Low-complexity deep learning and RBFN architectures for modulation classification of space-time block-code (STBC)-MIMO systems

Abstract

Due to the constraint of spectrum availability in the recent era, technologies such as cognitive radio and spectrum sensing have found a central stage in the communications arena. Automatic modulation classification (AMC) plays a pivotal role in most of these technologies. AMC for Multiple-Input Multiple-Output (MIMO) systems is a domain that still lacks the due attention. Keeping in view the recent advancements in the field of artificial intelligence, we propose two relatively simple neural network architectures, i.e., Sparse Autoencoders(SAE) based Deep Neural Network (DNN) and Radial Basis Function Network (RBFN) for AMC of Space-Time-Block-Codes(STBC)-MIMO systems. An assorted set of features is proposed on the basis of distribution analysis to train these classifiers. Limited-memory Broyden–Fletcher–Goldfarb–Shanno(L-BFGS) algorithm and the least square method is used for weight optimization of DNN and RBFN, respectively. The impact of channel estimation error on the classification outcome is thoroughly investigated for blind classification scenarios. Performance comparison of the proposed methods with one of the most recent works in this context validates the effectiveness of proposed features and classifiers. Moreover, comparison with optimal Maximum-Likelihood(ML) based classifier and state of the art machine learning algorithms including, Adaptive Boosting (AdaBoost) and Convolutional Neural Networks (CNN) is also carried out. Simulation results substantiate the viability of proposed architectures for three different configurations of the STBC-MIMO system for perfect channel state information (CSI) and blind scenarios amidst channel estimation error.