Learning structural damage fields from time-domain wave propagation data with Bayesian neural networks

This study investigates the application of Bayesian Neural Networks for regression tasks aimed at recovering structural damage parameters from time-domain wave propagation data. The raw wave propagation data are simulated in a one-dimensional elastic bar using the Spectral Finite Element Method, considering data collection from a limited number of sensors. The damage fields are modeled as rectangular profiles, described by three parameters: position, size (or width), and magnitude. The inverse problem of damage recovery is formulated within a Bayesian framework, where the wavefield data collected by the sensors serve as inputs, while the outputs consist of probabilistic distributions of the damage parameters, allowing the quantification of uncertainty in the damage parameters and providing a stochastic characterization of the damage state. The training data include deterministic values for system properties, while the validation and test datasets incorporate uncertainties in the elastic bar’s properties, thereby simulating modeling errors during the inference stage of the inverse model. The main contribution of this work is to demonstrate the effectiveness of Bayesian Neural Networks as surrogate inverse models for damage characterization, addressing challenges related to model training and generalization. The results highlight the potential of this approach for reliable damage assessment in Structural Health Monitoring applications under uncertainty.