Estimation of Numerical Data Reliability in Damage Detection Tasks

Numerical simulations with Finite Element models can produce vibration structural responses in large numbers that are necessary to train data-driven Structural Health Monitoring and Damage Detection classifiers. The numerical data has potential to substitute destructive and costly experiments required for high quality labeled training data. However, the simulation error usually contaminates the training data, leading the later classifiers in errors when they are presented with real experimental data. The present work deals with estimating the reliability of such numerically generated data for different monitoring tasks. Finite element models with perturbed properties are used to mimic the disagreements between experiment and numerical models. A neural network is used then to estimate the error of the monitoring classifier when presented with unseen perturbed data. The goal is to provide engineers with a future tool that can accompany data-based classifiers trained by simulation data.