Federated Conformal Approach for Trustworthy Uncertainty Quantification in Distributed Fleet Prognostics

Prognostics is essential for accurate predictions of future system states in various applications such as manufacturing and healthcare. Federated learning enables collaborative training of prognostic models without sharing sensitive data, yet existing methods often lack reliable uncertainty quantification, which is critical for trustworthy decision-making. Conformal prediction provides statistically rigorous uncertainty quantification but relies on the exchangeability assumption, which is violated in federated prognostics due to cross-fleet heterogeneity and within-unit degradation or progression patterns.We propose a novel conformal prediction methodology to produce rigorous uncertainty quantification for prognostic models in federated settings. We analyze the coverage gap caused by exchangeability violations and introduce a practical strategy to tighten this gap through importance weights derived from a multilevel hidden Markov model and a federated generalized expectation maximization scheme. Our approach is model-agnostic and distribution-free and thus can be applied post hoc to various prognostic models to enhance their uncertainty quantification with formal statistical guarantees.Case studies on the engine degradation and Alzheimer’s disease progression datasets demonstrate that our approach consistently outperforms standard federated learning and conformal prediction methods, achieving target coverage with sharper uncertainty intervals. The code will be made publicly available on GitHub upon the publication of this paper.