Bridge damage identification using unsupervised deep learning through variational autoencoders

Abstract

While supervised structural health monitoring (SHM) systems can potentially accurately detect and localize damage in bridge structures, they require labeled datasets containing healthy and damaged states to develop decision-making models. Obtaining supervised data from in situ bridges is often impractical. Unsupervised algorithms address this challenge by leveraging baseline (i.e., “healthy”) data, thereby eliminating the need for damaged datasets to detect anomalies. Conventional SHM techniques rely on sensor data to assess the structural condition; however, accurately detecting damage remains a challenge, especially when dealing with differing structure types and sizes and varying loading conditions. This study introduces an unsupervised damage detection methodology that aims to classify bridge health using structural response data under variable loading by leveraging deep learning. A variational autoencoder (VAE) network was employed to develop a data-driven methodology to identify structural anomalies based on response data collected during tests of an in situ bridge. Collected datasets included strain time histories representing three structural states (i.e., “healthy,” two levels of damage). The model was trained using strain data collected from healthy bridge loading tests and validated on separate healthy trials, while testing included both healthy and damaged scenarios to evaluate the method’s generalization and damage sensitivity. A VAE network was trained to learn underlying patterns of healthy behavior under moving live load. The training helped develop a compressed strain time-history representation using a network encoder and accurately reconstructed that input from a lower-dimension space. Normalized data reconstruction errors were used to define a damage index that permitted quantitative assessment of structural deterioration. In addition, clustering each test dataset corresponding as a function of structural state helped quantify response deviation from the healthy state. Results demonstrated that the proposed approach effectively detected increasing damage levels, thereby distinguishing different structural states from one another. Results highlighted the potential robustness and adaptability of the proposed approach to real-world SHM applications.