Backbone agnostic Pareto evidential networks for trustworthy fault diagnosis and out of distribution detection

Reliable decision-making under uncertainty is a central challenge in both engineering and scientific domains, yet remains difficult when systems encounter conditions beyond those observed during training. Such out-of-distribution (OOD) events are particularly problematic in intelligent fault diagnosis, where overconfident misclassification of novel faults can compromise safety and erode trust. Here, we introduce the Pareto-driven Evidential Dual-head Network (PEDNet), which unifies evidential uncertainty modeling with multi-objective optimization theory to address this challenge. PEDNet features a dual-head architecture-one head for fault classification and the other for epistemic uncertainty estimation-coupled via a dynamic Pareto weighting mechanism that resolves gradient conflicts without manual tuning, which provides a principled route to simultaneous improvement of predictive accuracy and OOD sensitivity, supported by theoretical grounding in Pareto stationarity. Across two industrial time-series datasets, PEDNet consistently enhances classification performance while markedly reducing OOD false alarms. Beyond its empirical advantages, the framework is lightweight, backbone-agnostic, and readily extendable to other domains requiring uncertainty-aware reasoning, offering a generalizable approach to trustworthy decision-making under distributional shifts.