-
UWB设备中的CIR数据有上千个采样点,如此长的CIR数据作为深度神经网络的输入必然会增加网络设计的困难,同时,CIR数据作为一种时序信号,网络设计过程中必须考虑CIR数据采样点位置之间的长距离依赖性,以下重点介绍深度神经网络模型设计的核心思想。
-
文中提出的深度神经网络中,用到的基本结构包括:卷积(Convolution, Conv)层、激活函数、批量归一化(Batch Normalization, BN)层、最大汇聚(Maximum Pooling, MaxPool)层、全局平均汇聚(Global Average Pooling, GAP)层。
Conv层的作用是提取一个局部区域的特征,不同的卷积核相当于不同的特征提取器,与全连接(Fully Connected, FC)层相比,Conv层可以明显减少神经网络参数的数量。
激活函数的作用是提升深层神经网络对输入的拟合能力。修正线性单元(Rectified Linear Unit, ReLU)函数是近年来使用广泛的激活函数,其主要优点是可以有效防止梯度消失问题,且计算速度快。ReLU函数的表达式如下:
$$ {\rm{ReLU}}({\boldsymbol{x}}) = \left\{ {\begin{array}{*{20}{c}} {{\boldsymbol{x}},{\boldsymbol{x}} \geqslant 0} \\ {0,{\boldsymbol{x}} < 0} \end{array}} \right. $$ (1) 式中:${\boldsymbol{x}}$为输入向量。
BN层的作用是对神经网络中任意中间层的数据进行归一化操作,提供有效的数据预处理方法,同时还可以提高网络的泛化能力。BN层的计算表达式如下:
$$ {\hat {\boldsymbol{x}}^{(l)}} = \frac{{{{\boldsymbol{x}}^{(l)}} - \mu }}{{\sqrt {{\sigma ^2} + \varepsilon } }} \odot {\boldsymbol{\gamma}} + {\boldsymbol{\lambda}} \triangleq {{\rm{BN}}_{\gamma ,\lambda }}({{\boldsymbol{x}}^{(l)}}) $$ (2) 式中,${{\boldsymbol{x}}^{(l)}}$为第$l$层神经元的输入;$\;\mu $和${\sigma ^2}$为输入的均值和方差;$\varepsilon $为一个非常小的常数;${\boldsymbol{\gamma}} $和${\boldsymbol{\lambda}} $分别代表缩放和平移的参数向量;$ \odot $为矩阵乘法。
MaxPool层的作用是计算局部区域的最大值,缓解Conv层对位置的过度敏感性;GAP层的作用是进行特征选择,减少参数数量,避免过拟合现象的产生,可视为一种结构化的正则化器。
-
为了提取输入数据的多维特征,深度神经网络往往需要具有足够多的层数,网络层数越多,参数越多,对输入数据的拟合能力则越强。但是,当深度神经网络的层数增加到一定数量后,网络性能不会上升反而会下降,这种现象使得无法通过单纯增加网络层数来提升网络性能。为了解决这个问题,文献[12]提出了ResNet网络,其核心思想是通过在深度神经网络中引入残差连接,使梯度可以有效地流向靠近输入层的早期层,从而避免网络性能随层数增加而降低。因此,文中借鉴ResNet网络核心思想设计深度神经网络。
ResNet网络的基本结构是RBU模块,利用多个RBU模块的串联连接可以构建深层神经网络,有效防止梯度消失和梯度爆炸等问题。根据是否包含$1 \times 1$卷积块,RBU模块可分成两类,其基本构成如图1所示,其中$f(x)$是要学习的目标函数。
根据所含卷积的次数,RBU模块也可分成两类,18、34层网络的基本模块记为Basicblock,包含2次卷积;50、101、152层网络的基本模块记为Bottleneck,包含3次卷积。相较于Basicblock,Bottleneck将两个$3 \times 3$的Conv层替换为两个$1 \times 1$,一个$3 \times 3$的Conv层,Bottleneck中$3 \times 3$的Conv层首先在一个$1 \times 1$的Conv层下降维,然后在另一个$1 \times 1$的Conv层下做了还原,这样既保持了精度又减少了计算量,两种基本模块的构成如图2所示。常用的ResNet-18网络结构如图3所示。
-
CIR数据是一种时序信号,其采样点位置之间存在长距离依赖性,具有全局特征。对于卷积等local模块,主要考虑的是提取局部区域特征,难以捕获全局特征。
为了解决local模块存在的问题,文献[13]提出了Non-local模块。该模块可保证输入和输出的尺度不变性,提取全局特征,并直接嵌入任意网络;与local模块相比,该模块只需更少的层数即可达到相同效果,计算效率高。
Non-local模块的通用计算公式如下:
$$ {y_i} = \frac{1}{{A(x)}}\sum\limits_{\forall j} {s({x_i},{x_j})h({x_j})} $$ (3) 式中:$i$和$j$代表的是输入数据的位置索引;$s$为计算输入两点相似度的函数;$h$为计算输入在某个位置特征的函数;$y$为输出;$A$为对输入进行标准化处理的函数。
Non-local模块计算流程如下:
形状为$\left[ {N,H,W,C} \right]$的输入${\boldsymbol{x}}$,经过3个$1 \times 1$卷积核得到${\boldsymbol{\theta}} $,${\boldsymbol{\varphi}} $,${\boldsymbol{g}}$, 其形状均为$\left[ {N,H,W,C/2} \right]$;对${\boldsymbol{\theta}} $,${\boldsymbol{\varphi}} $,${\boldsymbol{g}}$, 将$H$, $W$两个维度展开,得到其形状为$\left[ {N,HW,C/2} \right]$;对${\boldsymbol{\varphi }}$进行通道重排,得到其形状为$\left[ {N,C/2,HW} \right]$,随后将其与${\boldsymbol{\theta}} $进行矩阵乘法,得到形状为$\left[ {N,HW,HW} \right]$的矩阵,该矩阵计算的是相似度,随后经Softmax函数进行归一化,得到的矩阵与${\boldsymbol{g}}$进行矩阵乘法,得到形状为$\left[ {N,HW,C/2} \right]$的矩阵,然后将$HW$维度进行伸展,得到其形状为$\left[ {N,H,W,C/2} \right]$;再使用一个$1 \times 1$卷积对通道进行扩展,得到其形状为$\left[ {N,H,W,C} \right]$,与初始输入${\boldsymbol{x}}$的形状一致,最后,将该矩阵与输入${\boldsymbol{x}}$相加,得到最终输出${\boldsymbol{z}}$,该计算流程如图4所示。
