RNN-Based Low Complexity High Speed Channel Estimation Architectures for Vehicular Networks

Wireless channel estimation (CE) in vehicular communication networks is challenging due to mobility, and various data-pilot-aided (DPA) and deep learning (DL) techniques have been proposed to improve the reliability of CE without increasing the pilot overhead. However, the complexity and latency of such approaches are a concern, especially when the wireless physical layer (PHY) is deployed on resource-constrained edge devices. In this paper, we propose a low-complexity, high-speed architecture of Recurrent Neural Network (RNN) based Gated Recurrent Unit (GRU) augmented CE and compare it with RNN-based long short-term memory (LSTM) augmented CE and other DL-based CE on the system on chip (SoC). We compare their bit-error-rate (BER) performance using end-to-end PHY over a wide range of signal-to-noise ratio (SNR) and vehicular channels. We demonstrate a significant improvement of 21 in latency over LSTM-augmented CE for nearly identical ×BER. When compared to other DL-based CE, GRU-based CE offers a latency improvement of 1.26× and 21% reductions in multipliers with significantly improved BER. For identical BER, GRU-based CE offers a latency improvement of 1.28× and 25% reductions in multipliers over DL-based CE.