DD-Former: A Dual-Domain Fusion Transformer for Sensing-Assisted Beamforming Prediction

Next-generation wireless communication systems, particularly in vehicular networks, demand highly reliable and low-latency beamforming. Multi-modal sensor fusion, leveraging data from cameras, LiDAR, and radar, has emerged as a promising approach to predict optimal beam configurations by providing rich environmental context. Existing models primarily operate in the spatial domain, overlooking subtle but critical features embedded in the frequency domain. This vulnerability degrades real-world performance by preventing the model from processing corrupted high-frequency components, which undermines the final beam prediction. In this paper, we propose DD-Former, a Dual-Domain Fusion Transformer architecture that enhances multi-modal learning by integrating frequency-domain feature processing into a hierarchical framework. We adapt the Frequency-Domain MLP (Fre-MLP) module and insert it as a feature enhancement unit following each spatial feature extraction stage. This dual-domain approach enables the network to learn a more comprehensive joint representation from both spatial and frequency characteristics of sensor data, thereby improving the accuracy and robustness of beamforming prediction in dynamic environments.