Vibration-based damage detection of structures is increasingly widespread thanks to its ability to detect possible damage with minimal invasiveness. However, the processing of the data acquired by the monitoring system and the identification of the associated damage sensitive features are the most challenging tasks due to the variability of environmental conditions, noise on measures and uncertainties in extracted modal parameters. In this context, Machine Learning (ML) techniques showed promising results for their capability of feature discrimination even in presence of noise-corrupted data. Furthermore, damage identification with ML approaches results as a quick process, since the calibration of a physical model describing the structural behaviour is not directly involved.

The paper presents the application of a damage detection procedure based on neural networks to a railway bridge. The output of a network, that depends on the dynamic features given as input, allows to classify the structure state as undamaged, lightly damaged or severely damaged. The procedure employs only simulated data but includes a series of expedients to approach a real situation, like the stochastic modelling of measurement errors and the use of two different models to account for the model error. This last represents the residual discrepancy between model and reality that exists even at the end of the calibration. Moreover, the use of different data and feature extraction techniques is investigated. For each type of data and extraction technique, a specific network is considered. The performances of the networks are analysed with respect to datasets generated by the two different models, assessing their effective applicability. The results show the relevance of accounting for model error in the development of the damage detection procedure.