Deep Learning-Powered Computer Vision System for Selective Disassembly of Waste Printed Circuit Boards

The rapid growth of WPCBs poses significant environmental and resource challenges for their treatment due to their complex composition, which includes approximately 50 CRMs such as copper, beryllium and gallium. While WPCBs represent an environmental problem for their disposal, they also represent an important opportunity because their CRMs content is very high, even compared to that found in naturally occurring mines. Today, however, the recovery of high density CRMs from a circular economy perspective, through automated disassembly WPCBs have the potential because crushing and chemical treatment, is limited, for economic reasons, to a few CRMs (usually gold, copper, palladium and silver), resulting in the loss of the other CRMs. In order to make the extraction of CRMs from WPCBs cost-effective, this research proposes a new circular economy framework, exploiting the latest machine learning and computer vision techniques, aimed at selectively disassembling (or selecting on a conveyor belt) different types of electronic components to obtain different material, each with a high concentration of specific CRMs, so as to make chemical treatment aimed at their extraction efficient. The developed AI system combines object recognition models, including the YOLO and Transformer architectures, with a recyclability evaluation framework. The system is designed to detect and classify electronic components, quantify their recyclability based on material composition, and optimize CRMs recovery. The V-PCB dataset, curated and annotated as part of this work, serves as a benchmark for training and evaluating the proposed models. Data collection involved high-resolution imaging of V-PCBs using camera modules integrated with NVIDIA Jetson Nano, ensuring scalability for real-world applications. The methodology employs iterative training and domain adaptation techniques to improve model performance. Multi-stage transfer learning strategies were used to adapt the model to varying real-world conditions, significantly improving component recognition accuracy. In addition, the recyclability assessment integrates material analysis techniques such as XRD, providing a comprehensive assessment of the CRMs recovery potential. Experimental results demonstrate the superiority of the proposed system over traditional methods, achieving high mAP for component detection and ultimately increased high density CRMs recovery rates. The results highlight the feasibility of an automated, sustainable approach to WPCBs recycling that addresses key gaps in existing methods by focusing on component-level disassembly and reuse. This research makes a significant contribution to the field of e-waste management by providing a scalable and efficient solution for the recovery of CRMs. It aligns with circular economy principles by reducing environmental impact, minimizing waste, and promoting the reuse of functional components. The results have been disseminated at leading conferences and the proposed system is ready to transform industrial recycling workflows, making it a critical step towards sustainable electronics manufacturing.