Deep learning models for channel estimation and modelling in wireless networks

Abstract

This thesis focuses on studying radio propagation channels in relation to channel modelling and estimation using deep learning (DL) models. Specifically, there search examines challenges associated with path loss (PL) modelling, channel estimation, and radio-frequency electromagnetic fields (RF-EMF) analysis and modelling in future wireless networks. To address these challenges, this research employs DL methods to propose models that can improve the prediction accuracy compared to current state-of-the-art (SOA) methods. Firstly, the thesis proposes a DL-based multi-frequency PL model for urban and suburban non-line-of-sight (NLoS) scenarios in a fifth-generation (5G) wireless system that covers 13 frequencies from 0.8 GHz to 70 GHz. The proposed deep neural network (DNN) based PL model outperforms the conventional ABGPL model by achieving approximately a 6 dB improvement in terms of mean square error (MSE) over 13 frequencies. Secondly, this thesis investigates the channel estimation for a wireless system that integrates RIS (reconfigurable intelligent surface) and non-orthogonal multiple access (NOMA), referred to as the RIS-NOMA system. The convolutional long-short-term memory (CNNLSTM) model is proposed for channel estimation in the RIS-NOMA system. The proposed CNN-LSTM model is superior to other SOA DL models, including CNN (using a two-dimensional convolutional layer), LSTM, CNN1D-LSTM(using a one-dimensional convolutional layer), and CNN1D-BiLSTM (using bi-directional LSTM layer), in terms of estimation accuracy. Thirdly, the thesis studies DL models to forecast long-term RF-EMF based on the time-series data measured at 17 locations in Turkey. Three scenarios including single-step input and single-step output (SISO), multi-step input and single-step output(MISO), and multi-step input multi-step output (MIMO) are examined using DNN, CNN, LSTM, and transformer models. SISO has low accuracy, while CNN and LSTM outperform DNN in MISO and MIMO with limited input and label width. The transformer outperforms other DL models, especially in the MIMO scenario when label width and shift increase.

Date of Award	Nov 2023
Original language	English
Supervisor	Adnan Ahmad Cheema (Supervisor) & Omar Escalona (Supervisor)