Abnormal Signal Detection Methods Of UAVs Based On Convolutional Neural Networks

Unmanncd Aerial Vehicles(UAVs)have been widely used in military and civilian fields due to their small size,low cost,high flexibility,and strong maneuverability.In the military fields,UAVs are mainly used for missions with high risk,bringing substantial military benefits.UAVs mainly assist people in daily tasks in the civilian feilds,such as farmland irrigation,urban management,and geological exploration.However,various security threats gradually increase by continuously expanding the scopes of UAVs’ applications.Among all the security threats,false spoofing attack is a typical network threat faced by UAVs.At present,several solutions for solving the false spoofing attacks suffered by UAVs have been proposed at home and abroad.However,when detecting abnormal signals in false spoofing attacks of UAVs using deep learning,it is challenging to classify abnormal signals of UAVs due to similar data characteristics of abnormal signals and normal signals.Meanwhile,the collected signals of UAVs will be affected by interference,making it more challenging to classify abnormal signals for UAVs,thus affecting the detection accuracy of abnormal signals for UAVs.This thesis proposes abnormal signal detection methods for UAVs based on Convolutional Neural Networks to solve the problem.The main contents of this thesis are as follows:(1)In order to solve the problem that the detection accuracy of abnormal signal for UAVs is affected by the similar appearance and characteristics of normal signals data and abnormal signals data for UAVs,which makes it more challenging to classify abnormal signals.Therefore,this thesis proposes an abnormal signals detection method of UAVs based on Double Shortcuts Zero-Bias ResNet.First,the signals of UAVs are preprocessed by removing random noise,normalization,and Fourier transforms,then the frequency domain signals are obtained.Second,a new Double Shortcut Residual Network is constructed to extract characteristics of UAVs signals in the frequency domain.Finally,through the fusion of the Zero-Bias Fully Connected layer and the original Fully Connected layer,the network structure of the last layer of the Double Shortcuts ResNet model is optimized to obtain the final Double Shortcuts Zero-Bias ResNet,so as to improve the correlation degree of between the Fully Connected layer and Softmax layer.Simulation results indicate that when the data of ADS-B signals for UAVs set with good data quality is used,the proposed method can improve the detection accuracy compared with the existing abnormal signal detection methods.(2)In real scenes,there is inevitably a large amount of random noise in the collected UAV signals,which leads to interference of original signals of UAVs,thus affecting the detection accuracy of abnormal signals for UAVs.This thesis proposes an optimization detection method of abnormal signals of UAVs based on bimodal perception.First,the collected UAV signals are modally transformed to generate waveform images.Second,the Double Shortcuts Zero-Bias ResNet is used to extract bimodal features from the waveform image modal data and frequency domain signal modal data of original signals for UAVs.Then,the Bag-Stack ensemble fusion method is used to fuse the classification results of different modal data to alleviate the interference of random noise to the original signal of UAVs and improve the detection accuracy of abnormal signals on UAVs.Simulation results indicate that when the original ADS-B signal data set of UAVs with random noise is used,the proposed method can improve the detection accuracy compared with the existing abnormal signal detection methods.To sum up,the development and completion of this thesis have a particular theoretical significance and practical value for detecting abnormal signals of UAVs.