Ensemble Technique for Machine Learning with Application to Monitoring of Heritage Structures

Abstract

Recent studies have demonstrated the effectiveness of machine learning techniques in the context of Structural Health Monitoring (SHM), where they can be applied to distinguish operational and environmental changes of dynamic features from those related to the evolution of damage. For instance, the combination of these techniques with the cointegration, a theory usually employed in econometric studies, has led to promising results, even in the detection of damage in complex monumental buildings. Several algorithms, including Support Vector Machine and Relevance Vector Machine, can be used for implementation of the multivariate regression required in the method. The choice of the algorithm to apply and the parameters to set can drastically influence the results and can lead to a wrong perception of the structural health. This paper proposes a combinatorial selection of machine learning algorithms and their settings to define the most performing among several optimal results. The ranking problem is solved by using the Plackett–Luce (PL) model-based strategy. The final aim is to obtain the damage indicator as indifferent as possible to harmless environmental and operational variations but still sensitive to structural changes, suitable to investigate the dynamic response of a structure. Data recorded by the dynamic monitoring system installed on the Sanctuary of Vicoforte, which contains the largest masonry oval dome in the world, and a calibrated finite elements model of this structure are used to simulate a damaged condition, in order to demonstrate the proposed strategy for SHM.