Research On Algorithms Of Feature Selection And Online Learning For Radar Emitter Recognition

Radar emitter recognition is a key step of the modern electronic countermeasure. With more drastic changes in electromagnetic environment and more complex signal forms, traditional recognition methods face increasing challenges. The high sampling rate supported by the electronic equipment leads to high dimensionality of the existing feature set, which brings great difficulties to the subsequent classifiers. Besides, radar emitter recognition needs to handle data quickly in current electronic warfare, in order to support real-time decision-making. Concerned with above two issues, feature selection and online learning for radar emitter recognition are studied in this paper.For the problem of small number of samples and high dimensionality of features in emitter datasets, two"filter-model"feature selection algorithms, Relief and Simba, are studied respectively. Experimental results show that such algorithms can reduce the dimensionality of the real emitter data effectively and stably, which might help to speed up the classification process and is beneficial for engineering application.To solve the problem that SVM-RFE algorithm can not remove the redundant features, a max-relevant and min-redundancy theory based support vector machine recursion features eliminate algorithm (MRMR-SVM-RFE) is proposed. It embeds the MRMR criterion into the feature ranking strategies of SVM-RFE and builds a more reasonable criterion to rank the features and to optimize the structures. Experimental results show that the proposed MRMR-SVM-RFE algorithm is superior to the original SVM-RFE in terms of recognition accuracy and robustness, and can be applied to the feature selection for real emitter data successfully.At last, for the real-time requirement of radar emitter recognition system, several online learning algorithms are studied in this paper. Experiments on real radar data show that the Perceptron based algorithm is the fastest but has poor stability; The OISVM algorithm can maintain a high and stable recognition rate while its training time is much less than traditional SVM algorithms, We suggest it as a potential candidate in practical application.