A comparative study on NLOS error elimination methods based on channel measurement experiment

In order to study the performance of different elimination methods on the distance estimation forward error caused by the non-line-of-sight (NLOS) propagation of radio signals, this paper is based on the mean value, root mean square delay spread, skewness, kurtosis and peak-to-average ratio extracted from the channel state information (CSI), and combine it with the logarithmic estimated distance based on the time of arrival (TOA) as the feature input vector, through the establishment of Gaussian process regression (GPR), least square support vector machine regression (LS-SVMR) and BP neural network training model for experimental performance comparison. Through the actual measurement of the 2.4 to 5.4 GHz wireless propagation channel in the typical indoor environment, the error elimination experiment is carried out to compare the NLOS error elimination performance under different input characteristics, different bandwidths and different frequency bands. The experimental results show that the GPR model has the best NLOS error elimination performance, and the extracted CSI multi-features as the input of the GPR model can reduce the average absolute error and root mean square error by 71.12% and 81.36%, respectively. As the bandwidth continues to increase, the error elimination performance is gradually optimized. By increasing the bandwidth, the NLOS positioning error when the input features are less can be effectively improved. The positioning error of the low frequency band is smaller than that of the high frequency band under the multi-features, so the combination of all available frequency bands can eliminate the NLOS positioning error better than a single frequency band. 为了研究不同消除方法对无线电信号由非视距(NLOS)传播而产生的距离估计正偏误差的消除性能, 基于信道状态信息(CSI)提取出均值、均方根延迟扩展、偏度、峰度、峰均比特征, 并将其与基于到达时间(TOA)的对数估计距离相结合作为特征输入向量, 通过建立高斯过程回归(GPR)、最小二乘支持向量机回归(LS-SVMR)与BP神经网络训练模型进行实验性能比较。对实际测量的典型室内环境中2.4~5.4 GHz的无线传播信道进行误差消除实验, 比较不同输入特征、不同带宽和不同频带下的NLOS误差消除性能。实验结果表明：GPR模型表现出最好的NLOS误差消除性能, 且所提取的CSI多特征作为输入向量可以将平均绝对误差(MAE)和均方根误差(RMSE)分别减小71.12%和81.36%;随着带宽不断增加, 误差消除性能逐渐优化, 即可通过增大带宽有效地改善输入特征较少时的NLOS定位误差; 在多特征输入下, 低频带的NLOS测距误差与高频带不同, 因此将所有可用的频带结合可以比单频带更好地消除NLOS定位误差。