Bayesian Post-Repair Prognostics for Reliable RUL Prediction
M. Salinas-Camus (TU Delft - Aerospace Engineering)
M.A. Patrick (TU Delft - Aerospace Engineering)
J.A. Pascoe (TU Delft - Aerospace Engineering)
N. Eleftheroglou (TU Delft - Aerospace Engineering)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Prognostics aims to predict the Remaining Useful Life (RUL) of engineering assets and is essential for effective Predictive Maintenance (PdM). Unlike preventive maintenance, PdM offers substantial cost benefits by scheduling maintenance only when needed. However, most existing prognostic models assume that repairs return an asset to an “as-good-as-new” condition. In practice, repairs are often imperfect, as they only partly restore the asset and may change its subsequent degradation behavior. This mismatch represents a major limitation of current prognostic approaches, as poor prognostic performance can lead to unnecessary maintenance actions or unexpected failure.
This paper proposes a fully Bayesian prognostic model, named the Sequential Bayesian Semi-Markov Framework (SBSM), that explicitly accounts for imperfect repair while being trained exclusively on data from non-repaired assets. The framework combines a Hidden Semi-Markov Model (HSMM) to represent the degradation model with a particle filter for the predictive step. Repair actions are incorporated as prior distributions that represent repair effectiveness, enabling repair uncertainty and prognostic uncertainty to be treated in a unified manner. This formulation allows post-repair degradation trajectories to differ from pre-repair behavior without retraining the model.
The approach is experimentally validated using an in-house dataset of aluminum open-hole specimens subjected to constant-amplitude fatigue loading, repaired via cold spray deposition, and tested to failure. Baseline (non-repaired) specimens are used for training, while repaired specimens are used exclusively for testing. The proposed method is compared against a Convolutional Neural Network (CNN) baseline. Results show that the SBSM achieves lower prediction error and improved probabilistic calibration, particularly under significant distributional shifts induced by more effective repairs. The framework demonstrates robust post-repair RUL prediction and well-calibrated uncertainty estimates, highlighting its potential for real-world predictive maintenance applications involving imperfect repair.
help_outline