Reconfigurable Deep Learning Augmented Channel Estimation for Vehicular Networks on System on Chip

Accurate and fast Channel Estimation (CE) is critical for reliable vehicular communications. To address error accumulation issues in Time-domain Reliable Frequency-domain Interpolation (TRFI) and Spectral Temporal Averaging (STA)-based CE, the Deep Learning (DL) is being explored to enhance the performance of STA/TRFI or replace them. However, the high complexity of the DL especially for wireless radio which is deployed on resource-contained edge platforms is a concern. In this work, we propose: 1) A modified TRFI (TRFIM) achieving 87% higher accuracy than TRFI, with 99.4% faster execution and 12% lower complexity on Zynq System-on-Chip; 2) Hardware implementation of low-complexity DNN-augmented STA and TRFIM (referred as STADNN and TRFIMDNN, respectively) through hardware-software co-design and fixed-point optimization. Our implementations demonstrate significantly improved CE accuracy over conventional TRFI and TRFIM. Fixed point optimization reduces STADNN and TRFIMDNN latency and complexity by 29%, 68% and 19%, 57%, respectively. Though TRFIMDNN achieves better accuracy than STADNN over a wide range of signal-to-noise ratio (SNR), STADNN is 1,8 times faster. Also, STADNN is superior at a low SNR, while TRFIMDNN is superior at a high SNR. We propose an intelligent reconfigurable architecture that dynamically switches between STADNN and TRFIMDNN, requiring 40% fewer resources and 9.8% less power than non-reconfigurable architectures.