Channel estimation for IEEE 802.11p vehicular communication using deep learning

Abstract

Vehicular communication systems require precise channel estimation in dynamic propagation environments. Traditional methods struggle in high-mobility settings due to rapid. Doppler shifts and multipath effects. Recent deep learning (DL) approaches demonstrate improved accuracy but face challenges in generalisation across diverse operating conditions, inference latency, and interpretability. This thesis investigates DL-based channel estimation for IEEE 802.11p vehicular communications through three main contributions. First, different DL architectures are proposed and evaluated. A novel architecture that combines data pilot-aided (DPA) estimation with non-causal one-dimensional convolutions was found to achieve superior bit error rate (BER) performance with a median single-frame inference latency of 4.16 ms, satisfying the receiver processing budget for IEEE 802.11p continuous frame reception. The architecture employs a processing order where DPA first provides temporal tracking anchor points, whilst the temporal convolutional network (TCN) corrects errors through dilated convolutions. Secondly, training strategies that improve generalisation across diverse noise levels and propagation environments are investigated. Mixed signal-to-noise ratio (SNR) training substantially improves performance at low SNR across all evaluated architectures. Multichannel training enables a single model trained on combined data from three propagation models to generalise effectively to unseen environments, eliminating the need for channelspecific estimators. Thirdly, the Relevance-based Explainable Analysis for Channel estimation (REACH) framework extends layer-wise relevance propagation (LRP) for interpretability-driven optimisation. For feedforward networks, REACH enables 43.6% parameter reduction without significant performance degradation. For multi-channel TCNs, cross-channel relevance correlation exceeds 0.81, explaining the strong generalisation to unseen propagation environments. Analysis of cross-channel important features reveals that only 19.6% of input features are critical across six channel models. Combined, these contributions establish the DPA-TCN architecture as an effective estimator for IEEE 802.11p systems, combining efficient single-frame inference, generalisation across diverse conditions, and interpretability-driven optimisation.