Low-Complexity and Low-Energy RNN-Enhanced Kalman Filter for Channel Estimation: A Case Study

In radiocommunications, multipath propagation often induces Inter-Symbol Interference (ISI), which requires accurate channel estimation and equalization. Neural networks have recently shown promising performance for this task but remain limited by their high computational cost and lack of interpretability, hindering their adoption in mission-critical systems. To overcome these limitations, this study proposes a hybrid estimation architecture that combines a linear Kalman filter with two lightweight neural networks that dynamically and deterministically adjust the noise covariance matrices of the filter, enabling efficient and explainable channel tracking. Experimental results show performance gains of up to 5 dB at low SNR and 3 dB at high SNR compared with the Least Squares (LS) algorithm. This low-dimensional design enables real-time deployment on energy-constrained programmable platforms. Implementations on an ARM Cortex-A72 CPU and Artix-7/Virtex-7 FPGAs demonstrate throughputs of 53.7 Mbps and 8-20 Mbps, respectively, with FPGA energy consumption over 20× lower than on the CPU. These results confirm the effectiveness of the proposed NN-enhanced Kalman estimator for interpretable and energy-efficient channel estimation in embedded and defense applications.