Online Cladding Quality Assessment Based on Yolov8-Rf Model Using Molten Pool Images

Abstract

Laser directed energy deposition (LDED) holds great promise for high-end manufacturing. However, its unstable forming quality limits wider industrial adoption. The morphology of the molten pool is a key indicator of process conditions and forming quality, and its dynamic evolution offers critical cues for fault detection and real-time control. Yet, complex interactions among laser, powder, and substrate introduce strong background interference, impairing the accuracy of molten pool extraction and compromising prediction stability. To address these challenges, this study proposes a novel YOLOv8-RF based method for online assessment of cladding quality. Firstly, the model’s segmentation layer enables real-time tracking of molten pool morphology. By optimizing the loss function, the influence of poor annotations and geometric penalty artifacts is reduced, improving generalization. In addition, a global attention module is added to the segmentation head to enhance boundary recognition. Secondly, a cladding quality metric is defined based on dilution rate, aspect ratio, and wetting angle, with morphological evolution analyzed under varying process parameters. Finally, the quality assessment layer of the model establishes a nonlinear mapping between molten pool features and quality. Based on feature correlation analysis and optimization. five representative features are selected as input. Experimental results show the model achieves 98.87% accuracy with these five optimized features. Moreover, Feature visualization confirms the method’s effectiveness in reducing redundancy and computational cost. This work offers a novel and efficient approach for real-time tracking and quality assessment in LDED, providing critical technical support for quality control and process stability enhancement in additive manufacturing.

Keywords: LDED, Forming Quality, Attention Module, Feature Optimization