Hybrid Approach for Supervised Machine Learning Algorithms to Identify Damage in Bridges

In bridges, monitoring data usually correspond to normal operational and environmental conditions, resulting in a lack of damage-related data. For this reason, machine learning algorithms for damage detection are typically unsupervised. Conversely, numerical models are employable as surrogates for extreme events rarely encountered during the existence of a bridge and for common damage scenarios, enabling the training of supervised machine learning algorithms. In this paper, hybrid data obtained by integrating monitoring and numerical observations from the Z-24 Bridge benchmark are used for training supervised machine learning algorithms to classify damage. Special attention is dedicated to the numerical model that, while being simple enough to be used on thousands of runs, must capture the complex nonlinear behavior that typifies damaged conditions. A model updating technique is used for the preliminary calibration of the finite-element model. Numerical data are generated in a probabilistic manner starting from the initial finite-element model, assuming the Gaussian distribution of the uncertain parameters. Three undamaged scenarios and three damaged scenarios are modeled. Subsequently, several supervised learning classifiers are trained with the same hybrid database and a comparative summary of their effectiveness is presented. The paper introduces a novel soft classification method that accounts for the overlapping of observations belonging to different structural conditions in the feature space. This study proves that data generated from probabilistic-based finite-element models can be used for structural health monitoring and damage identification in the context of bridges, therefore providing a hybrid supervised approach that can be easily applied in practice.

Because bridges are one-of-a-kind, expensive structures, considerable efforts are made to ensure they remain in service for as long as possible. Consequently, monitoring systems are being installed on more and more bridges. However, monitoring data are usually related to healthy states of the structure rather than damage occurrences. Therefore, digital replicas can be employed to simulate damage scenarios that may occur in the lifespan of a bridge. Hybrid data can be obtained by blending monitoring data with simulated data from the same bridge. This paper illustrates the construction of such a hybrid database, which covers a wide range of possible scenarios for a well-known bridge benchmark. This hybrid database is then used to train machine learning algorithms to distinguish between various changes in the dynamic behavior of the bridge as a way to evaluate the condition of the structure and identify damage. In practice, the hybrid approach can be implemented on monitored bridges to raise flags when flaws are detected, to localize and evaluate the severity of damage, and even to make a prognosis on the remaining life of the bridge.