Anomaly Detection for Dispensing of Solder Paste on 3D Circuit Carriers Using Machine Learning

This paper addresses the challenges in manufacturing of 3D mechatronic integrated devices, particularly focusing on the anomalies that occur during the transfer of solder paste as an interconnection material. Utilizing machine learning techniques, the study extends prior research done on 2D circuit carriers to identify these anomalies in 3D substrates and predicts solder paste depot quality in general. The research employs a design of experiments to develop a comprehensive dataset, taking into account the unique properties of 3D substrates. The paper also presents a low-cost camera and segmentation algorithm setup to measure quality parameters like the area and offset of solder paste, as conventional systems fall short in this context. The system has a very good measurement capability for determining the area of the solder paste. Based on the data obtained, transfer learning techniques are utilized to leverage existing 2D data for model training, showing a significant improvement in model performance and especially reducing training times. Specifically, an R
²
score of 0.88 is achieved using transfer learning, compared to 0.83 with only 3D data, while an AutoML-model even reaches an R2 score of 0.94. The study suggests that AI-based methods offer cost-effective alternatives to established systems, enabling flexible manufacturing with fewer iterations and lower setup costs.