Neurosymbolic Artificial Intelligence for Printed Circuit Board Fault Detection

This paper presents a novel neurosymbolic approach for resource-efficient Printed Circuit Board (PCB) fault detection. The method combines classical image processing with a Convolutional Neural Network (CNN) and a Knowledge Graph (KG) to capture visual and structural relationships among PCB components. First, key PCB features are extracted and structured into a Knowledge Graph; then graph embeddings are fused with CNN-derived features through a hybrid loss function that enforces alignment between visual and symbolic knowledge. Experiments on the DeepPCB dataset show that the proposed framework improves the classification accuracy from 83.86% (baseline CNN) to 87.85%, while reducing false positives and improving the interpretability of the model. These results demonstrate that the integration of structured domain knowledge significantly improves performance and explainability, making the framework promising for deployment in resource-constrained industrial environments.