Adaptive Coalition-Based Decision Fusion for Distributed Hydrophone Leak Classification
Ján Skalka, Małgorzata Przybyła‐Kasperek, Martin Drlík, Dominik Halvoník · IEEE Access · 2026
This study investigates coalition-based decision-making as a strategy for integrating predictions from dispersed local classifiers in a hydrophone-leak classification task. The experimental setting assumes that data are distributed across multiple local partitions, reflecting heterogeneous monitoring conditions in which spatial variability, noise, and uneven information content may limit the reliability of global inference. Local models produce probability vectors that are fused either through standard non-coalition strategies, namely simple and accuracy-weighted summation, or through a coalition-based mechanism. In the proposed approach, pairwise relations between prediction vectors are quantified using Pawlak's rank-based conflict distance or cosine dissimilarity, enabling the formation of homogeneous coalitions based on prediction similarity and heterogeneous coalitions based on prediction diversity. For each test instance, the strongest coalition is selected, and its members' prediction vectors are aggregated to produce the final decision. Experimental results across multiple dispersion scenarios show that coalition-based fusion frequently outperforms non-coalition baselines, with the clearest advantages appearing at moderate and high levels of dispersion. Weighted heterogeneous variants repeatedly achieve the strongest performance, while homogeneous coalitions remain competitive when local predictions are still relatively aligned. Overall, the results suggest that coalition formation can mitigate the negative effects of distributed, non-IID partitions and provide a robust, interpretable framework for late decision fusion in hydrophone-based leak classification.