PCBA defect classification in real world with out-of-distribution detection

Inadequate chip solder joints can significantly impact the overall quality of finalised Printed Circuit Board (PCB) products. Detecting all solder joint defects in real time and adaptively during the actual production process poses a formidable challenge due to the diverse nature of these defects and the limited availability of anomaly data. In contrast to the conventional Printed Circuit Board Assembly (PCBA) defect classification task that is limited to pre-labelled defects, we propose adaptive classification algorithms capable of identifying new categories of defects and ensuring accurate classification of labelled categories. Specifically, we developed a multitasking network that utilises Swin-Transformer as the underlying architecture to classify labelled categories and identify novel defective categories. And neuronal activation coverage is designed to detect unseen types of PCBA defects. Furthermore, we propose an end-to-end unsupervised hashing algorithm that incorporates novel category discovery for images classified as previously unseen categories. Finally, we conducted experiments on a variety of different backbone networks in real PCBA defect datasets to demonstrate the effectiveness of our proposed method.