Comparative Analysis of Deep Learning-Based Defect Monitoring in Metal Additive Manufacturing
Md Hasib Zubayer,Chaoqun Zhang,Wang Yafei
DOI: https://doi.org/10.1109/icccr61138.2024.10585345
2024-01-01
Abstract:In metal additive manufacturing (AM), precise realtime defect detection is critical and significantly contributes to the effectiveness of this technology. The recent progress in additive manufacturing industry is increasingly reliant on AI-driven deep learning techniques, which demand substantial computational resources. Therefore, the optimization of software components and their alignment with appropriate hardware configurations are essential for the development of a system capable of realtime inference. This study conducts a comparative analysis of five advanced realtime image processing algorithms: Convolutional Neural Networks (CNN), Single Shot Detector (SSD), Faster Region-based Convolutional Neural Networks (Faster R-CNN), You Only Look Once (YOLO) version 5, and the latest YOLO version 8. The aim is to determine the most rapid and efficient algorithm for defect detection in metal AM. The results reveal that YOLOv8 outperforms CNN, SSD, Faster R-CNN, and other YOLO models in effectiveness, demonstrating a remarkable defect detection accuracy of up to 96% within approximately 288 seconds, a significant improvement over other automated recognition methods. While SSD boasts the fastest detection speeds, YOLOv8 leads in accuracy, highlighting its pivotal role in metal AM defect identification. The evaluation, based on metrics such as accuracy, precision, mean average precision (mAP), recall, and F1 score, assigns mAP values as follows: CNN (72%), Faster R-CNN (77 % ), SSD (73%), and YOLOv5 (89 % ). This comprehensive assessment highlights YOLOv8's outstanding capabilities, establishing it as a frontrunner, albeit slightly slower than SSD in speed, yet surpassing other deep learning models in accuracy and precision.