Reliability of probabilistic numerical data for training machine learning algorithms to detect damage 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 conditions during a certain period of time. However, the scarcity of information regarding the structural response under seasonal environmental variations and less frequent operational conditions makes the distinction between these undamaged states and damaged ones very challenging and may cause damage detection algorithms to yield false indications. To overcome this limitation, hybrid approaches for the training of machine learning algorithms have recently been proposed. Rather than relying exclusively on monitoring data, hybrid approaches use finite element models of the structure to generate numerical data for less frequent undamaged scenarios. The numerical data are used for the training of machine learning algorithms together with the monitoring data. This paper addresses the reliability of numerical data for the training of machine learning algorithms by quantifying the damage detection performance of an algorithm trained with numerical data only. Monitoring data are used only for the initial calibration of the finite element model, which does not need to be exceedingly detailed, as the probabilistic variation of the uncertain parameters is considered. The damage detection performance is quantified both in terms of quality (number of ill-classified observations) and robustness to sub-optimal choices of the training data and algorithmic parameters. A general procedure for the generation of model-based data for the training of machine learning algorithms to detect damage is given and validated using the well-known Z-24 Bridge benchmark.