Signal Detection in Intelligent Reflecting Surface-Assisted NOMA Network Using LSTM Model: A ML Approach

Abstract

Non-orthogonal multiple access (NOMA) is already considered a viable multiple access scheme in fifth-generation networks. However, the stochastic behaviour of a wireless channel becomes a key performance limiting factor. To combat this, and with the advancement of metasurface technology, NOMA networks are being integrated with intelligent reflecting surfaces (IRSs) to improve signal strength. But IRS complicates the detection accuracy of a NOMA system, which is dependent on the correctness of the successive interference cancelation (SIC) process. In this article, we propose a machine learning (ML)-based approach to perform joint channel estimation and signal detection in an IRS-enabled uplink NOMA network under efficient mitigation of SIC error propagation. The proposed scheme exploits a four layer deep learning (DL) model by employing a long short-term memory (LSTM) core structure. Further, to optimize the phase shifts of IRS, we exploit a low complexity iterative solution using the element-wise block coordinate descent (EBCD) method. Monte Carlo simulations are performed to analyze the performance of the proposed scheme, and the findings show a considerable improvement in channel estimation and signal detection using the LSTM based IRS-NOMA receiver. The comparison is made with a maximum likelihood detector employing conventional SIC scheme using least squares and minimum mean square error channel estimation approaches in a realistic path loss channel model with severe inter-symbol interference.