The Genetic Algorithm For Feature Selection Of Radio Abnormal Signal

Feature selection, one of the fundamental and challenging problems in radio abnormalsignal recognition, has attracted more and more attention in signal processing. Featureselection is the process that selects features according to their discrimination power. Byfeature selection, we can decrease the size of the data set and centralize information which hassignificant difference among classes, thus improve the performance of the overall recognitionsystem.The existing feature selection feature selection approaches can be grouped into twocategories: filter methods and wrapper methods. Acquiring no feedback from the classifier,the filter methods estimate the classification performance by some indirect assessments, it istime efficient but acquires lower precision. The wrapper methods, on the contrary, evaluatethe “goodness” of the selected feature subset directly based on the classification accuracy. Inspite of the good performance, the wrapper methods have limited applications due to the highcomputational complexity involved. Because of its strong stability and low requirement forthe analytic properties of the search space, genetic algorithm (GA), an random search strategywhich is created by imitating the mechanism of creature evolution, break the limitation of thetraditional complete search strategy and heuristic search strategy, find a new way for solvingcomplex optimization problems. Genetic algorithm is often used as an optimization tool infeature selection methods, GA-based feature selection model have been widely applied invarious areas.In this paper, we combine the advantages of filter methods and wrapper methods in theframework design of feature selection. And according to the framework, a feature selectionmethod based on genetic algorithm is proposed, in which, a class separability measurementF-Ratio is used as the fitness function of GA. By applying this method in the wholerecognition system, we get forward to find an equilibrium point of the accuracy and efficiencyof recognition. The main contents are as follows:Firstly, we use a class separability criterion F-Ratio which is defined by the ratio ofdistances to measure the classification validity of feature. Based on the analysis of theF-Ratio’s structure, a logarithmic or exponential function is added into F-Ratio’s numerator(inter-class distance) or denominator (intra-class distance) to adjust their effect to the wholeF-Ratio value, and thus a series of F-Ratio formula are obtained.Secondly, in the framework design of feature selection, the calculation formula ofF-Ratio is selected based on the information returned from the classifier at first, then theselected F-Ratio is applied as the fitness to evaluate the classification validity of the features.Compare to the framework, in which evaluation of feature subset is directly based on the classification accuracy or based on the F-Ratio criterion, this semi-detached framework canbetter balance the requirements of efficiency and accuracy.Thirdly, to avoid premature convergence of the genetic algorithm, we introduce astrategy consisting of competition and multi-population parallel to the standard GA and useadaptive genetic operators in GA. In compare with the standard GA, the improved GA has abetter performance in global search.Finally, we combined the F-Ratio adjusted adaptively with the improved geneticalgorithm to select the optimal feature subset for signal classification. The experimentalresults based on real-time data indicate that the proposed method improves the recognitionperformance in both time and accuracy.