Deep Learning-Based Automatic Modulation Classification With Blind OFDM Parameter Estimation

Abstract

Automatic modulation classification (AMC) is an essential factor in dynamic spectrum access to fulfill the spectrum demand of 5G wireless communications for achieving high data rate and low latency. Many deep learning (DL)-based AMC methods have achieved improved accuracy for classifying analog modulation schemes, single-carrier-based modulation schemes, and multi-carrier signals using several DL architectures such as the convolutional neural network (CNN) and long-short term memory (LSTM). However, most conventional DL-based AMC methods have confused the orthogonal frequency multiplexing division (OFDM)-based signals with different OFDM useful symbol lengths. To resolve the issue, we propose a CNN model operating on the fast Fourier transformation window bank (FWB) to extract the useful symbol length in OFDM, which represents the identification of each OFDM-based wireless communication technology. After extracting the OFDM useful symbol length, we propose a DL-based AMC system combined with FWB and in-phase and quadrature-phase signals to classify the OFDM symbol length and single-carrier modulation schemes simultaneously. Furthermore, we explore the constraints of the FWB parameters according to the length and the fast Fourier transformation (FFT) size of the OFDM signal to achieve good classification accuracy through the experiment. We constructed a dataset by generating OFDM signals of different lengths while changing the FFT size in a fixed bandwidth and selecting only quadrature amplitude modulation (QAM) schemes from RadioML2016.10a. Experimental results show the improved classification accuracy by about 30% over conventional classifiers in additive white Gaussian noise, synchronization impairments, and fading environments.