Authors:
Conceição N. Silva
1
;
Neandra P. Ferreira
2
;
Sharlene S. Meireles
2
;
Mario Otani
2
;
Vandermi J. da Silva
1
;
Carlos A. O. de Freitas
1
and
Felipe G. Oliveira
1
Affiliations:
1
Institute of Exact Sciences and Technology (ICET), Federal University of Amazonas (UFAM), Itacoatiara, Amazonas, Brazil
;
2
Cal-Comp, Institute of Research and Technological Innovation (ICCT), Manaus, Amazonas, Brazil
Keyword(s):
Ball Bond Inspection, Automatic Visual Inspection, Deep Learning.
Abstract:
The growing demand for increasing memory storage capacity has required a high density of integration within the semiconductor encapsulation and, consequently, has made this process more complex and susceptible to failures during the production stage. In the semiconductor encapsulation area, the costs of materials and equipment are high and the profit margin is narrow, making it necessary to rigorously inspect the process steps to keep the productive activity viable. This work addresses the problem of quality control in silicon wafers soldering procedure, allowing error detection before the epoxy resin molding process, generating useful infor-mation for correcting equipment configurations and predicting failures from the raw materials and inputs used in the process. We propose an approach to classify solder balls, in the soldering process of silicon wafers on Ball Grid Array (BGA), contained in the Printed Circuit Board (PCB) substrates. The proposed methodology is composed of two mai
n steps: i) Solder ball segmentation; and ii) Solder ball classification through deep learning. The proposed predictive model learns the relation between visual features and the different soldering conditions. Real and simulated experiments were carried out to validate the proposed approach. Results show the obtained accuracy of 99.4%, using Convolutional Neural Network (CNN) classification model. Furthermore, the proposed approach presents high accuracy even regarding noisy images, resulting in accuracy of 92.8% and 75.7% for a Salt and Pepper and Gaussian noise, respectively, in the worst scenario. Experiments demonstrate reliability and robustness, optimizing the manufacturing.
(More)