Unsupervised deep learning method for bridge condition assessment based on intra-and inter-class probabilistic correlations of quasi-static responses

Abstract

Data-driven methods for structural condition assessment have been extensively investigated using deep learning (DL). However, studies on quasi-static response data-based structural health diagnoses are relatively insufficient. The difficulty is that quasi-static response data contain coupled effects of structural parameters and external loads. Considering that the correlation between quasi-static responses subjected to identical external loads is only a function of structural parameters and independent from the external loads, the correlation can therefore be employed as an indicator of the structural condition. This study proposes a condition assessment approach for cable-stayed bridges based on correlation modeling between the deflection of girders and tension in cables. The correlation is modeled by an unsupervised DL network comprising two variational autoencoders (AE) and two generative adversarial networks (GANs). The input and output are marginal probability density functions (PDFs). The DL network is trained as the reconstruction and translation processes to model the intra-class and inter-class correlations. Assumptions of shared latent space and cycle consistency are taken to ensure mutual modeling capacity. The Wasserstein distance between the predicted and ground-truth PDFs of tension in cables is used as an indicator of the structural condition. Using probabilistic correlation of quasi-static responses only requires the PDF of external loads to be identical and does not need the external loads to be precisely identical at any moment, thus relieving time-synchronization restrictions for different sensors. The results show that the predicted PDFs agree well with the ground-truth values under normal conditions. Furthermore, the Wasserstein distance is sensitive to damage and shows noticeable variations when the damage of the stay cable occurs.

Typical advances of artificial intelligence in civil engineering

“…Except for the temporal correlation modeling, probability distribution variation informed condition assessment has also been investigated to discover the probabilistic correlation pattern embedded in monitoring data of different structural responses. A DL network comprising two variational autoencoders (VAEs, E 1 - G 1 and E 2 - G 2 ) and two GANs ( G 1 - D 1 and G 2 - D 2 ) was established to model the probabilistic correlations of quasi-static responses of bridges (Xu et al, 2022), as shown in Figure 12. VAEs were designed to model intra-class correlations among either GVDs or CTs, and GANs were designed to model inter-class correlations between GVDs and CTs.…”

“…Weight sharing (shown as red blocks) was further implemented to ensure that the shared latent space could connect two VAEs. In addition, if the model functions well, PDFs of GVD and CT were supposed to be translated in a cycled manner, which would obey a cycle-consistency constraint (shown as blue arrows).

Figure 12.Schematic of intra-class and inter-class probabilistic correlation modeling between two response groups (Xu et al, 2022).

…”

“…Figure 13 shows that the predicted PDFs matched the ground-truth ones well and that specific modes of bridge condition change induced by SHM system upgrade, cable damage, sensor malfunction, and traffic jam could be successfully identified. The results showed that the Wasserstein distance was very sensitive to damage and presented noticeable variations when the damage of the stay cable occurred.

Figure 13.Representative results of probabilistic correlation informed bridge condition assessment (recognized patterns of bridge condition in (c) from top left to down right: system upgrade, cable damage, sensor malfunction, and traffic jam) (Xu et al, 2022).

…”

“…Figure12. Schematic of intra-class and inter-class probabilistic correlation modeling between two response groups(Xu et al, 2022).…”

“…Figure13. Representative results of probabilistic correlation informed bridge condition assessment (recognized patterns of bridge condition in (c) from top left to down right: system upgrade, cable damage, sensor malfunction, and traffic jam)(Xu et al, 2022).…”