Binary Block-Based Neural Network Decoder for F-GSM Signal Detection

Abstract

Fully-generalised spatial modulation (F-GSM) was presented as an energy and spectral efficiency modulation scheme. In F-GSM, almost all possible transmitted antenna indices combinations are utilized in the data transmission process, which improves its spectral and energy efficiency on one side. On the other side, this adds more challenges to the F-GSM decoder due to using maximum Likelihood Estimator (MLE). Although MLE achieves optimum performance, its computational complexity (CC) increases exponentially with the F-GSM achievable rate. Therefore, this paper proposes a simple binary block-based neural networks (BBNN) F-GSM decoder with lower CC. In particular, a simple binary classifier neural network architecture is proposed to detect the activation status independently of each antenna or spatial bit, which reduces the offline training time and required training data size. Moreover, it makes the F-GSM decoder more reliable and improves adaptation capability when the number of antennas is changed. After that, a low complexity Euclidean distance-based estimator is used to detect the signal constellation. Simulation results show that the bit error rate (BER) performance of the proposed BBNN decoder outperforms the performance of the conventional block zero-forcing (BZF) and block minimum mean squared error (BMMSE) systems. In contrast, it archives a comparable ABER to that of traditional MLE but with lower CC, as it achieves on average ~81.5 % CC reduction compared to MLE.