Electrical & Computer Engineering, Department of


First Advisor

Prof. Hamid R. Sharif-Kashani

Date of this Version



@article{ghsaemzadeh, title={Deep Learning and Polar Transformation to Achieve a Novel Adaptive Automatic Modulation Classification Framework}, author={Ghasemzadeh, Pejman}, year={2020} }


A THESIS Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfillment of Requirements For the Degree of Master of Science, Major: Telecommunications Engineering, Under the Supervision of Professor Hamid R. Sharif-Kashani. Lincoln, Nebraska: May, 2020

Copyright 2020, Pejman Ghasemzadeh


Automatic modulation classification (AMC) is an approach that can be leveraged to identify an observed signal's most likely employed modulation scheme without any a priori knowledge of the intercepted signal. Of the three primary approaches proposed in literature, which are likelihood-based, distribution test-based, and feature-based (FB), the latter is considered to be the most promising approach for real-world implementations due to its favorable computational complexity and classification accuracy. FB AMC is comprised of two stages: feature extraction and labeling. In this thesis, we enhance the FB approach in both stages. In the feature extraction stage, we propose a new architecture in which it first removes the bias issue for the estimator of fourth-order cumulants, then extracts polar-transformed information of the received IQ waveform's samples, and finally forms a unique dataset to be used in the labeling stage. The labeling stage utilizes a deep learning architecture. Furthermore, we propose a new approach to increasing the classification accuracy in low signal-to-noise ratio conditions by employing a deep belief network platform in addition to the spiking neural network platform to overcome computational complexity concerns associated with deep learning architecture. In the process of evaluating the contributions, we first study each individual FB AMC classifier to derive the respective upper and lower performance bounds. We then propose an adaptive framework that is built upon and developed around these findings. This framework aims to efficiently classify the received signal's modulation scheme by intelligently switching between these different FB classifiers to achieve an optimal balance between classification accuracy and computational complexity for any observed channel conditions derived from the main receiver's equalizer. This framework also provides flexibility in deploying FB AMC classifiers in various environments. We conduct a performance analysis using this framework in which we employ the standard RadioML dataset to achieve a realistic evaluation. Numerical results indicate a notably higher classification accuracy by 16.02% on average when the deep belief network is employed, whereas the spiking neural network requires significantly less computational complexity by 34.31% to label the modulation scheme compared to the other platforms. Moreover, the analysis of employing framework exhibits higher efficiency versus employing an individual FB AMC classifier.

Advisor: Hamid R. Sharif-Kashani