基于5G子基站的室内定位卷积神经网络模型  

Indoor positioning CNN model based on 5G sub-base station

作　　者：谢海情[1] 汪章紫璇 陆俊霖 宜新博 曾梦琳 文勇军[1] XIE Haiqing;WANG Zhangzixuan;LU Junlin;YI Xinbo;ZENG Menglin;WEN Yongjun(School of Physics and Electronics Science,Changsha University of Science and Technology,Changsha Hunan 410114)

机构地区：[1]长沙理工大学物理与电子科学学院,湖南长沙410114

出　　处：《首都师范大学学报（自然科学版）》2022年第1期34-40,共7页Journal of Capital Normal University：Natural Science Edition

基　　金：湖南省教育厅科学研究项目(20K007);长沙市科技计划重点项目(kq1901102);湖南省柔性电子材料基因组工程重点实验室开放研究基金项目(191906)。

摘　　要：针对目前GPS技术在室内定位中精度差的问题,本文将5G技术与卷积神经网络算法相结合,提出基于5G全新无线空口(NR)参数的室内定位方案.通过采集5G NR数据,与参考点编号形成指纹数据存入指纹库,以精确率、召回率和微平均等值作为评价指标,采用卷积神经网络算法对指纹库进行训练以获得定位模型,并使用Adam方法进行模型优化.该方案使用的总数据集为2 400个,其中训练集大小为2 160个,测试集大小为240个;采用定位模型进行了1 000次训练,每批次训练数据量为512个,最终实现平均误差为1.33 m的室内定位效果.To solve the problem of poor accuracy of the current GPS technology of indoor positioning,5 G technology is combined with convolutional neural network algorithm to proposes an indoor positioning scheme based on 5 G new radio(NR)parameters. By collecting 5 G NR data,the fingerprint data is formed with the reference point number and stored in the fingerprint database. Taking the precision,recall and micro-score value as the evaluation index,the convolutional neural network algorithm was used to train the fingerprint database to obtain the location model,and the Adam method was used to optimize the model. The total data set used in this scheme is 2 400,of which the training set size is 2 160 and the test set size is 240. The positioning model was used for 1 000 training sessions,with 512 training data for each batch,and the indoor positioning effect with an average error of 1.33 m was finally achieved.

关 键 词：室内定位 神经网络 5G NR 位置指纹算法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]