基于边缘加速节点的直线加速器束流轨道参数预测技术研究  

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

作　　者：侯明阳 郭玉辉[2] 杨旭辉 杨贵进 HOU Mingyang;GUO Yuhui;YANG Xuhui;YANG Guijin(College of Physics and Electronic Engineering,Northwest Normal University,Lanzhou 730070,China;Institute of Modern Physics,Chinese Academy of Sciences,Lanzhou 730000,China)

机构地区：[1]西北师范大学物理与电子工程学院,兰州730070 [2]中国科学院近代物理研究所,兰州730000

出　　处：《原子核物理评论》2024年第4期975-983,共9页Nuclear Physics Review

基　　金：国家重点研发计划项目(2022YFF0704900,2019YFA0405400);中国科学院先导专项B类(XDB25030103)。

摘　　要：在目前能源紧缺的国际背景下,核能发展为重要清洁能源,质子加速器成为处理核废料的重要设备。在加速器束流轨道校正过程中,需要束流轨道参数辅助计算。神经网络目前已被广泛应用在工业界各个领域,可以实现准确率高的数据拟合预测。因此,本工作提出了一种基于边缘智能计算节点的直线加速器束流轨道参数预测技术。该方法通过反向传播神经网络(Back Propagation Neural Network,BPNN)对过往数据学习,生成模型并部署到边缘加速节点以预测束流位置监测器BPM(Beam Position Monitor,BPM)参数,从而使用边缘计算节点加速校正束流位置。实验结果表明,现场可编程门阵列(Field Programmable Gate Array,FPGA)硬件加速器作为边缘加速节点仿真推理速度达到55μs,其能效比分别是GPU和CPU的23.13倍和553.15倍左右,预测结果误差平均为0.5%,时延和精度达到预期目标。s:Under the international background of energy scarcity,nuclear energy has emerged as a crucial clean energy source,with proton accelerators serving as vital equipment for nuclear waste treatment.In accelerator beam orbit correction processes,accurate beam trajectory parameter prediction becomes essential for computational assistance.Given the widespread application of neural networks in industrial sectors for high-accuracy data fitting and prediction,this study proposes a beam orbit parameter prediction technique for linear accelerators based on edge intelligent computing nodes.The methodology employs a BPNN to learn historical operational data,subsequently deploying the trained model to edge acceleration nodes for predicting BPM parameters.This approach enables accelerated beam position correction through edge computing resources.Experimental results demonstrate that the FPGA hardware accelerator,functioning as an edge acceleration node,achieves a simulation inference speed of 55μs.The energy efficiency ratio outperforms GPU and CPU platforms by factors of 23.13 and 553.15 respectively,with an average prediction error of 0.5%.Both temporal performance and prediction accuracy meet predetermined requirements,indicating promising potential for real-time beam control applications in nuclear energy facilities.

关 键 词：FPGA 轨道参数预测 神经网络加速器 BPM 

分 类 号：TL53[核科学技术—核技术及应用]