Low-Complexity High Speed Residual Network-Augmented Channel Estimation for mmWave Massive MIMO

This work addresses the crucial challenge of efficient design of computationally intensive channel estimation (CE) on resource-constrained edge platforms for millimeter-wave (mmWave) massive multiple-input-multiple-output (m-MIMO) in beyond 5G systems. We critically examine the performance limitation and computational bottlenecks of existing statistical CE methods such as least squares (LS) and linear minimum mean squared error (LMMSE) along with deep learning (DL) based CE. We propose a novel low-complexity high-speed CE algorithm based on Residual Network (ResNet) augmented LS (LS-ResNet) that significantly outperforms LS, LMMSE, and DL-based CE with an order of magnitude lower memory and complexity requirements than DL-based CE. The LS-ResNet offers signal-to-noise ratio (SNR) gains of 8 dB, 3 dB, and 2 dB over LS, LMMSE, and DL-based CE, respectively, for bit-error-rate (BER) of 10
⁻³
. The proposed LS-ResNet achieves remarkable speedups of nearly 1600 and 420 over DL-based CE on edge processors such as Cortex-A53 and Cortex-A9, respectively. An acceleration of existing DL-based CE with dedicated hardware is still slower than LS-ResNet on an embedded processor.