Defect Detection Using Deep Learning-Based YOLOv3 in Cross-Sectional Image of Additive Manufacturing

Deposition defects like porosity, crack and lack of fusion in additive manufacturing process is a major obstacle to commercialization of the process. Thus, metallurgical microscopy analysis has been mainly conducted to optimize process conditions by detecting and investigating the defects. However,...

Full description

Saved in:
Bibliographic Details
Published inArchives of metallurgy and materials Vol. 66; no. 4; pp. 1037 - 1041
Main Authors Choi, Byungjoo, Choi, Yongjun, Lee, Moon Gu, Kim, Jung Sub, Lee, Sang Won, Jeon, Yongho
Format Journal Article
LanguageEnglish
Published Warsaw Polish Academy of Sciences 01.01.2021
Subjects
Additive manufacturing
Algorithms
Artificial intelligence
Commercialization
data augmentation
Defects
deposition defect
Human error
Machine learning
Manufacturing
Metallurgical analysis
object detection
Object recognition
Performance evaluation
Training
Verification
yolov3
Online AccessGet full text

Cover

More Information
Summary:Deposition defects like porosity, crack and lack of fusion in additive manufacturing process is a major obstacle to commercialization of the process. Thus, metallurgical microscopy analysis has been mainly conducted to optimize process conditions by detecting and investigating the defects. However, these defect detection methods indicate a deviation from the operator’s experience. In this study, artificial intelligence based YOLOv3 of object detection algorithm was applied to avoid the human dependency. The algorithm aims to automatically find and label the defects. To enable the aim, 80 training images and 20 verification images were prepared, and they were amplified into 640 training images and 160 verification images using augmentation algorithm of rotation, movement and scale down, randomly. To evaluate the performance of the algorithm, total loss was derived as the sum of localization loss, confidence loss, and classification loss. In the training process, the total loss was 8.672 for the initial 100 sample images. However, the total loss was reduced to 5.841 after training with additional 800 images. For the verification of the proposed method, new defect images were input and then the mean Average Precision (mAP) in terms of precision and recall was 0.3795. Therefore, the detection performance with high accuracy can be applied to industry for avoiding human errors.
ISSN:1733-3490
2300-1909
DOI:10.24425/amm.2021.136421