Improved U-net network with temporal adaptive Module for precise melt pool detection in LDED additive manufacturing

Melt powder laser directed energy deposition (LDED) is a process that involves heating and melting metal powders with a laser, which then causes the powders to condense into molten pools on the substrate. The features of these pools have a direct impact on the components’ material and physical qualities. As a result, precise and timely detection of molten pool shape is critical for efficient quality monitoring and feedback management. In order to partition the metal molten pool region at the pixel level in LDED additive manufacturing process, an improved U-Net network based on the temporal adaptive module (TAM) is presented in this research. First, a database of melt pool morphology is created, which includes a variety of processing factors and metal powder components. Then, an advanced U-Net network is created to semantically partition the melt pool region in order to collect exact melt pool border information. A comparison study of multiple semantic segmentation models is also performed, and the specific contributions of each model component are confirmed via a series of ablation experiments. According to the testing results, the improved U-Net model achieves 95.8% accuracy and 32.5 FPS, which represents a significant improvement in accuracy and performance over other network architectures. Finally, the study investigated the fluctuation of melt pool morphological characteristics in response to various processing factors, as well as the interaction between melt pool morphology and laser power, scanning speed, and powder feeding rate.