SHM under varying environmental conditions: an approach based on model order reduction and deep learning

Data-driven approaches to structural health monitoring (SHM) have been recently shown to be a powerful paradigm, helping to lead to an evolution of traditional scheduled-based maintenance methodologies towards condition-based ones. Nevertheless, only few of them provide monitoring scenarios accounting for the varying loading and environmental conditions, which can potentially lead to misinterpretations of the structural state. In this paper, we propose a damage localization strategy that efficiently exploits vibration and temperature data to account for the effects of temperature fluctuations on the structural response. By allowing for a finite number of a priori defined damage scenarios, deep learning techniques are used to handle the damage localization task as a supervised classification, conditioned on temperature data. The training dataset is generated through a parametrized thermo-mechanical model of the structure, under a prescribed variability of loading and thermal conditions. To relieve the computational burden associated to the data generation process, a parametric order reduction strategy is also exploited. Results relevant to two case studies, a cantilever beam and a portal frame, are adopted to testify the capability of the proposed procedure to locate the damage, also when characterized by a rather small reduction of the local stiffness properties.