International Journal of Electronics and Computer Applications

Abstract

Automated quality inspection of Printed Circuit Boards (PCBs) is a critical requirement in modern electronics manufacturing. Traditional manual inspection methods are slow, inconsistent, and error-prone, while industrial Automated Optical Inspection (AOI) machines are prohibitively expensive for small and medium enterprises. This paper presents an AI-based PCB fault detection system using a custom Convolutional Neural Network (CNN) trained on the DeepPCB benchmark dataset from Peking University. The proposed approach introduces XML annotation-guided patch extraction that crops 128×128 pixel regions centered on annotated defect locations, enabling the classifier to focus exclusively on defect morphology rather than irrelevant background regions. The system classifies seven categories: missing hole, mouse bite, open circuit, short circuit, spur, spurious copper, and no-defect. Data augmentation expanded the training set from 2,953 to 25,472 samples. The model achieves 9697% validation accuracy on 680 unseen samples, outperforming comparable CNN-based methods in the literature. Grad-CAM (Gradient-weighted Class Activation Mapping) provides visual explainability by highlighting defect locations with heatmap overlays. The complete system is deployed as a Flask web application with a Neon PostgreSQL cloud database for persistent inspection logging. Experimental results demonstrate the feasibility of replacing expensive hardware AOI systems with an affordable, software-based AI inspection solution.