Model-Free Channel Estimation for Massive MIMO: A Channel Charting-Inspired Approach

TL;DR

This paper introduces a novel model-free deep learning approach for massive MIMO channel estimation that reduces pilot overhead and improves accuracy by learning low-dimensional channel representations.

Contribution

It combines autoencoders and LSTM networks with a channel charting-inspired loss to effectively estimate channels without relying on explicit models.

Findings

Achieves up to 9 dB improvement in normalized mean square error

Maintains scalability across different MIMO configurations

Outperforms traditional least-squares methods in challenging scenarios

Abstract

Channel estimation is fundamental to wireless communications, yet it becomes increasingly challenging in massive multiple-input multiple-output (MIMO) systems where base stations employ hundreds of antennas. Traditional least-squares methods require prohibitive pilot overhead that scales with antenna count, while sparse estimation methods depend on precise channel models that may not always be practical. This paper proposes a model-free approach combining deep autoencoders and LSTM networks. The method first learns low-dimensional channel representations preserving temporal correlation through augmenting a channel charting-inspired loss function, then tracks these features to recover full channel information from limited pilots. Simulation results using ray-tracing datasets show that the proposed approach achieves up to 9 dB improvement in normalized mean square error compared to the least-squares methods under ill-conditioned scenarios, while maintaining scalability across MIMO configurations.