Low-Complexity Residual Network-Augmented Channel Sounding in mmWave Massive MIMO on System on Chip

Accurate and fast channel sounding (CS) for beam identification is essential for reliable beamforming in millimeter-wave (mmWave) massive multiple-input multiple-output (mMIMO) systems. CS using conventional channel estimation (CE) techniques, such as least squares (LS), suffer from low performance, while the minimum mean squared error (MMSE) is complex, especially with the high-dimensional channel matrix in multi-carrier mMIMO systems. Recently, deep learning (DL) techniques have shown improved CE. However, their computational complexity is a concern, especially on resource-constrained edge platforms. In this work, we propose a lightweight residual network-augmented CE method (LS-ResNet) for CS, where we augment an LS CE with a DL model. We also consider algorithms-to-architecture mapping of existing and proposed CE on the Zynq system-on-chip (SoC) via hardware-software co-design. The LS-ResNet CE outperforms LS, MMSE, and multilayer perceptron (MLP)-based CE by 8 dB, 3 dB, and 2 dB in signal-to-noise ratio (SNR) at a bit-error-rate (BER) of 10−3. LS-ResNet operates 10× faster, requires 23% fewer resources than MLP-based CE, and is robust to wireless channel variations. We explored depthwise separable convolutions (DSConv) based LS-ResNet architecture, which reduces the execution time further by 2.5× on the edge processor while maintaining comparable BER performance.