Research On Wireless Network Devices Recognition Method Based On Transfer Learning

In the modern battlefield,electronic warfare has become the key to victory,and the most important part of electronic warfare is the reconnaissance of the enemy,so individual recognition of wireless network devices has become a top priority.With the development of electronic technology,a variety of wireless network devices has been applied to the battlefield.The electromagnetic environment on the battlefield has become more complex and changeable.These have added challenges to individual recognition tasks.Under this background,this paper conducts research with the goal of achieving individual recognition of wireless network devices,and applies transfer learning and deep learning to the field of individual recognition.The main research contents are as follows:(1)Firstly,the fingerprint sources of individual wireless network devices and the types of wireless frames are analyzed.Then,we build a hardware experiment platform according to actual environment and collect various wireless frame signals generated by individual wireless network devices.At last,the data preprocessing work are performed.(2)Individual recognition of wireless network devices based on transient characteristics are analyzed.Firstly,the positioning of the starting point of the signal is studied,and two starting point positioning methods of short-term energy detection method and fuzzy entropy method are used and compared.Then the transient signal envelope is obtained using the method based on Hillbert transform and least squares polynomial fitting,and then the mathematical characteristics of the transient signal envelope are extracted,and the weight of the feature is determined using the entropy value weighting method.We get characteristic vector of transient signal to study.Finally,experiments were conducted on individuals of the same model and different models of wireless network devices to prove the effectiveness of the method.(3)We discuss the existence of both the target domain and the source domain,and propose an inductive instance-based transfer learning method-an improved Tr Ada Boost algorithm.This algorithm compares the Tr Ada Boost algorithm with two aspects of initial weight setting and weight update rate.Optimized to achieve the transfer of source domain knowledge to the target domain,the risk of Tr Ada Boost algorithm negative transfer and overfitting are reduced.Experimental results show that the recognition effect of this algorithm is significantly improved compared with that before the improvement.(4)We discuss the situation where there is no source domain to transfer knowledge and the target domain data is difficult to label.An individual recognition method combining convolutional neural network and generative adversarial network and transfer learning.First,a convolutional neural network that can classify the data in the target domain is designed,and then the unlabeled data in the target domain is input into the generative adversarial network for unsupervised pre-training,then the model of the discriminant network is transfered to the convolutional neural network.In the network,the labeled data is finally used to fine-tune the convolutional neural network.Experiments show that this method can effectively solve the problem of low recognition rate in the case of "small samples" and "less labels".