基于组合神经网络的UWB室内定位方法研究  

Research on UWB indoor localization methods based on combined neural networks

作　　者：潘镐铖 范皓然 陈建飞 PAN Haocheng;FAN Haoran;CHEN Jianfei(College of Electronic and Optical Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学电子与光学工程学院,江苏南京210003

出　　处：《电子设计工程》2024年第21期1-7,共7页Electronic Design Engineering

基　　金：国家自然科学基金(61601237);江苏省研究生科研与实践创新计划项目(SJCX21_0280)。

摘　　要：由于室内环境复杂,传统的超宽带(UWB)室内定位技术仅仅采用几何算法对待测目标的位置进行解算,误差较大且不可控。为了提高物体在运动状态下的定位精度,该文提出了基于CNNLSTM组合神经网络的定位预测模型。为了提高模型预测性能,在数据预处理阶段利用MLP对海量的UWB信道数据进行学习,训练NLOS/LOS分类算法。剔除NLOS数据后将各基站解算的测距信息按时间顺序作为整个预测网络的输入,借助CNN层提取时间序列上表征能力强的高层特征,由LSTM层处理具有时间连续性的定位信息,并应用了自适应学习率算法加快收敛速度。通过与单一LSTM神经网络和BP神经网络的对比验证了CNN-LSTM网络模型定位精度更高,相比单一LSTM神经网络误差控制性能提升了约69%,平均精度误差控制在0.06 m左右。Due to the complexity of indoor environments,traditional Ultra Wide Band(UWB)indoor positioning technologies only use geometric algorithms to calculate the position of the target with large and uncontrollable errors.In order to improve the positioning accuracy of moving objects,a location prediction model based on a combination of CNN⁃LSTM neural networks is proposed.To improve the predictive performance of the model,a MLP is used in the data preprocessing stage to learn and train NLOS/LOS classification algorithms for massive UWB channel data.After removing NLOS data,the ranging information solved by each base station is used as the input to the entire prediction network in chronological order.The high⁃level features that characterize the time series are extracted using the CNN layer,and the LSTM layer processes the location information with time continuity.An adaptive learning rate algorithm is applied to accelerate convergence speed.Compared with a single LSTM neural network and a BP neural network,the CNN⁃LSTM network model in this paper has higher positioning accuracy,with an error control function improvement of approximately 69%compared to the single LSTM neural network,and an average precision error controlled at around 0.06 m.

关 键 词：室内定位 卷积神经网络 长短时记忆网络 深度学习 自适应学习率 

分 类 号：TN961[电子电信—信号与信息处理]