Intelligent Recognition Methods For Radar Emitter Signals

Radar emitter signal recognition is one of the key procedure of signal processing in ELectronic INTelligence (ELINT), Electronic Support Measures (ESM) and Radar Warning Receiver (RWR) systems in electronic warfare. It is also the precondition and foundation of electronic interfering. The state of the art of radar emitter signal recognition corresponds to the technical merit of electronic reconnaissance equipment. As countermeasure activities in modern electronic warfare become more and more drastic, advanced radars increase rapidly and become the main component of radars gradually. Complex and changeful signal waveform weakens greatly the validity of traditional recognition methods and makes the validity lose gradually. Radar emitter signal recognition has been confronted with strange challenges. At present, there is a big gap between the state of the art of recognizing advanced radar emitter signals in our country and that in some advanced countries such as the United States of America. The prime reason lies in low level of theoretic research about radar emitter signal recognition in our country and lack of theoretic basis for improving counter-radar equipments. Southwest Institute of Electronic Equipment (SWIEE) is a key base of scientific research in electronic warfare in our country. In the process of studying modern electronic equipment, the experts of SWIEE profoundly point out that the currently technical merit of counter-radar equipments cannot adapt the complex environment of modern electronic warfare in which advanced radars will have a decisive influence, and the existing methods for recognizing radar emitter signals lag behind rapid-development radar technique. Therefore, the recognition of advanced radar emitter signals becomes a key problem to solve urgently in electronic warfare. Only some new and valid approaches are explored, can the technical merit of ELINT, ESM and RWR systems be enhanced radically.In recent years, although radar emitter signal recognition is paid much attention and some recognition methods were presented, using conventional 5 parameters, traditional recognition methods and their improved methodsencounter serious difficulties in identifying advanced radar emitter signals. Unfortunately, the existing intrapulse characteristic extraction approaches only analyze qualitatively two or three radar emitter signals without considering the effects of noise nearly. So the approaches cannot meet the intelligentized requirements of modern information warfare for electronic warfare reconnaissance systems. For the difficult problem of recognizing complicatedly and changefully advanced radar emitter signals, multiple methods need be used to extract the valid features from radar emitter signals from multiple different views firstly. The features construct a high-dimensional feature vector. Then, some feature selection methods are proposed to select the most discriminatory features and to eliminate redundant features so as to lower the dimensionality of the feature vector. Finally, efficient classifiers are designed to fulfill automatic recognition of radar emitter signals. Through theoretic research systematically, the theory can be established to improving the techniques for recognizing radar emitter signals effectively.Aiming at the key issue to solve urgently in signal processing of electronic warfare in our country, intelligent recognition model and algorithms for advanced radar emitter signals are studied systematically and exploringly in this dissertation. Theoretical fruits are as follows.1 A new model structure for recognizing radar emitter signals is proposed and consequently a fire-new thinking is introduced to solve the difficult problem of advanced radar emitter signal recognition. The existing matching technique using 5 conventional parameters identifies advanced radar emitter signals difficultly. According to the characteristics of advanced radar emitter signals and the drawbacks of the existing methods, this dissertation uses a new way, including feature analysis and extraction, feature evaluation and selection, classifier design, to solve the difficult issue of radar emitter signal recognition. The algorithms for implementing the proposed model structure and testing experiments validate that the new model structure is more effective than the current model structure.2 When Signal-to-Noise Rate (SNR) varies in a certain range, Severalfeatures of radar emitter signals are analyzed and studied quantificationally. From different views, different feature extraction methods are used to explore the intrinsical characteristics. Several methods including resemblance coefficient feature extraction, entropy feature extraction, complexity feature extraction and wavelet packet transform feature extraction are proposed to extract features from radar emitter signals. Noise-suppression performances of resemblance coefficient features, entropy features and complexity features are analyzed qualitatively or quantificationally. Experimental results show that these features are not nearly affected by