Publication: Machine learning study of void detection using machine learning
Loading...
Date
2024-07
Authors
Chew, Cheng Kai
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
As BGA chips continue to shrink in size, voiding has become a significant challenge in soldering, particularly during underfilling. Manual inspection for voids is not only labour-intensive but also prone to human error, making it difficult to accurately evaluate the void percentage for acceptance. This research introduces a deep learning CNN network image processing model, Mask R-CNN, developed using Detectron2, to detect voids in TSAM images of chips with epoxy underfill. Various datasets and ratio were compared using evaluation metrics to determine the best performance. Additionally, comparison between Roboflow training and Mask R-CNN model is conducted as well as evaluations between different Mask R-CNN. A regression study of the correlation between underfill encapsulation parameters and void percentage was conducted using Orange Data Mining and Minitab Statistical Software. The result is the highest mAP value of 0,537 achieved by 70-image dataset with ratio of 8:1:1. The Roboflow training shows overall higher mAP value compared to Mask R-CNN training with the highest mAP achieved at 70-image dataset, 0.584. The highest mAP value of 0.566 achieved by X101_FPN model shows the best performance while lowest time achieved by R101_FPN model shows the fastest training with 25.49 minutes. The regression study result shows valve pressure is the most significant parameters affect the void percentage of the chip.