Deep Learning-Based Modulation Classification for OFDM Systems Without Symbol-Level Synchronization

Abstract

Deep learning (DL)-based modulation classification of incoherently received orthogonal frequency division multiplexing (OFDM) signals is studied. We propose a novel preprocessing algorithm to build features characterizing the modulation of OFDM signals, which are insensitive to synchronization error. With obtained features, pilot subcarrier indices used for CFO correction may also be estimated. The features obtained with the proposed algorithm are classified with a convolutional neural network (CNN)-based classifier. We have evaluated classification performance with simulated and hardware-generated data. Using these features, the modulation classifier outperforms existing DL-based classifiers which assume symbol-level synchronization with up to 25% classification accuracy performance gain.