A Unified DNN-Based Channel Estimator For Massive MIMO Systems Under Various 5G Channel Scenarios

An accurate channel estimate is crucial for 5G and future wireless communications to achieve the performance gains offered by Massive MIMO systems. When dealing with less-than-ideal propagation conditions, standard estimation approaches such as Least Squares (LS) and Minimum Mean Square Error (MMSE) might be problematic in real-world deployment scenarios. In this paper, a unified model for channel estimation using Deep Neural Networks (DNNs) is presented. This model shows good performance in many real-world channel scenarios, including pilot contamination, time-varying channels, OFDM-based frequency-selective fading, spatially correlated MIMO channels, Rayleigh fading, and environments polluted by impulsive non-Gaussian noise. We model and evaluate the DNN's performance in comparison to LS and MMSE estimators, with normalized mean squared error (NMSE) serving as the primary metric. The findings show that the proposed DNN consistently outperforms traditional methods, with an NMSE improvement of up to 6 dB in challenging scenarios such as impulsive noise and pilot contamination. Another aspect of the DNN that proves its adaptability and durability is its strong generalization across different channel types and SNR levels. With only one DNN architecture trained on a diverse dataset encompassing all six channel conditions, scenario-specific estimators are unnecessary. Among the many feasible and scalable options for deployment in varied 5G networks, this research demonstrates that a unified DNN-based channel estimator may provide low inference latency and high estimate accuracy. This study's findings support the future use of data-driven approaches to communication system design, particularly in contexts where analytical modelling fails to provide satisfactory results.