Fully convolutional networks for structural health monitoring through\n multivariate time series classification

Abstract

We propose a novel approach to Structural Health Monitoring (SHM), aiming at\nthe automatic identification of damage-sensitive features from data acquired\nthrough pervasive sensor systems. Damage detection and localization are\nformulated as classification problems, and tackled through Fully Convolutional\nNetworks (FCNs). A supervised training of the proposed network architecture is\nperformed on data extracted from numerical simulations of a physics-based model\n(playing the role of digital twin of the structure to be monitored) accounting\nfor different damage scenarios. By relying on this simplified model of the\nstructure, several load conditions are considered during the training phase of\nthe FCN, whose architecture has been designed to deal with time series of\ndifferent length. The training of the neural network is done before the\nmonitoring system starts operating, thus enabling a real time damage\nclassification. The numerical performances of the proposed strategy are\nassessed on a numerical benchmark case consisting of an eight-story shear\nbuilding subjected to two load types, one of which modeling random vibrations\ndue to low-energy seismicity. Measurement noise has been added to the responses\nof the structure to mimic the outputs of a real monitoring system. Extremely\ngood classification capacities are shown: among the nine possible alternatives\n(represented by the healthy state and by a damage at any floor), damage is\ncorrectly classified in up to 95% of cases, thus showing the strong potential\nof the proposed approach in view of the application to real-life cases.\n