-
将Non-local模块插入ResNet网络的基本结构RBU模块中,并参考ResNet网络中模块的排列方式,即可构建文中提出的深度神经网络,并将其命名为NLO-ResNet,其基本模块NLRBU如图5所示,根据是否包含$1 \times 1$卷积块,将其分别命名为NLRBU-Iden和NLRBU-Conv。
图 5 NLO-ResNet网络的两类基本模块:NLRBU-Iden (a)和NLRBU-Conv (b)
Figure 5. Two basic types of modules for NLO-ResNet: NLRBU-Iden (a) and NLRBU-Conv (b)
NLO-ResNet网络可通过增加基本模块的数量来提升网络性能,但当网络层数增加到一定数量后,网络性能提升不再明显,且由于参数的增加,网络训练时间会显著增加。文中通过多次实验,并考虑训练时间和网络性能之间的平衡,所使用的NLO-ResNet网络共有22层,其网络结构如图6所示。
-
为了更好地评价所提出的深度神经网络的NLOS传播影响抑制性能,文中选取了其他几种网络进行性能比较,选取的网络分别是SVM、多层感知机(Multi-Layer Perceptron, MLP)、CNN、ResNet。
其中,SVM和MLP是传统的基于机器学习的网络,这两种网络的输入均采用6个人工提取的波形统计特征:RSSI、第一路径RSS值、平均超额延迟、均方根延迟扩展、峰度和偏度。对于SVM,将其核函数设置为径向基函数;对于MLP,将其设置为3层结构,其中,输入层共6个神经元,隐藏层共128个神经元,输出层共1个神经元。
CNN、ResNet和NLO-ResNet均为基于深度学习的网络。CNN、ResNet、NLO-ResNet这3种网络的结构相同,主要区别是组成网络的基本模块不同,3种网络的基本模块对比见图9。相较于CNN,ResNet增加了残差连接,NLO-ResNet增加了残差连接和Non-local模块。
图 9 CNN (a)、ResNet (b)和NLO-ResNet (c) 3种网络的基本模块对比图
Figure 9. Comparison of the basic modules of CNN (a), ResNet (b), and NLO-ResNet (c)
为了评估模型在数据集上的性能,引入了平均绝对误差(Mean Absolute Error, MAE),其定义如下:
$$ MAE = \frac{1}{n}\sum\limits_{k = 1}^n {\left| {{{\hat m}_k} - {m_k}} \right|} $$ (4) 式中:$\hat {\boldsymbol{m}} = \left\{ {{{\hat m}_1},{{\hat m}_2}, \cdots ,{{\hat m}_n}} \right\}$为预测值;${\boldsymbol{m}} = \left\{ {{m_1},{m_2}, \cdots ,{m_n}} \right\}$为真实值;$n$为样本数量。
-
通过重复、多次实验,并衡量准确性和效率的重要性,最终确定文中的超参数设置如下:优化算法选择Adam算法,其动量参数设置为:$\;{\beta _1} = 0.9$、${\;\beta _2} = 0.999$、$\varepsilon = {10^{ - 8}}$;学习率设置为learning rate=$3 e - 4$,批量大小设置为batch size=32,损失函数设置为MAE,迭代次数设置为epoch=50。
该深度神经网络通过Pytorch框架实现,服务器的配置为:32 GB RAM、GeForce RTX 3090 GPU。
-
实验结果如表1所示,SVM和MLP等传统的基于机器学习的网络,虽计算时间短,但性能较差。基于深度学习的网络普遍性能更优,原因是其网络层数更多、泛化能力更强。相较于SVM和MLP, CNN网络的性能提升不明显;ResNet网络的性能明显,其MAE可达0.0715 m,相较于CNN网络降低了近7.9%;在所有网络中,NLO-ResNet网络性能最优,其MAE可达0.0681 m,相较于CNN网络下降了近12.2%,相较于ResNet网络下降了近4.8%。
表 1 网络性能对比实验结果
Table 1. Experiment results of network performance comparison
Network MAE/m Raw data 0.1242 SVM 0.0782 MLP 0.0815 CNN 0.0776 ResNet 0.0715 NLO-ResNet 0.0681 训练过程中,NLO-ResNet网络在测试集上的损失函数变化曲线如图10所示。可以看到,其学习过程快速收敛。
NLO-ResNet网络在测试集上输出的预测测距误差和数据集记录的实际测距误差的对比如图11所示。从图中可以看出,在大多数情况下,预测测距误差与实际测距误差非常接近。原数据集中由于设备等因素的影响,出现了个别异常值,这些异常值影响了网络性能,未来的研究工作可以首先对数据集的数据进行平滑处理,剔除异常值,进一步提升网络性能。
综上,文中提出的基于NLO-ResNet网络的UWB NLOS传播影响抑制方法的主要优势如下:与基于机器学习的方法相比,该方法无需对输入数据进行额外处理,节省了人工提取UWB信号波形统计特征的时间,且网络性能更好;与文献[10]提出的基于CNN的方法,文献[11]提出的基于REMnet的方法相比,该方法考虑了CIR数据的时序特征,可以捕获CIR数据位置之间的长距离依赖性,提取CIR数据的全局特征,网络性能进一步提升。
Deep learning-based impact mitigation method for UWB NLOS propagation
-
