Deep learning aided OFDM receiver for underwater acoustic communications

Abstract

In this study, we propose a deep learning (DL)-based orthogonal frequency division multiplexing (OFDM) receiver for underwater acoustic (UWA) communications. Compared to existing deep neural network (DNN) OFDM receivers composed of fully connected (FC) layers, our model tailors complex UWA communications with precision. To this end, it utilizes a convolutional neural network with skip connections to perform signal recovery. The stacks of convolutional layers with skip connections can effectively extract promising features from received signals and reconstruct the original transmitted symbols. Then, a multilayer perceptron is used for demodulation. To demonstrate the performance of the proposed DL-based UWA-OFDM communication system, the training and testing sets are generated using the strength of the measured-at-sea WATERMARK dataset. The experimental results show that the proposed model with skip connections can outperform the existing approaches (i.e., traditional UWA-OFDM with least squares channel estimation, and FC-DNN-based framework) in terms of both accuracy and efficiency. This is prominent in harsh UWA environments with strong multipath spread and rapid time-varying characteristics.