other models, achieving the highest accuracy (0.87), F1 score (0.85) and AUC (0.92) demonstrating its suitability for defect detection, particularly in data-limited scenarios where transfer learning is advantageous. The custom CNN, while less complex and computationally efficient, performed competitively but requires more data for optimal tuning. ResNet50, however, struggled with convergence due to the noisy data despite denoising efforts.
From Volvo's perspective, the study’s findings provide a promising proof of concept for automating defect detection, offering potential long-term cost savings by reducing labor and avoiding production inefficiencies. While the hardware setup is minimal, the models showed potential for scalability and implementation in a real-world manufacturing environment. The research contributes to the field by validating the use of transfer learning in defect detection and deep learning denoising but also highlights the need for further optimization and dataset expansion to ensure broader applicability. Future work will focus on refining the models, increasing dataset diversity and automating the labelling process.