Development of CNN-LSTM Hybrid Deep Learning Network for the Joint Detection of Non-Orthogonal Multiple Access Signals in 5G Uplink Receivers

Abstract

This paper presents a novel approach for signal detection in non-orthogonal multiple access (NOMA) uplink receivers. We propose converting incoming packets into a stream of 2D image-like vectors. Thereby, converting a signal detection problem into a video classification problem. Our approach is true end-to-end learning where no manual feature engineering and/or pre-processing is required. Detection is done blindly in a joint fashion with no explicit channel estimation and equalization steps required. Successive interference cancellation (SIC) is the default method for detecting NOMA packets, but it requires perfect channel estimation to be maintained. Deep learning approaches have shown great promise. However, they suffer from overfitting and/or poor performance. We show how to improve performance by a better training dataset generation procedure and hyperparameter optimization along with the use of CNN as a feature filter. The CNN-LSTM hybrid network has registered a training and testing accuracies of 88.61 and 85.36 which are higher than the state-of-the-art approaches and its symbol error rate (SER) vs signal-to-noise ratio (SNR) performance is higher by about 9dB than the LSTM approach, maximum likelihood and other standard SIC based approaches like the minimum mean square error (MMSE) and least square (LS). This suggests deep learning-based receivers as strong candidates for the upcoming generations of wireless communication systems.