HYBRID SUPERVISED MACHINE LEARNING APPROACH FOR DAMAGE IDENTIFICATION IN BRIDGES

Structural health monitoring (SHM) of bridges often involves machine learning algorithms, trained based on two independent learning strategies, namely unsupervised and supervised learning, depending on the type of training data available. When unsupervised learning strategy is employed, the algorithms are normally trained with data gathered from monitoring systems, corresponding to normal operational and environmental conditions. The lack of information regarding the dynamic response of the structure under extreme environmental and operational conditions, as well as under damage scenarios, may lead to flaws in the damage detection process, namely the rise of false indications of damage. In order to overcome this drawback, finite element models can be used as structural proxies to generate data that correspond to scenarios unlikely to be recorded by the monitoring systems, such as extreme temperatures or structural damage. The use of both monitoring and numerical data in the framework of a hybrid approach greatly improves the quality of the training process, as recently shown by the authors. The hybrid approach also enables the use of the supervised learning strategy if numerical data corresponding to damage scenarios are available. Therefore, this paper assesses the reliability of a hybrid approach for the supervised training of machine learning algorithms using numerical data corresponding to extreme temperatures and several damage scenarios. The damage scenarios comprise various degrees of settlement of a bridge pier and a landslide near the same pier. Monitoring data are used for the testing of the algorithms and 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 taken into account. The procedure was applied to the Z-24 Bridge, a well-known benchmark consisting of one year of continuous monitoring and including progressive damage readings.