Deep Learning Model-Driven Pore Detection for Laser Directed Energy Deposition under Varying Brightness and Image Size Conditions

Le résumé fourni par la source

Pore formation is a significant challenges in the advancement of laser additive manufacturing (LAM) technologies. To address this issue, image data-driven pore detection techniques have become a research focus. However, existing methods are constrained by reliance on a single detection environment (e.g., consistent brightness) and fixed input image sizes, limiting their predictive accuracy and application scope. This paper introduces an innovative a pore detection method based on a deep learning model for laser-directed energy deposition (L-DED). The proposed method leverages the deep learning model’s ability to extract feature information from melt pool images captured by a high-speed camera, enabling efficient pore detection under varying brightness conditions and diverse image sizes. The detection results demonstrate that, under varying brightness levels, the proposed model achieves a pore detection accuracy of approximately 93.5 % and a root mean square error (RMSE) of 0.42 for local porosity prediction. Additionally, even with changes in input image size, the model maintains robust performance, achieving a detection accuracy of 96 % for pore status detection and an RMSE value of 0.09 for local porosity prediction. This study not only addresses the limitations of traditional detection techniques but also broadens the scope of online detection technologies. It highlights the potential of deep learning in complex industrial settings and provides valuable insights for advancing defect detection research in related fields.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.