Increasing opportunities for component reuse on printed circuit boards using deep learning

Abstract

With the increasing volume of discarded printed circuit boards, there is an urgent need for efficient classification and reuse of electronic components to mitigate environmental risks and recover valuable materials. Current solutions face challenges due to high computational requirements and inefficiencies in detecting reusable components before they are destroyed. This study introduces PCBNet, a lightweight deep learning model based on a modified YOLOv8-tiny architecture, optimized for electronic component classification. PCBNet incorporates novel knowledge distillation strategies involving three teacher models using a projection head that dynamically updates the teacher model weights to enhance performance without increasing computational complexity. The optimized version, with α = 0.3 and β = 0.7 during the nowledge distillation process, achieves an mAP@50 of 0.467 and an mAP@95 of 0.368 with 0.5 million parameters and 1.7 billion floating-point operations, achieving an optimal balance between performance and computational efficiency. A prototype system using a Raspberry Pi, an automated conveyor, and a monitoring camera has been developed to verify PCBNet's effectiveness in detecting and classifying electronic components in PCBs. The results demonstrate that PCBNet is not only capable of accurate classification of electronic components but is also deployable on low-configuration devices, making it an effective solution for real-time e-waste recycling and component reuse. The results show that PCBNet accurately classifies electronic components and can be deployed on low-configuration devices, providing an effective solution for real-time e-waste recycling and component reuse.

Graphical abstract