Deep learning-based solder joint defect detector

Abstract In the recent years, due to the technological advancements brought by Industry 4.0, flexible in-line quality control processes that analyze 100% of produced parts targeting Zero Defect Manufacturing have become a reality. The electronics industry in general and the manufacturing of PCBAs in particular have also adopted this paradigm, where quality inspection applies especially to the most critical and error-prone feature: the solder joints between PCBs and components. When the literature related to AI-based solder joint quality inspection is analyzed, we find a lack of applications in real production lines, since most of them are limited to lab scenarios. Therefore, we present a deep learning-based method for solder joint defect detection, which is deployed in a real manufacturing line. The mentioned system is trained using the database we captured using an ad hoc robotic system with onboard front and lateral cameras designed also in this work, which we also describe in detail. The deployment of the defect detector in the real production line is done using that same robotic system too, ensuring the coherence and compatibility between the training data and the real inference scenario.