Long Short-Term Memory Network for Co-Frequency Co-Time Full-Duplex Digital Domain Interference Suppression

Abstract

The co-frequency co-time full duplex (CCFD) system transmits and receives signals at the same frequency simultaneously, which can effectively improve the communication spectrum utilization of satellite and solve the problem of satellite frequency resource constraint. However, due to a large number of devices in a relatively small space, there is strong self-interference between transmitting and receiving antennas. Besides, the received signal intensity is not stable due to other factors such as multipath effect, electromagnetic interference, and so on. To solve this problem, this paper proposes a machine learning method for self-interference suppression of CCFD systems in digital domain. A signal sequence decomposition method based on the number of poles is used to solve the problem of insufficient suppression of nonlinear interference signal caused by signal strength instability. By using Hilbert transform, the input signal dimension of machine learning is reduced, which is to solve the high computational complexity. The short-term and long-term neural network architecture is adopted to predict and reconstruct the interference, and Bayesian optimization method is used to optimize the hyperparameters of the network, and introduces the acquisition function into the loss function, to improve the comprehensive training optimization of machine learning networks. The experimental results show that the proposed algorithm can effectively analyze and suppress the interference signal, and the interference suppression capability is improved by 10.7 dB compared with the traditional Long Short-Term Memory algorithm and 21.1 dB compared with the Least Mean Square algorithm. The algorithm presented in this paper is significant for the application of CCFD system on satellite.