A robust LSTM-based defect identification methodology for imbalanced data of solder paste inspection

Nowadays, the demand for high-quality electronic products has increased substantially, compelling printed circuit board (PCB) manufacturers to improve the quality of surface mount technology (SMT) processes. Due to the high sensitivity settings of solder paste inspection (SPI) machines, the number of samples classified as defective has significantly increased. As a result, quality inspectors must perform additional manual re-inspections, which increases labor costs and raises the risk of misclassification. To address this issue, this study integrates principal component analysis (PCA), the borderline synthetic minority over-sampling technique (SMOTE), the edited nearest neighbor (ENN) method, and the long short-term memory (LSTM) deep learning model to develop a robust methodology for identifying defective samples in the solder paste printing (SPP) stage. To facilitate more efficient data analysis, PCA is first employed to analyze the information content of the feature variables for SPI data, thereby extracting the most informative features. Because defective SPP samples are relatively scarce, the borderline SMOTE method is applied to generate additional minority-class observations. During the training process of the deep learning model, mislabeled data may adversely impact the classification accuracy of the proposed model. To mitigate this issue, the ENN method is utilized to remove potentially mislabeled samples. Once high-quality samples are obtained, an LSTM deep learning model is employed to establish a robust defect identification model for SPP samples to boost the PCB manufacturing yield. Finally, this study uses the SPI data to validates the proposed model, achieving an outstanding classification accuracy of 99.9%.