Linear and nonlinear classification of quadrature modulation signals

Automatic modulation classification (AMC) is a form of modern technology integrated into communication receivers to automatically define the modulation type of a received signal. In this work, two new AMC algorithms based on signal feature extraction and a pattern-recognition method are introduced. First, AMC methods are summarised. Subsequently, up-to-date AMC algorithms, including their development, extracted features, their ability to recognise modulated signals, operating configurations, and some recent AMC projects are discussed from their preliminary historical development. Furthermore, state-of-the-art AMC methods are classified according to their feature-selection group, classifier structure, number of modulation signals, channel system, and simulation tool. Second, a new AMC method based on a hierarchical threshold classifier structure is presented. The proposed system comprises multiple tree-based on logarithmic thresholds. The system uses a logarithmic modification of the high-order cumulants (HOCs) of the modulated signal as discriminant features to distinguish among the different digital-modulation schemes. Moreover, the proposed system is tested for classifying QPSK and MQAM modulation types in additive white Gaussian noise (AWGN) and flat-fading environments. The simulation results demonstrate that a very good classification rate is achieved at a low SNR of 5 dB, under conditions of statistically noisy channel models. This shows the potential of the logarithmic-classifier model for application in M-QAM signal classification. Further, in this thesis, we propose the leveraging of novel higher-order spectra features (HOSF) in classification algorithms based on neural-network properties, to mitigate the modulation recognition problems specified in the M-APSK DVB-S2X modulation signals standard. The HOSF characteristics of signals are extracted under the AWGN– channel, and four individual parameters are defined for distinguishing modulation– signals from the set, {16, 32, 64}-APSK. This approach makes the recogniser more intelligent and improves its classification success rate. The results illustrate the excellent classification accuracy obtained at a low SNR of 0 dB, which demonstrates the potential of combining these proposed features with a neural-network classifier for M-APSK modulation classification. Finally, in both proposed algorithms, extensive simulations revealed that a significant improvement in classification accuracy and reduction in system complexity is achieved compared to the previously proposed systems in the literature.