Split-complex feedforward neural network for GFDM joint channel equalization and signal detection

Abstract

This paper presents a novel approach for channel equalization and signal detection in generalized frequency division multiplexing (GFDM) systems, designed for dispersive channels and capable of handling nonlinearities. In digital communications systems, deep learning (DL) techniques have emerged as a promising alternative to traditional adaptive digital signal processing. Although DL is a trending topic and has been applied to areas such as beamforming, channel estimation, equalization, and decoding, there is limited research on the use of complex-valued neural networks (CVNN), particularly in the context of GFDM systems. In this work, we propose a joint channel equalization and signal detection approach for GFDM based on the fully connected CVNN split-complex feedforward neural network (SCFNN). The proposed SCFNN effectively equalizes the dispersive 5G channel while concurrently detects the symbols non-orthogonally multiplexed in frequency, handling both scenarios with and without clipping, all within a single SCFNN. Results are compared with classical equalization and detection algorithms, as well as with the fully connected real-valued neural network (RVNN) approach. The proposed SCFNN solution presents superior symbol error rate (SER) performance while maintaining computational complexity on par with conventional methods.