Work

Physics-informed time series transformers for uncertainty-aware remaining useful life estimation under data scarcity

Year: 2026
Type: article
Abstract: Remaining useful life (RUL) estimation under limited run-to-failure data is challenging because purely data-driven models may overfit and provide unreliable uncertainty estimates. This study proposes two physics-informed patch time series transformer (PatchTST) architectures for uncertainty-aware bushing-wear prognostics. Physics-Informed PatchTST predicts a time-varying wear coefficient that is introduced into the Archard equation, whereas Residual Physics-Informed PatchTST (R-PatchTST) learns a residual correction around a nominal Archard-based wear trajectory. Both models use a probabilistic output head to represent aleatoric uncertainty, while predictive uncertainty is evaluated by comparing Monte Carlo (MC) dropout, deep ensembles, and conformal prediction. The methodology is assessed using six accelerated degradation experiments, leave-one-unit-out validation, and multiple input-window and forecasting-horizon configurations. Both physics-informed models improve wear forecasting relative to the standard PatchTST baseline in all six leave-one-unit-out cases on average, while R-PatchTST provides the best balance between accuracy and computational cost. Among the uncertainty methods, MC dropout achieves the best overall trade-off, with a mean root mean squared error of 0.0595 and a mean Winkler score of 0.497. The resulting framework obtains mean RUL errors of 0 . 97 ± 0 . 28 h and 2 . 00 ± 0 . 50 h at 12- and 24-h forecasting horizons, respectively. These results demonstrate the value of incorporating physical wear knowledge and explicit uncertainty evaluation into transformer-based prognostics under data scarcity.
Language: English
Open Access status: hybrid
Estimated APC paid: $4,400
OpenAlex ID: W7237339684
FWCI:
Cites: 59
Cited by:
Related to: