Radar Target HRRP Recognition Based On Rough Set Theory

An effective HRRP recognition method is expected to have the ability to mine hidden, novel and potentially valuable information from such a dataset as is enormous and at the same time contaminated by noise. As a matter of fact, such ability can be provided exactly by the emerging data mining technology. As a kind of data mining method which is employed extensively at present and extraordinarily promising in the future, rough set theory (RST), and also its applications in diverse areas, is arousing more and more attention and interest in the academe. Consequently, it has become a new direction of HRRP recognition research to use RST as a tool to discover the class information hidden in HRRP to support the recognition process. In this background, we investigate systematically the RST-based HRRP recognition method in this dissertation. To sum up, several main aspects are included in the dissertation and are listed as follows.In chapter one, the great significance of radar automated target recognition (RATR) is first introduced. And then, the unique advantages of HRRP-based RATR are analyzed in detail. Subsequently, the existing recognition methods are reviewed systematically.Chapter two introduces the basic concepts, principles and methods of the standard RST in a systematic way. Specifically, the central function, namely knowledge reduction, of the RST is introduced, and accordingly the two basic concepts to fulfill this function are given, which are called core and reduct, respectively. In addition, according to the application requirements in HRRP recognition, the existing reduct calculation algorithms are summarized systematically, and meanwhile the aspects to be tackled in the process of finding the HRRP reduct are analyzed accordingly. In chapter three, a series of problems encountered in the standard RST-based HRRP recognition are investigated in sequence, including data preprocessing method, reduct calculation algorithm, rule induction algorithm and rule fusion method. Firstly, the necessity of data partition is analyzed, and a partition strategy is given; the existing continuous attributes discretization methods are summarized systematically, and a method applicable to HRRP discretization is introduced. Secondly, a reduct calculation algorithm which is able to find several HRRP reducts is proposed. By employing this algorithm and the data partition scheme together, the NP-hard problem in finding all reducts can be successfully avoided. Thirdly, the decision tree classifier (C4.5 classifier) and the majority voting method are adopted to fulfill rule induction and rule fusion, respectively. Based on the above steps, we finally construct a standard RST-based HRRP recognition system. It is demonstrated by computer simulations that this system exhibits good classification rate and quite powerful noise suppression quality.In order to avoid the disadvantages of the standard RST-based system caused by the standard RST's incompetence in directly dealing with continuous attributes, chapter four focuses the attention of the dissertation on how to improve the performance of the standard system by using the improved rough set theory. Specifically, the continuous rough set theory is applied to the HRRP recognition area in this chapter. The attribute reduction method and rule induction algorithm are intensively investigated. Firstly, aεprecision reduct calculation algorithm is proposed. Theεprecision reduct can maintain the inclusion degree-based classification rules of the continuous information system after the attribute reduction. The algorithm is able to find severalεprecision reducts. Secondly, a new reduct definition based on the continuous RST is proposed by extending the reduct definition in the standard RST. An algorithm which can find several such reducts is also proposed. According to the different characteristics of the two kinds of reduct, the rule induction methods of inclusion degree-based rule induction and C4.5 classifier-based rule induction are adopted respectively, only to produce two kinds of continuous RST-based recognition systems. It is demonstrated by the simulation results that each system outperforms the standard RST-based system both in robustness and recognition time cost.Finally, the dissertation is concluded in chapter five. Several aspects for future work are also pointed out.