FAULT DETECTION IN METALLIC PRODUCT USING MACHINE LEARNING TECHNIQUES

Abstract

In the realm of manufacturing, ensuring product quality is critical for maintaining operational efficiency and meeting customer satisfaction standards. Defect detection and anomaly identification are key elements of quality control processes. This thesis proposes a machine learning-based approach to the detection and segmentation of faults in metallic products, utilizing three advanced techniques: Mini Batch Dictionary Learning with Sparse Coder, a custom U-Net model, and DeepLabV3+ algorithms. The research focuses on six distinct metallic objects—cable, grid, metal nut, screw, transistor, and zipper—using the MVTec AD anomaly detection dataset, which includes both defective and defect-free images. For unsupervised anomaly detection, the Mini Batch Dictionary Learning method is employed, demonstrating high precision with an Average Precision (AP) score of 0.976 across the selected metallic objects. Additionally, a custom U-Net model is developed and trained for fault segmentation, providing detailed pixel-level detection of defects on metallic surfaces. The U-Net model achieved high accuracy levels, ranging from 87.92% to 99.59% for different object types, indicating its strong applicability in industrial environments. Finally, DeepLabV3+ model is incorporated to improve segmentation accuracy through the enhancement of defect detection and classification capabilities. The results of this study validate the effectiveness of machine learning algorithms in automating Defect detection in industrial products, thereby reducing human error, minimizing waste, and improving overall production quality. A comparative analysis demonstrates the competitiveness of the proposed approach against alternative algorithms. Future research should focus on refining the segmentation of faulty regions and exploring additional performance metrics. This study contributes significantly to the field of industrial anomaly detection, offering valuable insights into enhancing quality control procedures within industrial settings.