Uncertainty-guided alignment for unsupervised domain adaptation in regression

In prognostics and health management systems, models must reliably predict asset health conditions across varying operating conditions, equipment manufacturers, or degradation patterns. However, obtaining labeled data for every new operational context is often impractical, particularly for run-to-failure trajectories. This work addresses this challenge through Unsupervised Domain Adaptation for Regression, which enables adaptation from a labeled source domain to an unlabeled target domain. Traditional feature alignment methods (such as adversarial or moment matching) underperform in PHM regression tasks due to the inherent correlation among learned features, leading to unreliable prognostic predictions when operating conditions change. This work proposes Uncertainty-Guided Alignment (UGA), a novel method that explicitly integrates predictive uncertainty into the feature alignment process. UGA employs Evidential Deep Learning to predict both target values and their associated uncertainties, using this uncertainty information to guide domain alignment and regularize the embedding space. This directly addresses key PHM requirements: (1) quantifying confidence in predictions when deployed in new operational conditions, (2) enabling reliable cross-operational context deployment. The approach is validated on two computer vision benchmarks and a real-world PHM case study of battery state-of-charge prediction, where domains are defined by different manufacturers (LG, Panasonic) and operating temperatures (-20 ∘C to 25 ∘C). Across 52 transfer tasks, UGA outperforms existing state-of-the-art methods on average. Our approach not only improves adaptation performance but also provides well-calibrated uncertainty estimates.