Throughput optimized channel smoothing

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Electrical Engineering | Master's thesis

Department

Mcode

Language

en

Pages

55

Series

Abstract

Wireless communication systems have traditionally relied on carefully designed receiver algorithms, such as channel estimation, to mitigate impairments and ensure reliable data transmission. Although these traditional methods have been refined over decades, they often fall short when faced with the complexity of real-world wireless environments. In recent years, Artificial Intelligence (AI) has become a promising tool to enhance or even replace conventional receiver components. However, many AI-based methods in channel estimation still rely on distance-based loss functions for channel smoothing. These objectives may not correlate well with overall system performance and can lead to overfitting, especially when estimation errors have little impact on throughput. This thesis investigates the use of AI for channel estimation, especially channel smoothing, in the 5G receiver chain with a particular focus on the Physical Uplink Shared Channel. A novel neural network is proposed that integrates Convolutional Neural Network and Residual neural Network, trained with a system-level, throughput-oriented objective function. By leveraging a fully differentiable receiver chain, the neural network is able to directly optimize decoding performance at the bit level. This method is compared against a conventional AI method using the same neural network architecture but optimized with a loss based on ideal channel conditions. Both methods are trained under simulated 5G channel environments Clustered Delay Line (CDL) A/B/D and evaluate under CDL C/E, using the Sionna simulation framework. Results demonstrate that the throughput-optimized neural network consistently outperforms traditional least squares estimation in terms of Bit Error Rate and Block Error Rate, and in certain scenarios also surpasses the neural network trained with ideal channel conditions. Because ideal channel conditions are only available in simulation, these results highlight the practical value of learning directly for throughput and point toward a viable path for AI-enhanced receivers in real-world 5G deployments.

Description

Supervisor

Vorobyov, Sergiy

Thesis advisor

Olives, Jean-Luc
Susan, Eeli

Other note

Citation