ABSTRACT
Real-time damage diagnosis and prognosis are fundamental tasks for effective health management of engineering structures. A number of physics-based and data-driven methods have been proposed in the literature to address a variety of damage diagnosis and prognosis challenges. Among them, neural networks have proven their strength and flexibility when suitable physics-based models are not accessible, and field data are available for sufficiently accurate regressions/classification. Their flexibility is particularly beneficial for prognosis problems, where unaccounted external variables may introduce new, un-modeled system dynamics. In this work, we propose to train neural network models in real-time by embedding them in a particle filtering scheme. The proposed algorithm learns from the monitored structure, and is capable of sequentially adapting itself to changes in the degradation dynamics. For validation purposes, the method is demonstrated with reference to a numerically-simulated fatigue crack growth in an aeronautical panel.
