Thompson, John

Yaghoobi Vaighan, Mehrdad

Luan, Dianxin

2024-05-09T12:53:50Z

2024-05-09T12:53:50Z

2024-05-09

Channel estimation is a fundamental challenge in wireless communication systems, as accurate knowledge of the communication channel is essential for achieving reliable data transmission. Conventional channel estimation methods often rely on statistical mathematics models and prior assumptions that may not accurately capture the dynamic and nonlinear property of wireless channels. In recent years, machine learning approaches have emerged as promising alternative solutions for channel estimation, as they basically aim for a local optimization and only requires available channel knowledge for prediction. Moreover, conventional methods cannot meet the demands for fifth-generation (5G) communication system and beyond while future communication systems need precise channel estimate to establish reliable and high-through links. This is where neural networks can provide improvement. Therefore, research on machine learning for channel estimation, especially neural network solutions, is attracting a growth of interests. In this thesis, I investigate the wireless channel estimation problem by using neural network solutions for orthogonal frequency-division multiplexing waveforms. I first concentrate on the low-complexity neural network design, aiming to achieve improved performance while simultaneously reducing complexity to enable practical implementation. I propose a neural network with low time and space complexity, which deploys bilinear interpolation layer for double-interpolation. Compared with other neural network solutions, the simulation results indicates improved performance for our approach. An encoder-decoder neural architecture is also proposed that incorporates self-attention mechanism encoder to emphasize the most crucial input information. This architecture utilizes a transformer encoder block as the encoder to capture vital elements and employs a residual convolutional neural network as the decoder. From the simulations, I demonstrate that our proposed approach achieves superior performance for both mean squared error (MSE) and bit error ratio (BER)} over other neural network solutions under consideration. The encoder-decoder architecture is then optimized for improved MSE performance, and also employs customized weight-level pruning to reduce the number of tunable parameters in the trained neural network. This enables parameter reductions up to 70\%, while maintaining nearly identical performance compared with the complete Channelformer network. Further, I propose an effective online training method, which only needs the available information at the receiver to update the neural network weights during its real-time operation. Using industrial standard channel models, the simulations of attention-based solutions show superior MSE performance compared with other candidate neural network methods. All current neural network solutions cannot be used in real wireless applications because they achieve narrow artificial intelligence. Narrow artificial intelligence often refers to a specific type of artificial intelligence (AI) in which learning algorithms are designed for a particular kind of task, and any knowledge gained from performing that task will not automatically be applied to other tasks. To fundamentally solve this, the training datasets are customized for training neural networks to achieve robust generalization to a wide range of channels, and this conceived training dataset itself also specifies the applicable channels. Further, I propose an benchmark training dataset for channel estimation task. Moreover, neural networks being trained with the actual and complete channel matrix still retain the denoising and interpolation capability on these previously unseen channels. To prove the applicability to general neural networks, one attention-based neural network and two convolutional neural networks are employed for testing. The trained neural networks are shown to achieve robust generalization on fixed power delay profile channels and variable delay channels. The effects of additive white Gaussian noise in the training dataset are studied. In addition, the effect of removing the cyclic prefix and the weight-level pruning on the proposed method are considered. In the end, I will discuss open research challenges and potential directions for further advancements in neural network-based channel estimation. These include the integration of domain knowledge, addressing the need for explainable models, and enhancing the resilience of neural network estimators against adversarial attacks.

https://hdl.handle.net/1842/41763

http://dx.doi.org/10.7488/era/4486

en

The University of Edinburgh

Luan, Dianxin, and John Thompson. "Low complexity channel estimation with neural network solutions." WSA 2021; 25th International ITG Workshop on Smart Antennas. VDE, 2021

. Luan, Dianxin, and John Thompson. "Attention based neural networks for wireless channel estimation." 2022 IEEE 95th Vehicular Technology Conference:(VTC2022-Spring). IEEE, 2022.

Luan, Dianxin, and John Thompson. "Achieving Robust Generalization for Wireless Channel Estimation Neural Networks by Designed Training Data." 2023 IEEE International Conference on Communications (ICC), Rome (Italy), IEEE, 2023.

Luan, Dianxin, and John Thompson. "Channelformer: Attention based Neural Solution for Wireless Channel Estimation and Effective Online Training." IEEE Transactions on Wireless Communications (2023)

machine learning

channel estimation

wireless communications

deep learning

attention mechanism

generalization

OFDM

Neural networks for channel estimation

Thesis or Dissertation

Doctoral

PhD Doctor of Philosophy