Identification Method Of Radar Emitter Type Based On Transfer Learning

With the significant improvement of the confrontation degree in the modern electromagnetic battlefield,the radar data intercepted by the reconnaissance receiver presents the complex characteristics of big data and small samples.It is one of the key problems in the field of radar reconnaissance that how to effectively perceive the types of radar emitter by using radar data with this feature.Therefore,this paper introduces the idea of transfer learning into the field of radar type identification and studies the theory and technology of radar type identification based on transfer learning to meet the needs of radar threat perception under complex data conditions.The main research contents and achievements of this paper are as follows.1.The knowledge transfer mechanism of radar is analyzed.Based on the characteristic parameters of radar signals,the differences and correlations of radar samples between different types and different working modes of the same type are studied,and the necessity and feasibility of applying knowledge transfer to model identification are expounded.2.Aiming at the problem that the characteristic parameters of radar signals have different distributions in different working modes,which leads to the difficulty of radar types identification,a radar emitter type identification method based on characteristic knowledge transfer is proposed.Firstly,a feature reconstruction model combining spatio-temporal information is constructed,which makes the sample feature distribution have better intra-class aggregation and inter-class divergence characteristics,and provides feature distribution knowledge for subsequent transfer learning.Then,a distribution measurement formula based on higher-order moments is proposed to adaptively determine the overall distribution difference among samples.By reducing the distribution difference,the distribution alignment(characteristic knowledge transfer)of sample characteristic parameters in different working modes is realized.After the transfer learning,the classifier trained by a single working mode sample can identify all samples of the same type,thus reducing the difficulty of constructing the classification model.The experimental results show that the type identification accuracy of this method is 87.92% in the scenario set in this paper,which is 26.93% and 2.68% higher than the non-transfer identification method and the existing transfer identification method,respectively.When25% of the pulses are missing in radar data,the recognition accuracy of this method drops to 74.78%.When the false pulse is 25%,the offset recognition effect of this method is not ideal and the recognition accuracy is only 43.96%.3.Aiming at the problem that the number of samples intercepted by reconnaissance is too small to meet the training requirements of recognition model,which leads to poor classification performance,this paper proposes a radar emitter type identification method based on model parameter transfer.Based on the Model-Agnostic algorithm in meta-learning framework,this method embeds relevant radar domain knowledge into model parameters effectively.After the model parameters are fine-tuned(parameter transfer)by small samples,they can be transformed into the classification model of the samples to be identified.At the same time,in order to solve the problem of poor feature extraction ability of the original network of the algorithm in the case of small samples,this paper proposes a backbone network based on mixed attention mechanism as the knowledge embedding model of the algorithm.This model has better knowledge transfer performance.The experimental results show that this method has a good recognition effect under the condition of small samples,and the less samples,the more obvious the recognition performance advantage.Under the condition of 10 samples,the classification accuracy reaches 80.71%,which is 11.3% higher than the existing migration method Fine-tune.Under the case of 25% missing pulse and 25% false pulse,the recognition accuracy of this method drops to 70.52% and 65.81% respectively.