Spectral Efficiency Improvement in THz Massive MIMO System Through Federated Learning Based channel Training

Abstract

This work proposes an efficient channel training and estimation framework for Terahertz (THz) massive MIMO systems using a Federated Learning (FL)-based approach integrated with Convolutional Neural Networks (CNN). Although THz communication enables ultra-high data rates for next-generation wireless systems, it faces challenges such as severe path loss, noise sensitivity, and rapidly varying channel conditions, making accurate channel estimation difficult. Traditional centralized methods require collecting pilot signals at a central server, leading to high communication overhead, increased computational complexity, and limited scalability. To overcome these limitations, the proposed approach employs a distributed learning framework in which multiple users locally train CNN based channel estimation models using their respective pilot signals. Instead of sharing raw data, only model parameters are transmitted and aggregated using the Federated Averaging (FedAvg) algorithm to form a global model. This ensures reduced communication overhead, improved scalability, and enhanced data privacy. Simulation results demonstrate that the proposed Fed-CNN model outperforms conventional techniques such as Least Squares (LS) and Minimum Mean Square Error (MMSE). The method achieves lower 1 Normalized Mean Square Error (NMSE), faster convergence, and higher spectral efficiency, reaching approximately 6.7 bps/Hz at high Signal-to Noise Ratio (SNR) levels. Additionally, the model exhibits strong robustness under dynamic channel conditions. Overall, the results validate the effectiveness of combining federated learning with deep learning for accurate and efficient channel estimation in THz massive MIMO systems.