Fatigue damage diagnosis and prognosis of an aeronautical structure based on surrogate modelling and particle filter

Abstract

A key issue affecting the performances of every human-conceived engineering system is its degradation, fatigue crack growth being one of the major structural deterioration phenomena. Fatigue crack growth is usually modelled as a stochastic process: uncertainty sources lie both in the item and in the physical degradation process variability. Fatigue crack growth deserves close attention, especially considering that condition-based maintenance methodologies are recently experiencing a major drive to increase their technology readiness level, requiring validated diagnostic and prognostic methodologies which should be capable of operating online and in real-time. In this regard, particle filters provide a consistent Bayesian framework, where the posterior distribution of the system degradation state is recursively approximated based on a time-growing stream of observations measuring the system response, enabling, in general, increasingly informed lifetime estimates. However, the real-time operation capability of such methods is hindered by their requirements in terms of computational power, which is mainly due to the complexity of the structural models they rely upon. Within this work, a comprehensive particle filter framework, able to deal with fatigue crack growth uncertainty sources while simultaneously addressing the computational burden issue, is proposed. The algorithm structure enables to simultaneously perform the diagnosis and prognosis of fatigue crack growth, while the adoption of the augmented state formulation allows to address scenarios where the degradation process of fatigue crack growth fails to meet the degradation model ruling the particle filter. Artificial neural networks–based surrogate modelling is adopted at different stages and embedded within the particle filter algorithm, relieving the computational burden associated with the evaluation of the trajectory likelihoods as well as enabling a fast estimation of the remaining useful life. Both simulated and experimental data sets regarding fatigue crack growth in an aluminium aeronautical panel are used for the algorithm testing, additionally proving the validity and effectiveness thereof by means of common prognostic performance metrics.

Fatigue modeling using neural networks: A comprehensive review

“…NN models are widely used as prognostics and health management tools to obtain a clear picture of the health state of the material (diagnosis) and an estimation of the remaining fatigue life (prognosis) 169 . This section is organized by fatigue diagnosis followed by fatigue prognosis.…”

Particle filter‐based hybrid damage prognosis considering measurement bias

“…3,44 In implementing the PF, at each time step, N p process noise samples will be sampled from the distributions in Table 2 and assigned to the N p particles. The noise distributions in this study are determined from Corbetta et al and Cristiani et al, 3,9,10 while the means and STDs are selected based on a trial-and-error procedure.…”

“…From the perspective of how the prognostic models are formulated, they can be classified into physics-based, [1][2][3] data-driven, [4][5][6] and hybrid methods. [7][8][9][10][11] Physics-based methods utilize specific mechanistic knowledge and theories to formulate pure physics-based models, which describe the structural degradation phenomena and the links between the damage states and the SHM measurements. Data-driven methods, resorting to data-driven modeling techniques such as neural networks 4,6 and Markov chains, 5 attempt to use a large amount of data to build the relationship between the internal degradation behavior and the external observations.…”

Particle Filter-Based Hybrid Damage Prognosis Considering Bias

“…All the PF parameters used in this study are reported in Table 3. The strategies about the distributions of initial intervals for C and m and the process noise ω are determined from [3,11,12].…”

“…In the last decades, a great variety of damage prognosis techniques have been developed in SHM depending on the availability of physics knowledge and data. From the perspective of how the prognosis models are formulated, they can be distinguished into physics-based [1][2][3], data-driven [4][5][6][7][8] and hybrid methods [9][10][11][12][13]. Physics-based methods utilize specific mechanistic knowledge and theories to formulate a pure physics-based model, which describes the structural degradation phenomena as well as the links between the damage states and the SHM measurements.…”