A Novel Multi-Scale Convolutional Attention Network Based on Meta-Transfer Strategy for Solder Paste Position Offset Prediction

Solder paste position offset is a critical stencil printing quality indicator, the prediction of which using available data is important for printing quality improvement. However, existing data-driven works on solder paste position offset prediction often suffer from their poor adaptation to changing printing stages and small training samples. To address the above problems, we propose a novel multi-scale convolutional attention network based on meta-transfer strategy for solder paste position offset prediction in surface mount technology assembly lines. First, to better capture the fluctuating trends of printing sequence in different cleaning cycles, we propose a multi-scale convolutional attention network, in which a multi-head attention with position encoding is designed at each level to adaptively capture the changing trend of solder paste position offset at different stages. Then, to improve the precision of prediction model under insufficient samples, we introduce a meta-transfer strategy. Specifically, the parameters of the model are updated in the meta-training process through the bi-level optimization and the parameter fine-tuning method is used in the meta-testing process to improve the prediction performance of proposed model under complex working conditions. The proposed method is verified over practical dataset and compared with other advanced methods. The results show that the proposed method can accurately and robustly achieve solder paste position offset prediction, especially in small sample dataset.