基于多尺度窗口和区域注意力残差网络的无线电力终端身份识别方法

Wireless Power Terminal Identification Method Based on Multiscale Windowed Deep Residual Network

针对现有无线通信设备信号识别方法需对信号进行域变换、增加网络输入数据维数的问题,该文提出基于多尺度窗口区域注意力残差网络的无线电力终端身份识别方法。首先,通过所提多尺度窗口模块完成信号前导码在各个周期尺度下的信息交互,使网络能够直接处理并识别原始无线通信信号数据;然后,设计区域注意力模块,以显著特征区域均值为评价指标对通道资源进行重新分配,提高了网络对信号局部特征的学习能力;最后,以池化分类器替代全连接层,采用Adam优化器进行梯度更新完成训练过程。实际采集无线信号数据实验结果表明,设计的各模块可显著提升网络的训练与识别性能,相同型号设备识别准确率提高至97.316%,非法设备的检测率达82.8%,可有效增强电力系统的无线通信安全。

The technique of wireless terminal identification based on the differential characteristics of wireless signals is currently an important physical layer security mechanism. However, traditional wireless signal identification methods generally require signal-signal domain conversion. Therefore, the dimensionality of the data and the arithmetic power requirements are enhanced. This can increase the application cost of this security mechanism. To solve this problem, a residual network-based wireless power terminal identification method is proposed. With the designed multiscale window module and area attention module, it can directly process the signal raw data to accurately identify the legal device identity and illegal device detection.First, the proposed multiscale window module completes the information interaction of the signal precursor code at each cycle scale, enabling the network to directly process and identify the raw wireless communication signal data. Then, the regional attention module is designed to reallocate channel resources with the mean value of significant feature regions as the evaluation index, which improves the network’s ability to learn local features of signals. Finally, a pooling classifier is used to replace the fully connected layer, and the Adam optimizer is used for gradient update to complete the training process. In this model, the multiscale window module makes use of the leading code subframe feature, which can directly process the original signal data and greatly improve the learning performance recognition effect.The experimental results on the actual collected wireless terminal signal data show that the multiscale module improves the recognition accuracy by 31% compared with the traditional residual network due to the comprehensive consideration of the information of each subframe of the leading code signal. The recognition accuracy and learning performance of the network are significantly improved, which verifies the effectiveness of the module on network performance improvement. The addition of the regional attention mechanism further improves the recognition accuracy while improving the training performance. It is verified that the region attention mechanism can effectively improve the performance of the network in recognizing signal distortion features. The recognition accuracy is up to 97.316% for 30 identical models of commercial devices. Also, the maximum value of the output probability corresponding to the label is selected as the identification result of the identity, and the threshold of the output probability is defined as the confidence level, which can also detect the illegal devices while further scientific evaluation. Five experiments of uninvolved training devices are selected for illegal detection, and the results show that when the confidence level is 99%, the detection rate of illegal devices reaches 82.8%.The following conclusions can be drawn from the analysis of the experimental results:① The constructed multiscale module avoids the loss of local difference features caused by the analysis of the signal as a whole because it considers the information of each subframe of the leading code signal, which significantly improves the training performance and recognition accuracy.② The proposed regional attention mechanism further enhances the learning of regional difference features by dividing the channel into regions, which achieves a significant improvement in the training performance and This achieves further improvement of training performance and recognition accuracy.③ The proposed identification method based on multi-scale windowed regional attention residual network achieves self-learning of the difference features of the original I/Q data of wireless signals. The recognition accuracy of the same model device can reach 97.316%, and the recognition rate of illegal devices reaches 82.8%.