Deep learning-based receiver structure with superimposed channel estimation for autonomous underwater vehicles

Abstract

Although autonomous underwater vehicles can achieve complex tasks in harsh and inaccessible marine environments, they face several challenges related to the harsh channel effects and limited available bandwidth, making reliable channel estimation crucial task for achieving robust communication. Therefore, to track channel effects, a significant portion of bandwidth is usually reserved for overhead, which dramatically reduces the already limited bandwidth efficiency. In this paper, we present a real signal orthogonal frequency division multiplexing (OFDM) method based on deep learning to accurately track channel effects without losing bandwidth efficiency. The proposed scheme adopts the discrete Hartley transform to produce a real signal modulation, and a unitary neural network (UNN) at the sending end to add extra packages. At the receiving end, we use a deep neural network (DNN) in conjunction with channel estimation and equalization to accurately detect the transmitted information data. Therefore, we jointly train both UNN and DNN to prevent interference between the data and pilot, as well as to address factors that impact data detection performance effectively. We also deploy the carrier frequency offset estimation technique without sacrificing any subcarriers. Consequently, we track channel effects without the need for dedicated pilot subcarriers and/or data detection degradation. The proposed method has a better bit error rate, spectral efficiency, and computational complexity than current benchmarks, as shown by both the simulation and the real experiments done in the sea over a distance of 300 m