Vibration-based building health monitoring using spatio-temporal learning model

Abstract

Vibration-based building health monitoring is a promising and feasible approach to assess the operational state of building structures in a remote, automated, and continuous fashion; however, efficiently handling high-dimensional vibration signals from multiple sensors and effectively coping with missing/noisy data represent two main technical challenges. In order to overcome these issues, this study proposes a novel, reliable and robust framework, abbreviated CLG-BHM, based on a hybrid deep learning architecture. First, the framework uses a 1D convolutional neural network layer to learn low-dimensional representation vectors of long sensor signals, which preserve underlying structures’ dynamic characteristics. Second, temporal relationships within data are distilled via a Long-Short Term Memory layer. Third, the representation vectors of sensors are aggregated with those of their neighbors in a principled way via a graph attention network layer, resulting in a new latent representation rich in both temporal and spatial information. Finally, the latter is gone through a fully-connected layer to provide damage detection results. The performance and viability of the present method are evidenced via various examples involving a simple lumped mass structure, a semi-rigid steel frame, and an experimental 4-story structure from the literature. Moreover, a robustness study is performed, showing that the method can provide reasonable results with the presence of noisy and missing data.