Testbed development of a Deep Learning-based OFDM scheme for Undersea RF Communication using Software Defined Radio Platform

This paper presents the testbed development of a deep-learning based OFDM scheme for short-range high performance undersea RF communication. Realization of an RF undersea communication system faces challenges such as low signal-to-noise (SNR) ratio, insufficient modelling of undersea wave propagation and characteristics of noises of underwater operating devices. This research investigates the concept of utilizing deep learning-based OFDM scheme, where a deep learning network is implemented at the OFDM receiver, instead of conventional channel estimation and equalization, and OFDM symbol detection and demodulation. The deep learning network is trained by feeding the network with receiving OFDM bit streams which inherit the effects of undersea / underwater channel characteristics, including devices and system noise effects, while varying the SNR, channel conditions and typical OFDM parameters such as cyclic prefix and pilot symbols. Deploying a fully trained network at the OFDM receiver enables demodulation without channel estimation and equalization. In this paper, we showcase the development of the proposed scheme using software define radio (SDR) platform, tested by transmission of data through a pair of in-house developed underwater antennas, in a laboratory water tank filled with seawater. The deep training data for the deep learning network is obtained through a series of experimental transmissions inside the laboratory water tank, and additional simulated data passed through simulated underwater RF channel. It is shown through initial experiments that the proposed scheme outperforms conventional channel estimation using least square (LS) method, in approximately 6 dB of SNR.