An efficient computational approach for generating synthetic data to train neural networks in concrete bridge monitoring

Machine-learning tools can automate inspection and monitoring of concrete bridges, but they require large, labeled datasets that encompass many damage scenarios. Conventional two-dimensional and three-dimensional nonlinear finite element models involve a high computational burden, which limits their practicality for generating large-scale datasets. This study proposes an efficient physics-based framework that couples a force-based fiber beam element with an enhanced damage-plasticity constitutive law accounting for partial crack closure, thus reproducing both nonlinear static responses and frequency shifts associated with beam cracking that underpin vibration-based Structural Health Monitoring. Validation against a prestressed beam laboratory test and the full-scale Alveo Vecchio viaduct demonstrates that the model matches load-displacement curves and crack-related frequency variations, while significantly reducing the computational burden compared to two-dimensional and three-dimensional finite element models. The resulting efficiency enables the execution of a large number of nonlinear simulations spanning elastic, cracking and yielding regimes. These synthetic responses train two neural networks for damage identification: (i) a Nonlinear AutoRegressive network that performs unsupervised novelty detection and (ii) a Long Short-Term Memory for supervised time series classification. Together the networks detect and classify damage with high accuracy in real time, illustrating how simulation-driven datasets can accelerate physics-informed Structural Health Monitoring of ageing bridge infrastructure.