Using neural networks in predicting defects in electronics manufacturing processes

Abstract

This study aims to develop a hybrid neural network model based on image and sensor data for defect prediction in electronics manufacturing. A hybrid model that combines CNN with MLP operates as a defect prediction system for electronics production lines by uniting data from image inputs with sensor outputs to boost predictive accuracy levels. Defect data is obtained from a simulated Surface Mount Technology (SMT) production line through experimental methods, while Adam optimizer and Categorical Cross-Entropy loss function perform the model optimization process. The proposed model demonstrates a 94.7% accuracy rate, which exceeds SVM (85.3%) and Decision Tree (82.1%) results while generating an AUC score of 0.96 to validate its high defect classification performance. The research demonstrates the value of multi-modal defect analysis for automated quality control, but existing issues on real-time deployment and computational efficiency require further improvement.