Multi-scale target detection of metal surface defects in additive manufacturing based on reinforcement learning

Abstract

Additive manufacturing technology has found wide application in aerospace, automotive manufacturing, and other industries. This technology builds complex geometric structures and high-precision parts by adding materials layer by layer, significantly reducing material waste and shortening the production cycle. However, the detection and repair of surface defects remain key challenges impacting product quality in the additive manufacturing process. The ambiguity of surface defects, constraints of fixed-scale features, and challenges in feature extraction accuracy are key obstacles to effective defect detection in machine vision target detection systems. To address these challenges, this study enhances the conventional feature extraction network by incorporating a reinforcement learning-based optimization strategy. A machine vision framework named Reinforcement Learning Multi-Scale You Only Look Once (RLMS-YOLO) is proposed for the automated detection and classification of surface defects. A high-resolution camera captures the surface image of the workpiece and in combination with a reinforcement learning algorithm, extracts the most distinctive multi-scale feature maps to analyze surface texture and morphology, accurately classifying and locating surface defects. Experimental results demonstrate that this method effectively detects tiny defects, significantly improves the quality and production efficiency of additive manufacturing products, and provides crucial data support for subsequent process optimization. In the task of additive manufacturing defect detection, the model improves the mean average precision at an intersection over union threshold of 0.5 for detecting surface hole defects by 12.7 %. This result demonstrates the significant effectiveness of reinforcement learning in enhancing multi-scale feature extraction and improving defect detection accuracy in complex industrial environments.