Automated Optical Inspection of Soldering Connections in Power Electronics Production Using Convolutional Neural Networks

Abstract

Automatic optical inspection (AOI) of solder joints is a common testing process in electronics production. Especially in power electronics production for electric drive systems, such inspection systems are employed for quality control of selective soldering processes for through-hole devices. Up to now, commercial systems rely on rule-based programming for the determination of soldering quality. However, this approach demands expert knowledge for setup and is very susceptible to changes in input data. To avoid error slip, thresholds are often defined very strictly, resulting in a high pseudo-error rate. Improvement is only possible through extensive expert input. As power electronics production is often characterized by a high variant and only medium quantities, this manual effort is critical. In this contribution, we benchmark a commercial AOI system with an adaptive approach utilizing convolutional neural networks based on a pre-trained VGG-16 algorithm with custom fully connected layers. Supervised learning is employed for each static region of interest with refined labeled data from the existing AOI system. To overcome the extremely unbalanced dataset, we employ data augmentation and data filtering. Our results show significant improvement in precision over the commercial system regarding the total recall. In addition, the adaptive system is also able to learn from pseudo-error classifications. We also show that our approach can not only output a binary classification but also identify process deviations that may still yield acceptable quality. Hence, this output might be used for an online control of process parameters in further research.