Fund Project:Project supported by the National Natural Science Foundation of China (Grant No. 61331007)
Received Date:08 March 2019
Accepted Date:29 May 2019
Available Online:01 September 2019
Published Online:05 September 2019
Abstract:Because of the strong non-linear fitting capability, the artificial neural network (ANN) can be used to establish the mapping relationship between the terminal position and the received signal for obtaining the channel characteristics at different locations. The accuracy of an ANN model is, in general, determined by the number of the training sets used in constructing the model. The more the training sets, the better the accuracy is. However, getting a large number of training sets by deterministic model is expensive. Therefore, under the same number of training sets, improving the accuracy of the model is crucial to develop an effective time reversal (TR) modeling method based on ANN. In this paper, a new TR channel modeling method based on the back propagation neural network is proposed. Genetic algorithm (GA) with excellent global search capability is used to optimize the weight and threshold of the ANN to avoid the possibility of the ANN falling into local minimum. According to the basic principle of time reversal, the peak characteristics are obtained by the fitting method. In order to improve the accuracy of the model, the peak value characteristics are integrated into the GA as empirical knowledge to change the fitness function. Meanwhile, the principal component analysis technology is utilized to process data, which reduces the data dimension and the training time of ANN while data characteristics are ensured. Once the terminal antenna positions are input to the proposed model, the accurate TR received signals can be quickly obtained. Finally, the deconvolution operation of the received signal is performed by the clean algorithm to obtain the channel characteristics. A simple indoor TR scenario is used as an example to demonstrate the effectiveness of the proposed method. The results show that the three channel characteristics obtained by the model, i.e., channel impulse response peak value, 15 dB multipath number, and average delay, have high accuracy. Furthermore, the proposed model has more excellent performance than the other two ANN models under the condition of the same number of training samples. Based on the basic principle of TR technology, the electromagnetic waves have better focusing effect in more complex environments. Therefore, the proposed method is also applicable to more complicated environments than the simple indoor scenario. Keywords:time reversal/ channel modeling/ artificial neural network/ empirical knowledge
基于经验知识的GA-BP模型获取TR电磁信道特性的流程图如图3所示. 首先利用训练样本的数据进行多项式拟合, 并得到不同终端位置${{{x}}_k}$所对应的接收信号峰值估计值, 将其作为经验知识用于计算遗传算法的适应度函数. 随后利用训练样本对GA-BP神经网络进行训练. 通过遗传算法的迭代优化, 不断调节神经网络的权值和阈值, 使神经网络误差函数不断减小, 最终确定模型内部参数. 最后利用模型得到对应于不同终端位置${{{x}}_k}$的$m$维输出数据${\tilde {{p}} _k}$. 将这些$m$维数据利用特征矩阵, 恢复为$M$维数据, 即获得对应终端位置的接收信号. 利用CLEAN算法[24,25]对接收信号进行处理, 即可获得对应的TR信道特性. 图 3 信道特性获取流程图 Figure3. Flowchart of the proposed model to obtain channel characteristic.
3.应用实例为了验证上述方法的有效性, 同时减少获取样本所耗的内存和时间, 对一个简单的室内场景进行信道建模, 场景如图4所示. 尺寸为100 cm × 100 cm × 100 cm的室内空间, 边界是相对介电常数为8、厚度为20 cm的材料, 用来模拟混凝土墙壁. 在室内放置4个TRM天线, 位置坐标如表1所列. 终端天线首先发射正弦调制高斯脉冲, 其中心频率为${f_0} = 5.5\;{\rm{GHz}}$. TRM天线将接收到的信号进行时间反演处理, 随后再发射, 最后在终端天线获得聚焦信号. 终端天线的位置如表2所列. 图 4 仿真场景俯视图 Figure4. Top view of the simulation scene.
X/cm
Y/cm
Z/cm
TRM1
2.5
0
0
TRM2
–2.5
0
0
TRM3
7.5
0
0
TRM4
–7.5
0
0
表1TRM天线位置 Table1.Location of the TRM antennas.
坐标最小值/cm
坐标最大值/cm
X
10
30
Y
10
30
Z
0
30
表2终端天线的位置 Table2.Location of the terminal antenna.
利用全波电磁仿真软件FDTD Solutions获得了44个样本数据, 将其中36个数据作为训练样本来训练神经网络, 剩下8个数据作为测试样本, 用来验证方法的有效性. 对36个训练样本数据进行处理, 随后采用最小二乘法获取多项式系数, 最终用11阶多项式近似表征了终端接收信号的峰值与距离$r$和辐射强度之间的函数关系, 如图5所示. 考虑到仿真所采用的理想偶极子天线的辐射特性以及TRM天线具有对称性, 我们用$\sin \theta $函数来近似表征偶极子辐射方向图, 其中, $\theta $为坐标原点指向终端天线的天顶角. 图 5 11阶多项式拟合结果与仿真数据的对比 Figure5. Comparison of the results of 11th order polynomial fitting and simulation data.
随后利用拟合函数对测试样本中的接收信号峰值进行粗略估计, 并将其作为经验知识用于遗传算法中. 采用单隐层神经网络, 并用Hecht-Nelson方法[21]根据输入向量维数$N$确定神经网络的隐层节点数为$2N + 1$. 利用训练样本对模型进行训练, 完成训练后, 获得测试样本对应位置的接收信号. 图6为采用本方法与采用电磁仿真软件得到的两个测试样本接收信号的对比. 可以看出本模型获得的波形与仿真波形存在一定的时移, 但整体波形基本一致. 此外, 因为PCA技术具有一定的降噪效果, 利用本模型获得的接收信号受到的噪声影响比仿真结果小. 图 6 利用本模型获得接收信号与仿真获得信号的对比 (a)测试样本1; (b)测试样本2 Figure6. Comparison of the signals of the proposed model and simulation: (a) Test sample #1; (b) test sample #2