Robust Multimodal Beam Prediction With Missing Modality

Utilizing deep learning to fuse multimodal data for millimeter-wave beam prediction can achieve significant performance improvement compared to the conventional unimodal approaches. However, current multimodal learning approaches highly rely on the availability of full modalities, which is usually not practical due to various non-ideal factors. In this letter, we focus on the design of deep multimodal learning based beam prediction robust to missing modalities. Specifically, a data augmentation technique consisting of available modality masking and missing modality imputation is proposed to improve data diversity for efficient model training. Moreover, the channel attention mechanism is customized to adaptively fuse the multimodal features in the presence of missing modalities. Simulation results demonstrate that our proposed approach significantly outperforms its conventional multimodal learning counterparts in handling missing modalities.