Research On Feature Selection Method Of Radar Signal Recognition

With the using of new system radars in the modern informationization battlefield,radar signal recognition has become the key issue that needs to be solved first in radar countermeasures.The radar signal recognition technology extracts and analyzes the parameters to judge the radar type,carrier,use and threat level of the target radar,affects the next operational decision.The development level of this technology has become an important indicator to measure the effectiveness of electronic reconnaissance system.Based on the feature selection algorithm,this paper studies the radar signal recognition methods,they overcome the weaknesses of large storage cost,complicated process,low recognition accuracy or unexplain characteristics in existing recognition technology.The subject extracts a variety of artificial features,merges them into an original feature set,and then uses feature selection algorithms based on Fisher-score or deep neural network to optimize the feature set.The specific research work is as follows:1.The paper studies several typical intrapulse modulation signals,analyzes the modulation characteristics through time domain waveform graphs and spectrums.This paper extracts complexity features in the frequency domain,extracts wavelet ridge frequency Cscade-Connection features in the time-frequency domain,extracts the information entropy cascade features in the Wigner-Ville distribution,extracts the ambiguity function features in the time-frequency domain,and merges them into the original feature set.In faced the problem of redundancy,the paper adopts the Fisher-score feature selection algorithm to score the inter-class and intra-class distances of each feature,and deletes redundancy features to improve classification performance.2.Deep neural network has the powerful ability of autonomous data learning.It is proposed to apply the deep feature selection network to radar signal recognition research.The model can capture the weightiness of each feature in the classification through training,measure if the feature is redundant or not,so as to implement weighting on the input level.The article optimizes the model parameters through parameter optimization experiments and obtains satisfactory results.It overcomes the problem that the features extracted by deep learning network are "black box" and unexplained,which embodies the superiority of the algorithm.3.Aiming at the problem that the gradient diffusion and the local optimal solution in the deep feature selection network,the feature selection network based on the stacked contractive autoencoder is proposed to perform radar signal recognition.The network uses greedy layer-wised unsupervised pre-training and supervised fine-tuning to train,effectively captures the sensitivity of each feature in classification task,and weights the inputs to improve the classification performance.In order to verify the universal applicability of the model,the experiment regards the data which is subjected to wavelet time-frequency transform as input.The result verifies that the model has valid universality.