← back to homepage

C-PERL

Chained Physics-Enhanced Residual Learning for the Prognostics of Space Assets.

Javier Garrues Apecechea
Mathematical Institute & Department of Computer Science, University of Bonn
Work carried out during an internship at the European Space Agency (ESA) — ESOC, Darmstadt, Germany.

Abstract

Physics-based modelling in Prognostics and Health Management (PHM) is emerging as a key tool to support operations and maintenance of spacecraft assets, as it enables reliable and credible monitoring of components’ health. However, its effectiveness in operational settings is often limited by three key challenges: physical model discrepancies or “missing physics”, limited data availability, and unobservable or latent health indicators. To address these issues, we propose a novel Chained Physics-Enhanced Residual Learning (C-PERL) framework. The method introduces an intermediate, physically grounded variable that explicitly links and propagates information between physics-based and data-driven components, enabling i) estimation of discrepancies and biases of a physics model under data availability constraints; and ii) the use of an attention-based architecture to forecast key physically grounded health indicators. The proposed framework’s effectiveness is validated on a critical sensor from two European Space Agency (ESA) missions. Results show that the framework accurately estimates the sensor’s degradation and reconstructs parameter evolution, providing physically traceable predictions critical for decision-making in high-stakes space operations.

Poster

Poster to be added.

Last modified: August 2026.