A Non-Parametric Mixed Learning Technique for Mitigating Environmental Effects on Structural Modal Frequencies

Abstract

Health monitoring of civil structures via machine learning is a powerful approach to the early detection of any damage pattern. Besides structural damage, also environmental and operational variabilities are known to affect the inherent structural properties. Although the induced variations in the monitored properties are not harmful, their confounding influence can lead to economic and human losses. For these reasons, a novel unsupervised learning strategy is here proposed, aiming to properly account for the environmental effects on the structural modal frequencies. The offered solution is a non-parametric mixed learning strategy resting on hierarchical clustering, local nonnegative matrix factorization, and Mahalanobis-squared distance (MSD). By means of the hierarchical clustering, training data consisting of modal frequencies relevant to the undamaged condition are subdivided into local clusters, which are then exploited in order to get rid of the environmental effects. The reconstructed data are finally used to train a non-parametric novelty detector based on the MSD, to obtain scores for decision making regarding the current state. To validate the proposed method, a set of modal frequencies of a steel arch bridge in its long-term monitoring has been considered; results show that the proposed methodology is effective in taking aside the environmental variability from the time history of the collected modal frequencies of the structure.