摘要: 随着智能技术和设备的不断发展,精确定位技术在军事领域的应用越来越广泛,其应用场景涵盖室外和室内环境。全球卫星导航系统定位技术在室外环境中定位精度高,提供信息丰富,应用十分普遍。然而,由于墙壁、树木、玻璃等障碍物的遮挡,其在室内环境中的定位精度明显下降。超宽带技术以其定位精度高、时空分辨率强、传输速率快、成本低而显示出明显的优势。在室内环境中,各种障碍物使超宽带系统的基站和标签之间的传播通道被阻挡,由于超宽带信号的非视距传播现象,超宽带系统的定位精度明显下降。文中基于深度学习技术,提出了一种深度神经网络用于超宽带非视距传播影响抑制,该深度神经网络兼具ResNet网络和Non-local模块的优点,既可防止网络层数增加时网络性能不升反降的问题,也可捕获输入数据的全局特征,建立超宽带系统原始信道脉冲响应和测距误差之间的映射关系。相关实验结果显示,该方法可将超宽带系统在非视距传播条件下的测距平均绝对误差从0.1242 m降低至0.0681 m。与传统方法相比,该方法可消除人工统计超宽带信号波形特征耗费大量时间的缺点,可进一步提高超宽带系统在非视距传播条件下的定位精度,具有鲁棒性强、应用范围广的优点,可为军事领域室内高精度定位提供技术支撑。
-
关键词:
- 超宽带技术 /
- 深度学习 /
- 非视距传播 /
- ResNet网络 /
- Non-local模块
Abstract:Objective With the continuous development of intelligent technologies and devices, precise positioning technology in the military field is becoming increasingly widespread, and its application scenarios cover both outdoor and indoor environments. Global Navigation Satellite System (GNSS) positioning technology is commonly used for its high positioning accuracy and rich information provision in outdoor environments; However, its positioning accuracy in indoor environments is significantly reduced due to the obstruction of walls and other obstacles. Ultra-Wideband (UWB) technology shows obvious advantages with its high positioning accuracy, firm spatial and temporal resolution, fast transmission rate, and low cost. These advantages make UWB technology particularly suitable for indoor high-precision positioning. In the indoor environment, various obstacles block the propagation channel between the base station and the tag of the UWB system, due to the Non-Line-Of-Sight (NLOS) phenomenon of UWB signals, the positioning accuracy of UWB systems is significantly reduced. Therefore, it is necessary to research the impact mitigation method for UWB NLOS propagation. Methods A deep neural network based on deep learning techniques is proposed for UWB NLOS propagation impact mitigation. This deep neural network takes the initial channel impulse response (CIR) of the UWB device as input and the ranging error of the UWB device as output. The experimental analysis shows that the characteristics of CIR data are significantly different under LOS and NLOS propagation conditions (Fig.7), which provides a solid theoretical basis for establishing the mapping relationship between CIR and ranging error using deep learning methods. Meanwhile, the network performance is related to the dimensionality of the input CIR data. The network performance is best when the input CIR data is 128 dimensions (Fig.8). When the input of the deep neural network is 128-dimensional data, too long input will lead to the structural design of the network becoming difficult. And the number of network layers is too small, the network performance can not meet the requirements to achieve good NLOS propagation impact mitigation effect; After the number of network