Thumbnail Image

Learning-Based Automatic Modulation Classification

Abdelmutalab, Ameen Elsiddig
A Master of Science thesis in Electrical Engineering by Ameen Elsiddig Abdelmutalab entitled, "Learning-Based Automatic Modulation Classification," submitted in May 2015. Thesis advisors are Dr. Khaled Assaleh and Dr. Mohamed El-Tarhuni. Soft and hard copy available.
Automatic Modulation Classification (AMC) is a new technology implemented into communication receivers to automatically determine the modulation type of a received signal. One of the main applications of AMC is in adaptive modulation systems, where the modulation scheme is changed dynamically according to the changes in the wireless channel. However, this requires the receiver to be continuously informed about the modulation type, resulting in a loss of bandwidth efficiency. The existence of smart receivers that can automatically recognize the modulation type improves the utilization of available bandwidth. In this thesis, a new AMC algorithm based on a Hierarchical Polynomial Classifier structure is introduced. The proposed system is tested for classifying BPSK, QPSK, 8-PSK, 16-QAM, 64-QAM and 256-QAM modulation types in Additive White Gaussian Noise (AWGN) and flat fading environments. Moreover, the system uses High Order Cumulants (HOCs) of the received signal as discriminant features to distinguish between the different digital modulation types. The proposed system divides the overall modulation classification problem into hierarchical binary sub-classification tasks. In each binary sub-classification, the HOC inputs are expanded into a higher dimensional space in which the two classes are linearly separable. Furthermore, the signal-to-noise ratio of the received signal is estimated and fed to the proposed classifier to improve the classification accuracy. Another modification is added to the proposed system by using stepwise regression optimization for feature selection. Hence, the input features to the classifier are chosen to give the highest classification accuracy while maintaining a minimum number of possible features. Extensive simulations showed that a significant improvement in classification accuracy and reduction in the system complexity is obtained compared to the previously suggested systems in the literature.
External URI