Universal light SAM algorithm for in situ melting pool monitoring of L-DED/L-PBF/PAM processes

Abstract

In additive manufacturing (AM), in situ monitoring of the melting pool and plasma arc status is crucial for ensuring workpiece quality. However, existing monitoring algorithms based on image segmentation typically require substantial amounts of high-quality training data. Furthermore, these algorithms lack the versatility necessary for deployment across various types of AM equipment. The recently proposed segment anything model (SAM) demonstrates strong generalization capabilities through its novel interactive image segmentation framework. However, its extensive network parameters hinder direct application in the field of AM in situ monitoring. To solve these problems, this paper presents a lightweight, general-purpose image segmentation algorithm called Light SAM, which is capable of performing cross-process feature extraction. The characteristics of the melt pool in the L-DED (laser directed energy deposition) and L-PBF (laser powder bed fusion) processes, as well as the plasma arc in the PAM process, can be extracted using only 190 training images. Mean intersection over union (MIoU) was chosen as the primary evaluation metric. Under the same training strategy, the four most advanced image segmentation algorithms demonstrated poor generalization compared with alternative methods, whereas Light SAM exhibited strong generalization, achieving an MIoU of 93.02%. Additionally, while its performance is comparable to that of the original SAM, its model size is only 23 MB. Moreover, the Light SAM algorithm can be integrated into the AI edge computing board, achieving an inference time of just 153 ms under Neural-network Processing Unit (NPU) acceleration following model pruning optimization, which meets the requirements of AM real-time monitoring. In addition to meeting the requirements of inference speed and accuracy, the algorithm can simultaneously extract multiple target features by manually inputting several bounding boxes to guide its processing. Furthermore, the algorithm significantly reduces training costs and time, and can be deployed in cross-process AM monitoring systems, thereby facilitating the widespread adoption of in situ monitoring technologies in AM.