layers increases to a certain degree, the network performance will decrease with the increase of the number of layers. For this reason, the ResNet network is selected in this paper, which enables the gradient to flow effectively to the early layers near the input layer by introducing residual connections in the deep neural network, thus improving the network performance with the increase of layers. At the same time, CIR data, as a time-series signal, correlates its data points. The global features of CIR data must be considered, while local module such as convolution can only extract local features. For this reason, this paper introduces the Non-local module, which can capture the long-distance dependence between locations and extract global information. In summary, the proposed deep neural network is constructed by inserting the Non-local module into the ResNet network's basic module while considering the CIR data's features, and named the deep neural network as NLO-ResNet. Results and Discussions In order to evaluate the NLOS propagation impact mitigation performance of the proposed deep neural network, four networks were selected for performance comparison. Four networks include two machine learning-based networks, SVM and MLP, and two deep learning-based networks, CNN and ResNet. Experimental results (Tab.1) show that, due to the increase in the number of layers of the network and the change in the input data, the performance of the deep learning-based network is generally better than that of the machine learning-based network; Among the deep learning-based networks, the CNN network has the worst performance, the ResNet network improves with the increase of the number of layers due to the introduction of residual connections, and the NLO-ResNet network has the best performance, which has the most comprehensive feature extraction of the input CIR data. The mean absolute error (MAE) is reduced by 12.2% compared to the CNN-based network and 4.8% compared to the ResNet-based network, and the learning process of this network converges quickly (Fig.10), and the predicted range error of this network is very close to the actual range error (Fig.11). Conclusions To improve the accuracy of UWB systems under NLOS propagation conditions, a deep learning-based NLOS propagation impact mitigation method is proposed, which constitutes a deep neural network by inserting a Non-local module into the basic module of the ResNet network. The method can reduce the MAE of the original data from 0.1242 m to 0.0681 m. The research provides technical support for indoor high-precision positioning in the military field. The related results can be applied in the autonomous takeoff and landing of military UAVs, and indoor positioning of military robots. -
Key words:
- Ultra-Wideband /
