DCEML: USV anomaly detection via Dual-Channel Explainable Multimodal Learning

Anomaly detection in Unmanned Surface Vehicles (USVs) is essential for maritime safety and autonomous navigation. Existing methods mainly rely on external observations such as GPS trajectories and navigational images, which are often insufficient for characterizing internal system states and explaining anomalous behaviors under complex marine disturbances. To address these challenges, we propose a Dual-Channel Explainable Multimodal Learning (DCEML) framework for tri-modal anomaly detection in USVs by integrating GPS trajectories, navigational images, and system logs. Firstly, we design a Log Dual-Channel Module (LDCM) that models log anomalies from structural and semantic perspectives to extract traceable internal evidence. Second, we develop a Tri-Modal Feature Fusion Module (TMFF) that aligns and integrates trajectory, image, and log features in a unified latent space through attention-based interaction, thereby improving robustness of anomaly discrimination under multisource ambiguity and conflict. Extensive experiments on a real-world USV dataset demonstrate that the DCEML framework consistently outperforms existing unimodal, general multimodal, and strong dual-modal baselines. Results show that introducing system logs not only improves detection accuracy and robustness, but also supports post-hoc traceability and diagnostic analysis. This work highlights the potential of jointly utilizing behavioral, environmental, and internal-state information for reliable anomaly detection in complex maritime environments. The code is publicly available at https://github.com/saxd-n/DCEML-TVCL .