基于变换核高斯回归模型的无线环境地图构建方法  

Radio Environment Map Construction Method Based on Transformed Kernel Gaussian Process Regression Model

作　　者：徐逸群 张邦宁 张晓凯 郭道省 XU Yiqun;ZHANG Bangning;ZHANG Xiaokai;GUO Daoxing(College of Communications Engineering,Army Engineering University of PLA,Nanjing 210007,China)

机构地区：[1]陆军工程大学通信工程学院,江苏南京210007

出　　处：《无线电工程》2022年第11期1908-1921,共14页Radio Engineering

基　　金：江苏省自然科学基金(BK20191328)。

摘　　要：针对无线环境地图(Radio Environment Map,REM)构建中复杂的地形和地物环境对电波传播影响的问题,提出一种基于变换核高斯回归模型的REM构建算法,并在此构建算法的基础上,提出了2种无线电频谱监测空间选点算法。将高斯过程引入到REM构建中,讨论了环境异质性的影响,设计了3种非线性核变换函数表征电波传播环境的异质性,从群智感知设备采集的数据中学习环境信息,并将这些信息隐含于核变换函数的非线性变换参数中以提升REM构建的准确性。在此基础上,通过优化频谱监测空间选点方案,提升REM中数据的有效性,以较少的数据达到更好的REM构建性能。将基于变换核高斯模型的构建算法与几种经典的REM构建算法进行了对比分析,仿真实现了2种空间选点方案并与传统的均匀随机选点的性能进行了对比。结果表明,当群智传感器的数量足够多、算法可以捕捉到局部环境的异质性时,所提出的变换核学习(Transformed Kernel Learning,TKL)算法优于其他经典的REM构建方法。在有限个可数点集范围内进行选点,互信息最大化(Greedy-based Mutual Information Maximum,GMIM)空间选点算法与均匀随机选点相比,REM构建的RMSE可以降低1.23 dB,连续分级概率评分(Continuous Ranked Probability Score,CRPS)降低0.64;对整个空间任意位置进行选点时,变分推断(Variational Inference,VI)空间选点算法与均匀随机选点相比,REM构建的RMSE可以降低1.56 dB,CRPS降低0.73。A radio environment map(REM)construction algorithm based on the transformed kernel Gaussian process regression(TKGPR)model is proposed.Meanwhile,two monitoring device placement algorithms are proposed based on the TKGPR model.It is the first time that Gaussian process is introduced into the construction of REM,and the impact of environmental heterogeneity is discussed.Three nonlinear kernel transformation functions are designed to characterize the heterogeneity of the radio signal propagation environment.The environmental information is learned from the data collected by the crowd-sourced devices,and this information is implicit in the nonlinear transformation parameters of the kernel transformation to improve the accuracy of REM construction.By optimizing the spectrum monitoring placement scheme,the REM data′s effectiveness is improved,and a better REM construction performance can be achieved with fewer data.The construction method based on the TKGPR model is compared with several classical REM construction methods.The results show that the proposed TKL algorithm outperforms other classical wireless environment map construction methods when the number of crowd-sourced sensors is large enough for the algorithm to capture the heterogeneity of the local environment.For sensor placement within a finite set of countable points,the GMIM sensor placement algorithm can reduce the RMSE of REM construction by 1.23 dB and the CRPS by 0.64 compared with uniform random sensor placement.For point sensor placement at any location in the whole space,the VI algorithm can reduce the RMSE of REM construction by 1.56 dB and the CRPS by 0.73 compared with uniform random sensor placement.

关 键 词：无线环境地图 频谱监测 高斯过程 核学习 

分 类 号：TN911[电子电信—通信与信息系统]