Deep learning based underwater acoustic OFDM communications

Abstract

In this paper, we present a deep learning based underwater acoustic (UWA) orthogonal frequency-division multiplexing (OFDM) communication system. Unlike the traditional receiver for UWA OFDM communication system that performs explicitly channel estimation and equalization for the detection of transmitted symbols, the deep learning based UWA OFDM communication receiver interpreted as a deep neural network (DNN) can recover the transmitted symbols directly after sufficient training. The estimation of transmitted symbols in the DNN based receiver is achieved in two stages: (1) training stage, when labeled data such as known transmitted data and signal received in the unknown channel are used to train the DNN, and (2) test stage, where the DNN receiver recovers transmitted symbols given the received signal. To demonstrate the performance of the deep learning based UWA OFDM communications, we generate a large number of labeled and unlabeled data by using an acoustic propagation model with a measured sound speed profile to train and test the DNN receiver. The performance of the deep learning based UWA OFDM communications is evaluated under various system parameters, such as the cyclic prefix length, number of pilot symbols, and others. Simulation results demonstrate that the deep leaning based receiver offers consistent improvement in performance compared to the traditional UWA OFDM receiver.