Lightweight Machine Learning for Real-Time Defect Detection and Categorization in Industrial Welds

This paper presents a deep learning-based approach for real-time classification of welded and non-welded aluminum samples using a multi-input architecture. The proposed model integrates a CNN branch for extracting visual features (such as cracks and color variations) and a Dense branch for analyzing numerical patterns (such as shear length). A fusion layer combines these features to enhance classification accuracy. Data augmentation techniques were applied to mitigate overfitting, achieving an accuracy of 85.71%, with 100% precision and an F1 score of 85.71%. Comparative analysis with state-of-the-art models highlights the effectiveness of our lightweight architecture for industrial weld inspection. Future work aims to refine classification by identifying weld types and enhancing model performance through advanced data augmentation.