Combining experimental and synthetic data in Deep Learning for damage detection
Please login to view abstract download link
Data-driven methods have demonstrated their strong capabilities in structural damage identification. Exploiting the measured data allows the characterization of the normal behavior of large structural systems to detect novelties and reduce the need for periodic inspections. However, providing a reliable and safe assessment requires incorporating knowledge of how damage affects the target system. Using a Finite Element (FE) parametrization that describes the physics behind the behavior of the structure is a common practice in damage identification. The FE model allows solving fast simulations under desired damage scenarios, widening the spectrum of responses beyond the healthy state~\cite{Ana2023}. This work proposes a feasible methodology to enrich a Deep Neural Network (DNN) training dataset with synthetic measurements from FE simulations accounting for environmental and operational conditions (EOCs). The methodology applies a clustering technique to the available long-term monitoring data to identify representative measurements that cover most of the variability occurring under normal operation. The selected measurements serve as the target response to calibrate a FE parametrization and build a database of damage scenarios labeled in terms of location and severity that complements the available experimental data. The methodology is implemented and validated on a full-scale operating case study: the Infante Dom Henrique bridge. Results reveal an increased accuracy in the severity and location estimates under environmental variability.
