Publication: Detection of solder joint defects on printed circuit board using pre trained convolutional neural networks
Loading...
Date
2023-07
Authors
Tan, Perng Yuan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Solder joint defects are common in electronics manufacturing and can cause serious problems, such as electrical shorts or open circuits. Traditional methods of detecting these defects can be time-consuming and unreliable, making automated defect detection system in high demand. Automated defect detection using CNN is a popular and effective approach for identifying and classifying defects. In this research, the pre-trained CNN model, namely Faster R-CNN, is proposed to classify images of solder joints into four different classes, namely ‘Normal’, ‘Over solder’, ‘Dirty solder’, and ‘Skipped solder’. The model is trained on the Google Colab platform with images that are pre-processed and augmented on the Roboflow website. The results of this research demonstrate the feasibility of using pre-trained CNN for the automated detection of solder joint defects on PCB. Faster-RCNN ResNet50 achieved a great performance in solder joint defect
detection with a 90.9% mAP, 93.8% mAR, and 92.33% F1-score. The optimized hyperparameters of training a Faster R-CNN model are a batch size of 4 and 10000 iteration steps. Faster R-CNN ResNet50 performance is being compared with other SSD pre-trained models. SSD MobileNet V1 FPN, SSD MobileNet V1 FPNLite, SSD ResNet50 V1 FPN, SSD ResNet101 V1 FPN, SSD ResNet152 V1 FPN achieved an F1-score of 90.60%, 83.45%, 87.09%, 80.68% and 76.01%
respectively. Solder joint defect detection systems can improve the efficiency and reliability of electronics manufacturing processes and reduce the risk of defects in the final products