MELF-XFD: Trustworthy and Interpretable Multi-Expert Fusion for Safety-Critical Fault Diagnosis in Autonomous Underwater Vehicles

Autonomous underwater vehicles (AUVs) operating in harsh, communication-limited underwater environments require robust fault diagnosis systems to ensure mission success and safety. Existing model-based approaches suffer from hydrodynamic modeling inaccuracies, while data-driven methods typically employ monolithic architectures that lack interpretability and real-time deployment awareness. This paper proposes <bold>M</bold>ulti-<bold>E</bold>xpert <bold>L</bold>ightweight <bold>F</bold>usion Model with Ethical &amp; <bold>E</bold>xplainable <bold>F</bold>ault <bold>D</bold>iagnosis (MELF-XFD), a novel model-free framework that addresses these limitations through physically-grounded multi-domain signal decomposition. MELF-XFD performs explainable fault diagnosis by fusing temporal, spectral, and statistical representations of multivariate AUV time-series data using lightweight expert networks with adaptive weighting for real-time onboard deployment. Experimental results on a public AUV benchmark with five fault classes show that MELF-XFD outperforms existing methods, achieving a 95.7&#x0025; macro-F1 score, 90.0&#x0025; severe fault recall, and 39 ms inference latency. It attains the highest composite diagnostic criterion around 93.6&#x0025;, balancing accuracy, safety-critical sensitivity, and computational efficiency. A low expected calibration error of 0.0880 ensures reliable confidence estimates for safety-critical deployment. Ablation studies confirm the critical role of temporal and frequency experts, while adaptive fusion enables interpretable fault attribution, establishing MELF-XFD as a practical and deployment-ready solution.