Multi-scale feature pyramid approach for melt track classification in laser powder bed fusion via coaxial high-speed imaging

Abstract

The randomness and low frequency of laser powder bed fusion defects are two important characteristics that can impact the quality and reliability of parts. Therefore, effectively detecting the forming quality of parts during the manufacturing process has become an important research problem in the field of intelligent additive manufacturing technology. In this study, the use of multi-scale and multi-feature manifold learning methods first demonstrated that the global optimal solution for predicting the forming morphology of the melt track cannot be obtained when the number of process phenomenon features in the laser powder bed fusion process is unknown. As an alternative, a multi-scale feature pyramid network is used for processing long sequence high-speed videos and predicting the forming morphology. Specifically, to address the randomness issue, this study used a coaxial high-speed imaging system to monitor the entire forming process and designed a 2D Transformer-based video understanding model to process high-speed video data and recognize key process phenomena. To solve the low frequency issue, physics-based simulation can quickly understand how process parameters affect the forming quality of parts to provide guidance for constructing multi-mode category datasets. The experimental results indicate that the model can accurately predict the forming morphology of the melt track, better control the entire forming process, and thus improve manufacturing quality and efficiency.