Ensemble of hybrid neural networks to compensate for epistemic uncertainties: a case study in system prognosis

Abstract

Despite the growing popularity of deep learning, even with recent progresses, many existing methods struggle with applications challenged by non-ideal datasets and partial understanding of input–output relationships. This is the case of prognosis and health management of industrial assets, where operating conditions might be recorded, but failure of equipment is only partially observed and partially understood. In this contribution, we address the issue of incomplete knowledge in models used to estimate time-dependent damage propagation. Specifically, unaccounted corrosion-fatigue of aircraft fuselage panels is presented as a case study. While degradation resulting from mechanical fatigue is accounted by damage models, unforeseen corrosion effects are not, yielding in large discrepancies between predicted and observed crack lengths. To address this model-form uncertainty, we propose using hybrid neural networks. Hybrid models are composed of physics-informed layers addressing well-understood aspects of damage accumulation, while data-driven layers are trained to act as correction terms. To improve overall accuracy, ensemble techniques are adapted to merge resulting predictions. Optimal ensemble weights are derived to help with the task of defining safe operation in the fleet. The proposed case study encompasses highly imbalanced data sets with uncertainties acting asynchronously. Main contributions of the proposed work are: (i) in a modeling perspective, an approach capable of compensating for model-form uncertainty by formulating the numerical integration of governing ordinary differential equations as hybrid recurrent neural network models; (ii) from a prognosis perspective, the proposed methods can be used to rank damage severity, allowing to prioritize aircraft for inspection and/or route reassignment.