Graph Convolutional Neural Networks Based Strain Estimation

Abstract

Infrastructural owners have to manage and control a huge number of structures that are subject to many phenomena that can alter their performance. During their service life, infrastructure like bridges are exposed to cycles of dynamic loads, causing potential fatigue failure. For the life-cycle assessment aim, most procedures require the use of strain data. The acquisition of strain data is not easy, especially in critical positions, and the high costs would restrict the wide use of such procedures. This study proposes a technique based on the graph theory to obtain strain from acceleration data. As shown in neuroscience and biological fields, this type of architecture is able to achieve good outcomes not relying on handcrafted design parameters. Besides, their capacity of grasping spatio-temporal dependencies turns out to be a key point in the damage detection framework. A numerical case study involving a two-span beam has been investigated to explore the potentiality of this architecture for acceleration-strain conversion. Here the graph nodes represent the sensor locations and the combination of geometric and signals aided to create a spatial-temporal GCN architecture. The obtained outcomes are promising and make the framework extendable to real case studies.