Research on Prediction Technology for Beamline Parameters of Linear Accelerator Based on Edge Computing Nodes

Subjects: Nuclear Science and Technology >> Particle Accelerator Subjects: Electronics and Communication Technology >> Electron Technology submitted time 2023-11-29

Hou Mingyang Guo Yuhui Yang Xuhui Yang guijin

Abstract：  In light of the current international energy scarcity, nuclear power has emerged as a crucial source of clean energy. Proton accelerators have therefore become a pivotal technology in nuclear waste management. During beamline orbit correction processes, precise calculations of beamline orbit parameters are required. Given the demonstrated effectiveness of neural networks in a wide variety of industry domains, they offer promising potential for high-accuracy data fitting and prediction. Hence, this study proposes a novel direct linear accelerator beamline orbit parameter prediction technique based on edge intelligence computing nodes. This technique leverages BPNN to learn from historical data and generate a powerful model that can be seamlessly deployed to edge computing nodes, thereby accelerating the prediction of BPM location parameters. Furthermore, the proposed approach may be complemented by an adaptive compensation system in the future, which, in combination with edge computing nodes, could enable automatic beamline position correction, thereby achieving beamline orbit correction. Our experimental results demonstrate that FPGA, as an edge acceleration node, can achieve an inference speed of 2.5us, which represents a remarkable performance enhancement of approximately 165.6 times compared to CPU and approximately 7.9 times compared to GPU. The predicted results exhibit an average error of only 0.5%, and they exhibit the desired latency and accuracy characteristics.

YES

 Hits 865  Downloads 192  Comment 0