Deep learning-based end-to-end receiver for the NOMA system

Orthogonal frequency-division multiplexing access (OFDMA) greatly improves the frequency utilization in fourth generation (4G) wireless communication system, data exchange rate and system capacity for multiple users by dividing frequency selective channel into several orthogonal subcarriers. H...

Full description

Saved in:
Bibliographic Details
Main Author: Liu, Yifan
Other Authors: Teh Kah Chan
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2021
Subjects:
Online Access:https://hdl.handle.net/10356/149378
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Orthogonal frequency-division multiplexing access (OFDMA) greatly improves the frequency utilization in fourth generation (4G) wireless communication system, data exchange rate and system capacity for multiple users by dividing frequency selective channel into several orthogonal subcarriers. However, when facing the explosive growth of communication data, more and more users in communication system and requirement of high quality of communication service, fifth generation (5G) communication system should be developed to solve these problems. As a key technology for the 5G system, non-orthogonal multiplexing access (NOMA) becomes a critical technique in next-generation communication. Compared with OFDMA, NOMA shows advantage in spectrum efficiency, user access capability and system capacity. However, perfect channel state information (CSI) is difficult to be obtained under the rapidly changing channel condition in practical communication system with traditional channel estimation method such as least square (LS) and minimum mean square error (MMSE). In addition, the high computational complexity in receiver restricts the performance of NOMA system. Because of the ability of finding the relationship of input and output of a system, deep learning (DL) method and neural network (NN) are widely applied in many areas including communication system. In this dissertation, two DL methods, i.e., back-propagation neural network (BPNN) and recurrent neural network (RNN) are introduced to classify the received signal after channel transmission and restore the original signal, which play a role of obtaining CSI and decoding symbol. The traditional channel estimation and equalization methods, i.e., LS and MMSE, are used as a comparison to evaluate the performance of NOMA system with DL methods. Furthermore, the effect of system parameters, such as cyclic prefix (CP) length and pilot arrangement, will be discussed to promote the NOMA system proposed in this dissertation. Results discussed in Chapter four show that II BPNN and RNN methods outperform conventional methods, with a lower bit-error rate (BER) at same signal-to-noise ratio (SNR). Also, computational complexity is reduced by using DL methods. Finally, cyclic prefix length and pilot arrangement show an obvious effect on the performance of all the channel estimation methods.