Deep learning-based receiver design for generalized frequency division multiplexing (GFDM)
Generalized frequency division multiplexing (GFDM) is a new efficient multi-carrier system in highly dispersive channels. In a conventional GFDM receiver, if the length of a GFDM block is the power of two (an even number), the equalizer can be implemented with low complexity using the fast Fourier t...
Saved in:
Published in | Physical communication Vol. 65; p. 102390 |
---|---|
Main Authors | , |
Format | Journal Article |
Language | English |
Published |
Elsevier B.V
01.08.2024
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | Generalized frequency division multiplexing (GFDM) is a new efficient multi-carrier system in highly dispersive channels. In a conventional GFDM receiver, if the length of a GFDM block is the power of two (an even number), the equalizer can be implemented with low complexity using the fast Fourier transform (FFT). On the downside, if the block length is even, the GFDM matrix will be singular, leading to severe performance loss. Thus, in the conventional GFDM system, high performance and low complexity cannot be achieved jointly. This paper proposes a new GFDM receiver based on deep learning (DL) methods that can achieve high performance regardless of the GFDM block length. In the proposed GFDM receiver, a deep neural network (DNN) is used to demodulate GFDM symbols. A particular case of a convolutional neural network (CNN) and a fully connected layer are employed in the DNN block. We simulate the proposed DL-based GFDM system in additive white Gaussian noise (AWGN) and multipath channels. Simulation results show that the proposed receiver achieves a significantly higher bit error rate (BER) performance than the conventional GFDM receiver when the length of the GFDM block is even in both AWGN and fading channels. Therefore, the proposed DL-based GFDM receiver can achieve low complexity and high performance simultaneously.
[Display omitted]
•Generalized frequency division multiplexing.•Demodulator design.•Deep learning.•Convolutional neural network.•Noise enhancement. |
---|---|
ISSN: | 1874-4907 1876-3219 |
DOI: | 10.1016/j.phycom.2024.102390 |