Deep Neural Network-Based Symbol Detection for Highly Dynamic Channels

Abstract

In extreme communication environments, a highly dynamic channel (HDC) often arises with quite challenging fast time-varying and nonstationary characteristics. Different from existing studies, in this work, we investigate the most intractable HDC case when the coherence time of the channel is smaller than the symbol period. We propose a deep neural network (DNN)-based symbol detector using the long short-term memory (LSTM) neural network. Particularly, the sampling sequence of the received signal per symbol is used as the input data of each LSTM unit, which can take advantage of all received information and thus achieve better performance. Furthermore, a preprocessing unit using the basis expansion model (BEM) is designed to dramatically reduce the number of parameters while training the neural network, and the BEM-DNN-based detector achieves almost the same performance as the DNNbased detector. Finally, simulation results are achieved using the highly dynamic plasma sheath channel (HDPSC) data measured from realistic shock tube experiments. The results show that the proposed DNN-based method outperforms conventional methods and requires no prior channel estimation or knowledge of channel models.