采用残差神经网络的无人机遥控信号识别监测算法

A Recognition and Monitoring Algorithm for Drone Remote Control Signals Using Residual Neural Network

为了解决无人机遥控信号在识别监测时易受随机噪声、窄带与宽带干扰影响,以及遥控信号跳频序列的跳频周期、跳频速率等特征难以准确提取的问题,提出了一种采用残差神经网络对无人机遥控信号时频谱图进行识别监测的算法(DRN-UAV)。首先通过滑动时间窗来读入时频谱图,并以联合自适应的方法计算信号频谱检测阈值;然后对已读入的时频谱图进行二值化、剔除干扰等预处理操作,构造待测谱图;接着将大量实测的不同机型遥控信号待测谱图作为数据集来训练和测试残差神经网络,进而避免跳频特征难以提取的问题;最终由训练好的网络来实时识别当前遥控信号是否存在及其所属机型。DRN-UAV算法能克服遮挡及无人机体积等不利影响,是对基于雷达或光学的反无人机系统的有效补充。实验结果表明:DRN-UAV算法使单次识别耗时约缩短为传统读入方式的1/25;在相同误检率下,DRN-UAV算法得到的信号频谱检测阈值相比传统方法降低了1.4 dBm,在不同硬件平台上都能有效增加检测范围;当信噪比高于5.5 dB时,在单个窄带定频信号和WiFi干扰下,检测错误率能达到0.01%以内。

An algorithm to monitor the control signal of drone using residual neural network is proposed to solve the problems that remote control signals of unmanned aerial vehicles(UAV)are usually susceptible to random noise and narrowband or broadband interference,and it is difficult to extract the frequency hopping period and rate of frequency hopping sequence of a remote control signal.Firstly,a time-spectrogram is obtained via a sliding time window,and a threshold of signal spectrum detection is calculated by a joint adaptive method.Then,pre-processing operations such as binarization and interference elimination are performed on the time-spectrogram to construct the time spectrum to be measured.Turther,a large number of processed spectrograms of different control signals are used as a data set to train and test the deep residual neural network,so as to avoid the problem of difficult extraction of frequency hopping features.Finally,the trained network is used to recognize the current remote control signal and its model in real time.The proposed DRN-UVA algorithm overcomes the adverse effects such as occlusion and UVA size,and is an effective supplement to anti-UVA system based on radar or optics.Experimental results show that the DRN-UAV algorithm shortens the single recognition time to about 1/25 of the traditional reading method.At the same error detection rate,the signal spectrum detection threshold obtained by the DRN-UAV algorithm reduces by 1.4 dBm compared with that of the traditional method and the detection range is effectively increased on different hardware platforms.When the signal-to-noise ratio is higher than 5.5 dB,the detection error rate can reach less that 0.01%under the interferences of single narrowband fixed-frequency signal and WiFi.