Advancing computer vision-based defect detection and traceability in iron foundries : Using synthetic data and durable, vision-readable marking methods

Sammanfattning

Ensuring high surface quality in cast metal components is crucial for performance and reliability in industrial manufacturing. Traditional visual inspection methods have been used for a long time, but they can become inconsistent when production speeds are high. Recent progress in artificial intelligence (AI) and computer vision offers promising options for automating quality checks and improving defect detection accuracy. This research looks into using deep learning-based computer vision methods to detect surface defects in cast iron bearing housings. The study follows a multi-phase approach. It begins with collecting real production images, manually marking defect areas, and creating synthetic data through data augmentation, 3D modeling, defect patching, and FastGAN. These methods aim to tackle issues related to limited defect data and changing imaging conditions often found in manufacturing environments. Both supervised and unsupervised models were developed and assessed using industrial-quality metrics like precision, recall, and F1-score. The effect of synthetic data was tested in situations like low lighting and infrequent defect appearances, showing improvements in detection reliability and model adaptability. Additionally, a new traceability system, Arrow Clock Marking, was created and tested to offer durable, vision-readable component identification during post-processing steps such as shot blasting. The results show that using synthetic data improves the effectiveness of supervised models in identifying both major and minor surface anomalies. The traceability approach also proved reliable under different conditions, allowing for consistent identification. Overall, this research connects methods in AI with practical use in manufacturing. It contributes to more scalable and resilient quality assurance systems.