Convolutional DPA Refinement for Real-Time Channel Estimation in High-Mobility Vehicular Networks

Abstract

Channel estimation in IEEE 802.11p vehicular networks must maintain reliable accuracy under severe Doppler conditions while meeting the receiver processing-time requirements of continuous frame reception. Although recurrent neural network (RNN)-based estimators can achieve competitive accuracy, their sequential hidden-state propagation leads to high per-frame latency that limits real-time applicability. To address this challenge, this paper proposes a two-part framework. First, a 1D residual dilated CNN (RDCNN) refines data pilot-aided (DPA) channel estimates across complete OFDM frames in a single parallel forward pass, avoiding the sequential per-symbol processing that dominates RNN-based latency. Second, a multi-channel mixed signal-to-noise ratio (SNR) training strategy exposes the estimator to diverse propagation environments and noise levels simultaneously, yielding a single model that generalises to unseen channel conditions without retraining or online adaptation. Simulation results across six Acosta-Marum vehicular channel models, including three reserved exclusively for testing, show that DPA-RDCNN with multi-channel mixed-SNR training surpasses RNN-based baselines in accuracy while reducing per-frame processing time by a factor of 16 to 32.