- deep learning /
- Non-Line-of-Sight /
- ResNet /
- Non-local
-
表 1 网络性能对比实验结果
Table 1. Experiment results of network performance comparison
Network MAE/m Raw data 0.1242 SVM 0.0782 MLP 0.0815 CNN 0.0776 ResNet 0.0715 NLO-ResNet 0.0681 -
[1] Subedi S, Pyun J. A survey of smartphone-based indoor positioning system using RF-based wireless technologies [J]. Sensors, 2020, 20(24): 7230. doi: 10.3390/s20247230 [2] Sahinoglu Z, Gezici S, Guvenc I. Ultra-Wideband Positioning Systems: Theoretical Limits, Ranging Algorithms, and Protocols[M]. New York: Cambridge University Press, 2008. [3] 文铠. 基于超宽带的室内定位与定向关键技术研究[D]. 武汉: 武汉大学, 2020. Wen Kai. Research on the key techniques of ultra-wideband based indoor position and orientation estimation[D]. Wuhan: Wuhan University, 2020. (in Chinese) [4] Che F, Ahmed Q, Fontaine J, et al. Feature-based generalized Gaussian distribution method for NLOS detection in ultra-wideband (UWB) indoor positioning system [J]. IEEE Sensors Journal, 2022, 22(19): 18726-18739. doi: 10.1109/JSEN.2022.3198680 [5] Wu S, Ma Y, Zhang Q, et al. NLOS error mitigation for UWB ranging in dense multipath environments[C]//Proc of IEEE Conference on WCNC, 2007: 1565-1570. [6] Venkatesh S, Buehrer R. NLOS mitigation using linear programming in ultrawideband location-aware networks [J]. IEEE Transactions on Vehicular Technology, 2007, 56(5): 3182-3198. doi: 10.1109/TVT.2007.900397 [7] Yang H, Wang Y, Seow C, et al. UWB sensor-based indoor LOS/NLOS localization with support vector machine learning [J]. IEEE Sensors Journal, 2023, 23(3): 2988-3004. doi: 10.1109/JSEN.2022.3232479 [8] Nguyen T, Jeong Y, Shin H. Machine learning for wideband localization [J]. IEEE Journal on Selected Areas in Communications, 2015, 33(7): 1357-1380. doi: 10.1109/JSAC.2015.2430191 [9] Yang X, Zhao F. Echo state network and echo state Gaussian process for non-line-of-sight target tracking [J]. IEEE Systems Journal, 2020, 14(3): 3885-3892. doi: 10.1109/JSYST.2020.2982516 [10] Bregar K, Mohorcic M. Improving indoor localization using convolutional neural networks on computationally restricted devices [J]. IEEE Access, 2018, 6(1): 17429-17441. [11] Angarano S, Mazzia V, Salvetti F, et al. Robust ultra-wideband range error mitigation with deep learning at the edge [J]. Engineering Applications of Artificial Intelligence, 2021, 102: 104278. doi: 10.1016/j.engappai.2021.104278 [12] He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proc of IEEE Conference on CVPR, 2016: 770-778. [13] Wang X, Girshick R, Gupta A, et al. Non-local neural networks[C]//Proc of IEEE Conference on CVPR, 2018: 7794-7803. [14] Angarano S, Salvetti F, Mazzia V, et al. Deep UWB: A dataset for UWB ranging error mitigation in indoor environments [EB/OL]. [2023-04-10]. https://www.zenodo.org/record/6611037#.ZBGEN3Yz.