On the reliability of finite element models for the training of machine learningalgorithms for damage detection in bridges

In Structural Health Monitoring of bridges, machine learning algorithms for damage detection are typically trained using an unsupervised learning strategy, with data gathered from monitoring systems, and assuming the structures are undamaged and functioning under normal operational state conditions during a certain period of time. However, the absence of information regarding the structural response under more extreme environmental and operational conditions, and under damage as well, makes the distinction between these states very challenging, and may cause damage detection algorithms to yield false-positive indications of damage. In the worst case, environmental and operational variability may hide actual damage, yielding false-negative indications. In order to overcome this limitation, the authors have recently proposed a hybrid approach for the training of machine learning algorithms. Rather than relying exclusively on measured data, the hybrid approach uses numerical models of the structure to generate data that corresponds to the undamaged scenarios that rarely occur in newly built bridges. This data is used for the training of the machine learning algorithms together with the monitoring data. While the hybrid approach was shown to successfully diagnose damage for a real bridge structure, the question of how good this performance would be in the absence of long-term monitoring data was left unexplored. Therefore, this paper pretends to assess the performance of a damage detection algorithm trained with numerical data only. Monitoring data is used only for the initial calibration of the numerical model, which does not need to be precise, since the probabilistic variation of uncertain parameters is taken into account.