CNN Based Channel Estimation in 5G systems and performance analysis using STM32 Platform

---

Abstract

Channel estimation is a major problem for 5G systems, especially when using embedded platforms with limited computing power. Traditional least squares and minimum mean square error (MMSE) estimators work well in lab conditions, but they have trouble with the 15-20 dB noise changes that happen in real networks. We built a convolutional neural network (CNN)-based method for the STM32H753ZI platform, dealing with the ARM Cortex-M7's memory limits through careful quantization. Our system cuts memory use by 75% with only 2% accuracy drop compared to full-precision models. Tests on single-input single-output (SISO) and multiple-input multiple-output (MIMO) 2×2 setups using Indoor Hotspot (InH), Urban Micro (UMi), and Urban Macro (UMa) channel models give good results, though MIMO performance changes a lot with pilot density. The method shows that neural networks can work on low resource 5G hardware, but power use and timing needs are still big challenges for widespread use.

---

Keywords

5G systems, channel estimation, convolutional neural networks, embedded systems.