Data-driven spectral damage estimator for non-stationary vibration loading

Abstract

Material fatigue describes the failure due to cyclic loading. To ensure structural integrity, i.e. to design structures against material fatigue, a fatigue assessment is essential. Such an assessment can be conducted following either a statistical or a sampling-based ’non-statistical’ approach, often denoted frequency- resp. time-domain approach. Currently, the highly-efficient statistical approach uses power spectral densities to characterize loading and stresses, while the sampling-based approach implies the computationally costly processing of time-domain realizations. However, real-world applications often involve non-stationary vibration loading, which commonly causes significant discrepancies between these approaches, leaving no alternative to the sampling-based approach. To bridge this gap, we propose a statistical approach that qualifies for non-stationary loading. This employs the non-stationarity matrix to characterize non-stationary loading and to statistically calculate response kurtosis for linear structures, which then serves as input for a machine learning (ML) model that predicts its effect on fatigue damage. Herein, we detail the process of generating training data and defining an appropriate model. We compare the predicted fatigue damage and computational effort to those of the established approaches. Our findings suggest that the integration of the non-stationarity matrix for structural dynamics in connection with an appropriate ML model can significantly enhance the prediction of fatigue damage under non-stationary loading conditions while retaining computational efficiency.

Topics & Concepts