noise and have good stability when SNR is above 5 dB or 6 dB. Using the several feature extraction methods, two resemblance coefficient features, two entropy features (approximate entropy and norm entropy), 8 wavelet packet decomposition features and 4 complexity features (Lempel-Ziv complexity, information dimension, box dimension and correlation dimension) are extracted from multiple radar emitter signals. Furthermore, the time complexities of the feature extraction methods, the within-class clustering and between-class separability of these features are used to evaluate comparatively the performances of the methods and the qualities of the features.3 After the features are extracted adequately from radar emitter signals, models and algorithms for exploring intrinsical features of radar emitter signals are studied further. Five feature selection methods, including class-separability, satisfactory rate of feature set, resemblance coefficient, rough set theory and discriminability coefficient, are proposed to analyze and evaluate the extracted features from different aspects and to eliminate redundant features so as to lower the dimensionality of feature space. Also, a large number of comparative experiments are made and it is proved that the several proposed methods are superior to several feature selection methods in the existing literatures greatly.Rough set theory can only deal with discreted attributes. The existing discretization definition based on cut-splitting is extended to generalized discretization definition based on class-separability. A novel method is presented to discretize interval-valued continuous attributes that cannot be processed effectively by the existing discretization methods. Also, time complexity of themethod is analyzed qualitatively.In the process of feature selection using class-separability, satisfactory rate of feature set and resemblance coefficient, a novel quantum genetic algorithm is proposed to search the best feature subset from original feature set. Comparing with traditional genetic algorithms, the novel quantum genetic algorithm has stronger searching capability, faster convergence speed, shorter computing time and better capability to avoid too-early convergence.4 Four classifier design methods, including Rough Set Neural Network (RSNN), Combination Support Vector Machine (CSVM), Huffman Tree Support Vector Machine (HTSVM) and Rough Set Support Vector Machine (RS-SVM), are proposed to recognize radar emitter signals automatically. The validities and superiorities of the 4 classifiers are shown in the comparative experiments.A large number of comparative experiments are made and several conclusions can be drawn: RSNN is superior to Radial Basis Function Neural Network (RBFNN), Probability Neural Network (PNN) and Back-Propagation Neural Network (BPNN) in classification capability and recognition efficiency; three popular multiclass classification Support Vector Machines (SVMs) including One-Against-All (OAA), One-Against-One (OAO) and Binary Tree Architecture (BTA), have better classification capabilities than RBFNN, PNN and BPNN; CSVM is superior to OAA, OAO and BTA greatly in classification capability and efficiency; HTSVM has faster recognition speed than OAA, OAO and BTA instead of lowering its classification performance; the introduction of rough set theory not only strengthens greatly classification capability of SVMs, but also enhances the generalization capability of SVMs.Unknown radar emitter signal recognition is also an important issue. This dissertation also discusses the applications of Competitive Learning Neural Network (CLNN) and Self-Organizing feature Map Neural Network (SOMNN) to recognizing radar emitter signals.5 One hundred and fifty-five radar emitter signals with different signal parameters are chosen to make recapitulative experiments to test the validity of the proposed model structure. Each of the 155 signals has one of 8 intrapulsemodulations. Experimental results show that the proposed model structure can recognize the advanced radar emitter signals more effectively than the current model structure, which proves that the introduced model structure is more practical than the current model structure. Also, CSVM and the feature subsets obtained by using respectively 4 feature selection methods including class-separability, satisfactory rate of feature set, resemblance coefficient and rough set theory are employed to make comparative experiments to identify the advanced radar emitter signals. About 88%, 85%, 93% and 91% correct recognition rates are achieved respectively, which testifies again that the validities and practicalities of the feature selection methods and CSVM. Moreover, some experiments are also made using CLNN and SOMNN respectively. Unsatisfying experimental results indicate that the performances of CLNN and SOMNN need improve further in future research.This work was supported partially by Electronic Warfare Technology Pre-research Foundation (NEWL51435QT220401), by the National EW Laboratory Foundation (No.51435030101ZS0502), and also supported by Doctorial Innovation Foundation (2003.12).