Deep Learning Model-Based Channel Estimation for THz Band Massive MIMO with RF Impairments

TL;DR

This work addresses robust channel estimation for THz-band ultra-massive MIMO under RF impairments by modeling a hybrid-field channel combining near-field and far-field effects and phase noise (PN). It introduces a BiLSTM-GRU neural network that processes sequential PN-affected pilot observations, using real-imag separation to form input features, and denoises and estimates the PN-augmented channel. The method outperforms LS, MMSE, DNN, and vanilla LSTM baselines across SNRs and PN variances, with notable NMSE gains at low SNR and with larger antenna arrays; training is performed offline with subsequent online evaluation. The approach provides a practical pathway to reliable THz UM-MIMO operation and can be extended to related high-frequency mmWave systems where PN and hybrid-field effects are significant.

Abstract

THz band enabled large scale massive MIMO (M-MIMO) is considered as a key enabler for the 6G technology, given its enormous bandwidth and for its low latency connectivity. In the large-scale M-MIMO configuration, enlarged array aperture and small wavelengths of THz results in an amalgamation of both far field and near field paths, which makes tasks such as channel estimation for THz M-MIMO highly challenging. Moreover, at the THz transceiver, radio frequency (RF) impairments such as phase noise (PN) of the analog devices also leads to degradation in channel estimation performance. Classical estimators as well as traditional deep learning (DL) based algorithms struggle to maintain their robustness when performing for large scale antenna arrays i.e., M-MIMO, and when RF impairments are considered for practical usage. To effectively address this issue, it is crucial to utilize a neural network (NN) that has the ability to study the behaviors of the channel and RF impairment correlations, such as a recurrent neural network (RNN). The RF impairments act as sequential noise data which is subsequently incorporated with the channel data, leading to choose a specific type of RNN known as bidirectional long short-term memory (BiLSTM) which is followed by gated recurrent units (GRU) to process the sequential data. Simulation results demonstrate that our proposed model outperforms other benchmark approaches at various signal-to-noise ratio (SNR) levels.

Figures (5)

• Figure 1: THz Massive MIMO.
• Figure 2: Proposed deep learning model with BiLSTM-GRU.
• Figure 3: NMSE vs. SNR (dB) for $N=64$.
• Figure 4: NMSE vs. SNR (dB) for $N=128$.
• Figure 5: Performance Analysis.