Distributed Fusion Using Attention Based Deep Distributed Learning with Multimodal Conformer

Distributed fusion has gained attention for its potential in tasks like tracking across sensor networks. Advances in communication and computational technologies enable deploying deep learning at network edges. This paper proposes a novel attention-based distributed fusion approach comprising three components: Optimal Distributed Set-Theoretic Information Flooding (ODSIF), an Encoding System (ES), and a Crossmodal-Conformer (CMCM). ODSIF exchanges and balances local and received data weights, while ES converts sensor node information into Gaussian heat maps. These heat maps are processed by CMCM to generate distributed fusion results independently at each node, eliminating the need for data association, a key step in traditional methods. The attention mechanism further mitigates outliers’ impact, ensuring robustness with less accurate sensor data. Simulations using Stone Soup, featuring three targets and seven sensors, demonstrate the proposed method’s superiority over DSIF, Confidence-Informed DSIF (C-DSIF), and Conformer.