Multiclass Classification of Modulation Formats in the presence of Rayleigh and Rician Channel Noise using Deep Learning Methods

Wireless communication technologies have revolutionized the communication scenario completely. In the last four decades, the transition from 1G to 5G communication systems has been instrumental in opening up many new possibilities. An important step for wide-range applications in intelligent Communication network is automated modular format identification. It is a significant step for an exact detecting of the signal on the receiver side. Deep learning methods are used extensively in the fields of speech recognition, image recognition, object detection and reinforcement learning etc. In this research, 2D and 3D Convolution neural networks (CNNs) were compared and contrasted for the purpose of automatic recognition of modulation formats. We have used random cat images that are modulated with 16 and 64 Quadrature Amplitude Modulation (QAM) signals and are classified using 2D and 3D CNN architecture after passing them through Rayleigh and Rician noise channels. We tried 5 and 10 fold cross-validation procedures to train multiclass (4-classes) classifiers. We observed that, in terms of reported metrics, the performance of the 3D CNN architecture trained using 10-fold cross validation procedure was the best and that of 2D CNN architecture trained using 5-fold cross validation procedure was the worst. Overall, we have found the performances of 3D architectures to be better than their 2D counterparts.