Data-driven Millimeter-wave Channel Modelling Based on Attention Mechanism

Abstract

This paper presents a novel data-driven millimeter-wave (mmWave) channel modeling approach enhanced by an attention mechanism for 16-QAM signal reconstruction. Leveraging the sparsity and dynamic nature of mmWave channels, we integrate an attention module into a Long Short-Term Memory (LSTM) network to dynamically prioritize dominant multipath components and critical temporal features. The proposed model learns the intricate mapping between transmitted and received signals, effectively mitigating distortions caused by multipath fading, phase noise, and frequency offsets. Simulation results demonstrate significant improvements, achieving a 52% reduction in Symbol Error Rate (SER) and a 10.6 dB gain in Normalized Mean Square Error (NMSE) compared to conventional LSTM models. This work establishes a robust framework for adaptive channel modeling in 5G/6G systems, enabling high-fidelity signal recovery in complex mmWave environments.