Particle filtering‐based adaptive training of neural networks for real‐time structural damage diagnosis and prognosis

Abstract

Real-time failure diagnosis and prognosis play key roles in prognostics and health management of critical engineering assets. Physics-based (or empirical–phenomenological) and data-driven methods have been proposed in the literature to solve damage diagnosis and prognosis challenges. Among data-driven methods, neural networks have often successfully proven their regression capabilities for the complex, nonlinear problems typical of this field research field. Indeed, the need to resort to machine learning techniques (e.g., neural networks) becomes stronger when physics-based (or empirical–phenomenological) models are lacking for describing the degradation behavior of the systems under consideration. In this work, we propose to exploit the flexibility offered by neural networks in order to adaptively, and in real time, learn from the monitored structure and derive models for diagnosis and prognosis of structural components subject to damage processes. To this aim, neural networks are embedded within a particle filtering scheme, and the training of the network is performed sequentially and in real time, as observations become available during the damage evolution. The structural health monitoring tool is capable of sequentially updating itself using the observations gathered during the system's operation, avoiding the typical limitations associated to off-line training schemes, which are, in general, not capable of capturing possible changes in the dynamical degradation behavior, when not properly included in the training patterns. For validation and comparison purposes, the method is demonstrated on simulated fatigue crack growths and real crack evolution dynamics observed in a metallic aeronautical panel loaded with a changing-amplitude fatigue load, during an accelerated laboratory test.