AI-Based Classification Method Identifying Burns on Used PCB Boards for Reuse in a Circular Manufacturing Process

The rapid growth of EOL and used products has intensified the need for circular manufacturing strategies that enable reuse and remanufacturing of products. One of the most critical challenges in this context is the visual inspection of EOL/used products to determine reusability. Traditional approaches rely either on manual inspection, which is often slow, inconsistent, and dependent on individual judgment, or on automated inspection systems based on rule-based image processing, which lack adaptability to new defect types, new products, and changing environmental conditions. This thesis addresses these challenges in the electronics domain by developing an AI-based classification method to classify burnt printed circuit boards (PCBs) from reusable PCBs. A dataset is developed by combining burnt and good PCB images collected from literature, online sources, and manually verified cases with synthetic images generated using generative AI tools such as ChatGPT (DALL·E) and Gemini. Data integrity is ensured through perceptual hashing and deep feature filtering with a pre-trained ResNet50 model to remove duplicate and augmented images. Finally, six targeted augmentation pipelines are also applied to introduce realistic variations in geometry, lighting, occlusion, noise, background, and compression. Several state-of-the-art deep learning architectures are fine-tuned on this dataset using transfer learning with pre-trained ImageNet weights. CNNs demonstrated the strongest performance: YOLOv11 and ResNet50 both achieve 98% accuracy with perfect precision for burnt PCBs, while EfficientNetB3 follows closely with an F1-score of 0.96 and perfect precision on good PCBs. In contrast, larger CNNs such as ResNet152 (F1 = 0.77) and EfficientNetB7 (F1 = 0.90) show weaker generalization despite their higher capacity. DeiT achieves competitive performance with an F1-score of 0.92, whereas the self-supervised DINO variants underperform (F1 = 0.72 for ViT, 0.54 for CaIT backbones). These findings highlight that in data-scarce industrial domains, lightweight CNNs outperform deeper or more complex models. The results demonstrate that AI-based visual inspection can significantly improve PCB reuse decisions in remanufacturing.