Statistical Methods in Modulation Classification
The interest in modulation classification has recently emerged in the research of communication systems. This has been due to the advances in reconfigurable signal processing systems, especially in the study of software radio. Published methods can be divided into two groups: maximum likelihood and pattern recognition approaches. In the maximum likelihood approach, decision rules are often simple but the test statistics are complicated and assume prior knowledge about the signals. In the pattern recognition approach, decision rules are often complicated whereas the features are simple and fast to calculate. Communication signals contain a vast amount of uncertainty due to the unknown modulating signal, modulation type, and noise. Therefore the modulation classification problem has to be approached by using statistical methods. The features and the test statistics may be derived from the known statistical characteristics of the modulated signals. Either implicit or explicit use of higher-order statistics has been studied previously in many communication applications. The higher-order statistics are often more preferable because second-order statistics suppress the phase information of the signal. Nevertheless, the estimation of the higher-order statistics requires long sample sets and has a high computational complexity. To overcome these problems, second-order cyclostationary statistics have been studied and the results seem promising. In this thesis, the feature extraction problem of the modulation classification is discussed. Useful characteristics and representations of the communication signals are presented as well as the relevant knowledge of statistical signal processing. The previous methods are presented in a literature review of the modulation classification. The first and second-order statistics including the cyclostationary statistics of digital modulated signals are studied, and a novel feature is proposed. Some previous methods and this novel feature are compared by investigating their discrimination performance in Matlab simulations.