Benefits of Neural Network-based Receivers in Underwater Acoustic Communication Systems

In conventional orthogonal frequency-division multiplexing (OFDM) communication systems, channel knowledge on data subcarriers is obtained through linear interpolation of the estimated channel information on pilot subcarriers. This operation degrades the accuracy of channel estimation, particularly when the multipath channel has a large delay spread as in underwater acoustic (UA) channels. In this paper, we show that such drawback can be overcome by a neural network (NN)-based receiver, which is effective in learning the nonlinearity in the channel information on subcarriers. A multilayer perceptron (MLP) network, a convolutional neural network (CNN), and a long short-term memory (LSTM) network architectures are attempted. These NNs are trained and tested by data collected in a real-world UA communication experiment conducted in a water tank. The results show that the MLP and LSTM network-based receivers achieve better bit-error-rate (BER) performance than the conventional OFDM receiver.