Deep learning-based receiver for downlink NOMA system

Non-orthogonal multiple access (NOMA) has grown to be an increasing significant part of wireless communication as it provides a higher spectral efficiency, massive connectivity, and other benefits. The successive interference cancellation (SIC) technique is typically implemented at the receiver in NOMA systems, where several users are decoded sequentially. Because of error propagation effects, the detection precision of SIC is heavily reliant on prior users' proper detection. This report describes the results of a preliminary study of deep learning (DL) in a NOMA system for detection and decoding in attempt to address this issue. The neural network in use is long short-term memory (LSTM) that is trained offline with simulation data and used to retrieve symbols from the transmission channel during the testing phase. Results from the simulation show that the DL model is able to outperform the traditional estimation methods in several scenarios of different test parameters. Some of the parameters include cyclic prefix, pilot symbols and modulation level. It is concluded that the DL model is able to improve the decoding precision in NOMA.Bachelor of Engineering (Electrical and Electronic Engineering