Soldering Quality Control With Semi-Supervised Deep Learning Model

Quality control through automated anomaly detection is crucial in manufacturing processes, particularly in the electronics industry where soldering defects on Printed Circuit Boards (PCBs) can significantly impact product reliability. In this study, we propose a semi-supervised deep learning approach to effectively detect anomalies in PCB solder joints. Leveraging the open-source Anomalib library, we identified EfficientAD as the optimal model for our application based on comprehensive evaluations. Our methodology evolved through several improvements: initial full-image training revealed limitations as the model focused on irrelevant PCB features rather than subtle soldering defects; we then developed a Region-of-Interest (ROI) approach that significantly improved detection accuracy by isolating specific soldering areas, though this introduced challenges with training time and aspect ratio distortion. Our final innovation—concatenating multiple ROIs into a single composite image—preserved the benefits of targeted analysis while addressing these limitations. Our trained small-sized EfficientAD model achieved a Pixel AUROC Score of 0.912 with practical inference times of 3.86 seconds, 66.62% CPU usage, and 833MB maximum memory consumption on a Raspberry Pi 4B with 2GB RAM. We successfully integrated our optimized model into a Raspberry Pi platform with a custom interface, demonstrating its viability for real-time quality control in resource-constrained industrial environments.