Research On Radar Emitter Signal Recognition Based On High Order Statistics

Radar emitter signal recognition technology is an important symbol of the performance ofradar countermeasure equipments. It impacts decision-making in battlefield such as attackingagainst the enemy and self-defense directly. However, the intense competition of the modernelectronic warfare makes the performance of traditional method which relies on conventionalfive parameters drop significantly, even can not use at all. For this issue, the main work of thispaper is as follows:1. The background and significance of the research of radar emitter signal recognition iselaborated. Then, this paper summarizes the current research situation of all aspects of radaremitter signal recognition.2. Some important types of low probability of intercept (LPI) radar signals is introduced,and analyzes these signals’ characteristics in time domain, frequency domain and time-frequencydomain. Then, the classifier which is used in this article-Support Vector Machine is introduced.3. Bispectrum has some merit such as strong anti-noise performance, but it has a largeamount of data. There were several methods to reduce the amount of data of bispectrum currently,on the basis of analyzing the advantages and disadvantages of these methods, a new method offeature extraction which is called diagonally integral bispectrum is proposed in this article. Inthis method, bispectrum is integrated along the lines which parallel to the diagonal line. It notonly avoids interpolation, optimization and coordinate axis transformation, but also takes boththe symmetry of the bispectrum and features into accounts, so utilizes more information ofbispectrum. At last, the diagonally integral bispectrum is used to extract the feature of the LPIradar emitter signal, theoretical analysis and experimental results show that the diagonallyintegral bispectrum outperforms existing methods.4. Cyclic bispectrum is a powerful tool to analyze the non-stationary of radar emitter signal,and contains a wealth of information. However, it has large amount of data, and the traditionaldimension reduction method lost most of the information. For this problem, the symmetry andperiodicity of cyclic bispectrum is proved, and on this basis five methods of feature extractionwhich are called radially integral cyclic bispectrum, axially integral cyclic bispectrum,surrounding-line integral cyclic bispectrum, sub-block integral cyclic bispectrum and diagonallyintegral cyclic bispectrum respectively are proposed in succession. These methods not only caneffectively reduce the amount of data, and have used all of the information of cyclic bispectrum.Under simulation conditions, the performance of the five new integral cyclic bispectrum iscompared with the traditional method when bimodal noise is added. The result shows that theperformance of new methods is far superior to the traditional one.5. There is a large amount of calculation, data redundancy and decline of recognition ratewhen the feature dimension is too high. For those problems, a new feature selection algorithmbased on weighted-J2criterion and similar coefficient is proposed. Firstly, the drawback of thetraditional J2criterion is analyzed and modified with entropy. Secondly, the resemblancecoefficient is used to characterize the separability of two types of sample, thereby establishing redundant vectors. The algorithm chooses and puts L features into the selected feature subsetaccording to the weighted-J2criterion in advance. And then deletes R features from selectedfeature subset which is more redundant, the cycle continues until the end condition is satisfied.Simulation results show that the algorithm is superior to mRMR and NMIFS algorithm.