Research On Key Technologies For The Open-set Recognition Of Radar Emitter

Radar emitter identification is the key means of winning electronic warfare.In practice,the electronic reconnaissance receivers are often faced with open electromagnetic environments.The type and number of radar emitters in the environment are constantly changing,but the capacity of the signal database is limited,so the new emitters which are completely unknown may appear during identification.Most existing methods do not have open-set recognition capability and may incorrectly identify a new emitter as some known one.When using the current widely used deep learning-based methods,if samples of a radar emitter signal class not included in the training stage appear in the recognition stage,it can cause recognition errors and seriously affect the reliability of the results.This dissertation focuses on solving the radar emitter identification problem in openset scenarios,and investigates the methods of radar signal modulation type open-set recognition and radar specific emitter open-set identification.And an incremental open-set recognition method is proposed to solve the problem of model update.The main work of the dissertation includes the following points:1.To solve the problem that the existing methods based on time-frequency images and deep learning for radar modulation type recognition cannot be adapted to open electromagnetic environments,the feature space is optimized based on reciprocal point learning.As a result,the samples of known classes are pushed to the edge and separated from each other,and those of the unknown class are distributed in the center.A method of determining adaptive threshold is proposed to achieve open-set recognition based on this feature.The attention mechanism module is added to the feature extraction network to make it pay more attention to the effective information in the energy-concentrated part of the time-frequency image.To address the problem of high complexity and information loss in computing the time-frequency image,this dissertation proposes to use the one-dimensional convolutional residual network instead of the two-dimensional convolutional network to extract the features and use the real and imaginary parts of the signal spectrum sequence as the input.The experimental results show that the proposed method has good adaptability in the open electromagnetic environment.And the method based on one-dimensional network has significant advantages under the low signal-to-noise ratio(SNR)conditions.2.To solve the problem that the fingerprint features of radar emitters are very fine and difficult to extract,the Selective Kernel Residual Network(SKRes Net)with multiple convolutional kernel size branches is proposed,which can change different branch weights adaptively according to the input data.The Open-SKRes Net model is obtained by using Open Max to improve the full connection layer of SKRes Net.It can complete the task of open-set recognition.The network is trained by combining center loss and cross-entropy loss to limit the range of known classes in the feature space and thus reduce the open space risk.A general preprocessing method for signal of radar specific emitter is proposed in this dissertation to reduce the redundancy of the input.It extracts the transient part of the signal pulse to replace the complete pulse as the input of the network.The experimental results show that SKRes Net has a strong feature extraction capability.The proposed method can effectively control the level of risk,and complete the radar specific emitter open-set identification.3.To solve the problem that the open-set recognition model cannot expand the identifiable categories according to the changing electromagnetic environment,an incremental open-set recognition method based on unbiased cosine classifier is proposed.Using the unbiased cosine classifier instead of the fully connected and Softmax layers for classification alleviates the classification preference of the classifier for the newly added categories.The prototypes of each class are used to represent the samples and make the value of the coefficient term of the corresponding class of the cosine classifier equal to the prototype vector.Thus,the outputs of the classifier are the cosine similarities between the samples and the prototypes of each class.Then set the adaptive threshold to achieve open-set recognition.In the initial training stage of model,the Arc Face loss is used to reserve more embedding space for future new classes and reduce the open space risk in the identification process.In the incremental stage,it is only needed to calculate the prototypes of the new classes and expend the classifier weight terms to achieve the model update.The experimental results on multiple data sets show that the proposed method can effectively solve the problem of radar incremental open-set recognition,and the number of labeled samples for the newly added classes is less required.