Real-Time Quality Inspection Based on Transfer Learning and Feature Clustering for Wave Soldering
Daoyuan Liu,Yu Guo,Jian Xie,Hanpeng Gao,Shaohua Huang
DOI: https://doi.org/10.1109/tim.2024.3365159
IF: 5.6
2024-03-02
IEEE Transactions on Instrumentation and Measurement
Abstract:Accurate quality inspection for wave soldering is crucial to ensure that the products moving into the next processing step are qualified, which plays a nonnegligible role in reducing production costs. When addressing the quality inspection problem of wave soldering, there is a need to enhance the accuracy of vision-based nondestructive examination methods. Hence, a novel inspection model based on EfficientNet is designed to solve the problem of low accuracy. First, model-based transfer learning is adopted to initialize the parameters of the feature extractor, improving the feature extraction ability. Second, a new loss function for high-order feature clustering is innovatively constructed and combined with the classification loss function that carries category weights, reducing the training difficulty and enhancing the performance of the inspection model. Finally, 12 404 images of wave soldering collected from an electronic factory are taken as the validation object. The experiments have shown that the proposed model is effective and superior to existing models, achieving an accuracy of 99.35%, a precision of 100%, and a recall of 98.03%. Meanwhile, the inspection time of the proposed model for assessing the quality of 30 solder joints is no more than 0.54 s, meeting the real-time quality inspection requirements for wave soldering.
engineering, electrical & electronic,instruments & instrumentation