From coupon to system : a deep learning and transfer learning framework for RUL prediction in multi-component structures
Loading...
Date
2026-06-26
Authors
Βλαχοπούλου, Ευαγγελία
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Prognostics health management (PHM) of metallic structures relies increasingly on data-driven methods for Remaining Useful Life (RUL) prediction yet the transition from coupon-level to system-level prognostics remains an open challenge, particularly under limited system-level data availability. Existing approaches typically address single-component degradation, yet a methodological gap remains in methods that can effectively leverage coupon-level knowledge to perform uncertainty-aware RUL prediction for interconnected multi-component systems. This paper proposes a transfer learning framework for system-level RUL prediction and uncertainty quantification of a three-coupon aluminum system, validated on experimental fatigue data comprising five coupons and six triplet specimens. A Genetic Algorithm (GA) is first employed to construct a health indicator from strain sensor measurements, providing a monotonic and informative input representation for the prognostic model. A Long Short-Term Memory (LSTM) network is then pre-trained on single-coupon fatigue data and subsequently adapted to the system level via Low-Rank Adaptation (LoRA), enabling parameter-efficient knowledge transfer under scarce system-level training data. System-level RUL (SRUL) prediction is evaluated under two failure criteria: first-coupon failure, representing a conservative operational limit, and last-coupon failure corresponding to total system collapse. Uncertainty quantification is performed through a dual decomposition strategy, separating epistemic and aleatoric uncertainty. The proposed framework is evaluated against two baselines: a Zero-Shot Baseline, which estimates system RUL as the minimum (or maximum) of coupon-level RULs at each time step without any system-level training, and a Standard Transfer Learning model that uses the pretrained LSTM on system-level data without LoRA adaptation. The LoRA-enhanced framework consistently outperforms both baselines, demonstrating that LoRA-based transfer learning from coupon- to system-level prognostics is an effective and data-efficient strategy for structural PHM, offering a principled path toward uncertainty-aware life prediction in interconnected structural systems.
Description
Keywords
Remaining useful life, Transfer learning, Prognostics health management