Data-driven structural identification of nonlinear assemblies: Uncertainty Quantification

Nonlinear model identification from vibration data is challenging due to limited measured data collected during the testing campaign and since the identified model should be capable of accounting for the uncertainties arising from the reassembly of the structure, environmental effects, and slight changes in parameters as a result of wear during vibration testing. In this paper, a new technique based on ensembling is proposed for uncertainty quantification during the identification of nonlinear assemblies using multiple data sets. First, an ensemble of parsimonious models is identified using a physics-informed nonlinear model identification method from subsets of measured data. Aggregate model statistics are then employed to calculate inclusion probabilities for the candidate model, which enable uncertainty quantification and a probabilistic estimate of the dynamic response. This results in a robust nonlinear model identification with physical interoperability. An application on a single-degree-of-freedom system idealised for an experimental structure with geometric and friction nonlinearities is presented. The results obtained demonstrate the substantial performance of the proposed technique in selecting accurate nonlinear models that capture the response over a large range of variability and repeatability for real-world data sets.