【作者】 刘琳；

【导师】 陈晓曙；

【作者基本信息】 东南大学 ， 信息与通信工程， 2020， 硕士

【摘要】 近年来,随着智能手机和移动互联网业务的高速发展,LBS在人们的生活、娱乐以及安全等领域的作用越来越明显。然而在5G技术与场景下,复杂多变的城市环境对移动终端定位的实时准确性提出了更多挑战。指纹定位方法能有效利用无线传播的多径特性实现高精度定位,其与诸如5G大规模MIMO技术相结合的应用前景更加宽阔,目前就两者结合的方法和技术研究仍在不断深入开展中。本文就大规模MIMO单站场景下基于机器学习的指纹定位方法展开了具体的研究,主要涵盖了以下内容:(1)针对城市环境的多变性导致定位性能变差的问题,提取三种具有不同鲁棒性的指纹。一是根据ADCPM指纹在角度域呈现的功率分布特点而提取的MACPV指纹,二是基于MACPV指纹进行降噪自编码器处理的DAE指纹,三是基于ADCPM指纹进行降噪卷积自编码器处理的DCAE指纹。对采用不同指纹定位精度的仿真分析研究表明随着散射体数目的减少,基于DCAE指纹的定位具有更优的性能效果。(2)针对5G网络指纹库更新代价过高的问题,提出了CAOA-PFUC更新算法,其利用基于中心AOA聚类的方法并结合基于概率的指纹更新准则来决定指纹库的更新与否。仿真研究的结果表明该方法在具有与暴力更新算法相当的定位精度前提下,能有效降低指纹库更新代价。(3)以实现实时定位和简化在线阶段流程为目标,提出了基于PQ算法的快速指纹定位方法,其通过基于ADCPM指纹的码本建立和码字搜索方式提高定位时效性。针对指纹库存储规模大幅增长时,PQ算法对定位速度性能提升不明显问题,给出了基于MTL的CNN快速定位改进算法,通过不依赖于指纹库规模的前馈神经网络计算可以非常高效的估计终端位置。仿真结果表明,当指纹数据库规模较大时,基于CNN算法优势明显,在保证定位精度接近于基于WKNN算法的条件下,将定位速度提升到毫秒级。更多还原

【Abstract】 In recent years,with the rapid development of smart phones and mobile Internet services,LBS plays increasingly important role in the fields of people’s lives,entertainment and security.However,under 5G technology and scenario,more challenges to the mobile terminal positioning instantaneous and accuracy have been put forward due to the complex and changeable urban environment.Fingerprint localization method is used to achieve high-precision positioning,in which the multi-path characteristics of wireless propagation could be effectively utilized,and it has a broader application prospect when combined with 5G massive MIMO technology.At present,methods and techniques on the combination of the two are under constant research.The specific research on the fingerprint localization method is carried out based on machine learning in massive MIMO single-station scenario.The main researches are as follows.(1)For the problem of poor positioning performance caused by the variability of urban environment,three kinds of fingerprints with different robustness are extracted.The first is the MACPV fingerprint extracted according to the power distribution characteristics presented by ADCPM fingerprint in angle domain,the second is the DAE fingerprint processed by denosing autoencoder based on MACPV fingerprint,and the third is the DCAE fingerprint handled by denosing convolution autoencoder based on ADCPM fingerprint.The simulation analysis using different positioning accuracy shows that as the number of scatterers decreases,the DCAE fingerprint-based positioning has better performance.(2)In view of the high cost of updating the fingerprint database in 5G network,the CAOAPFUC update algorithm is presented,which uses the method based on central AOA clustering and combines probability-based fingerprint update criterion to determine whether the fingerprint database is updated or not.The simulation results indicate that the cost of fingerprint database update can be effectively reduced with the method under the premise of the same positioning accuracy as the violent update algorithm.(3)Aiming at realizing real-time localization and simplifying the online phase process,the fast fingerprint localization method based on PQ algorithm is proposed,which improves the positioning efficiency through the codebook establishment and codeword search method based on ADCPM fingerprint.The improvement of positioning speed for the PQ algorithm is limited with the increasing size of database.Therefore,the modified CNN fast positioning method based on multitask learning is introduced which can estimate terminal position very efficiently by feedforward neural network calculation that does not depend on the size of database.The simulation results demonstrate that when the size of database is large,the advantages of the CNN-based algorithm are obvious,and the positioning speed is improved to the millisecond level under the condition that the positioning accuracy is close to the WKNN-based algorithm.更多还原