A Kolmogorov–Arnold-Network-centric mixture of experts transfer learning strategy for underwater target detection

This paper proposes a novel KANcMoE (Kolmogorov-Arnold-Network-centric Mixture of Experts) transfer learning strategy designed to achieve high-precision underwater target detection under data-scarce conditions. First, a comprehensive multi-domain feature extraction framework is established to synthesize time-domain, frequency-domain, and time-frequency characteristics for robust signal representation. Structurally, the KANcMoE architecture integrates a dynamic gating mechanism that adaptively modulates the contribution of KAN and MLP experts based on feature complexity, thereby synergizing their complementary strengths. Furthermore, a decoupled parallel collaborative fine-tuning mechanism is introduced to facilitate domain adaptation. This strategy moves beyond traditional global updates by simultaneously optimizing the routing logic and expert parameters, enabling the efficient reuse of few-shot data to preserve source knowledge while rapidly adapting to the target domain. Experimental evaluations on real-world sea trial datasets demonstrate that using merely 5% of labeled samples, the proposed method achieves detection accuracy improvements of up to 13.13% compared to mainstream baselines. This approach offers an effective solution for underwater acoustic target detection in environments with limited labeled data. • Multi-domain feature extraction captures underwater signal patterns. • A novel hybrid framework combines Kolmogorov-Arnold Networks and MLPs. • An adaptive gating mechanism enables collaborative expert optimization. • Transfer learning enhances model convergence on small-sample datasets. • Strategies are validated using real-world underwater acoustic datasets.