Eliminating thermal responses of stay cables by partially missing sensors toward more reliable and robust SHM

Abstract

Abstract In structural health monitoring (SHM), data normalization is important since the features sensitive to potential damages may be buried by the operational and environmental effects. Although cables are critical components in cable-stayed bridges, their temperature effects have not been sufficiently studied. To bridge this gap, this study proposed a framework to capture and eliminate the thermal response (TR) of cable forces to enhance the reliability of SHM. Specifically, empirical wavelet transform was used to approximate the TR, and gated recurrent unit was used to learn to infer the TR based on field monitored temperature field (TF) data. In addition, a recovery mechanism called generative adversarial imputation nets (GAIN) was introduced to recover the TF from partial sensors to enhance the robustness of this method. Four cables of a bridge, covering long or short, side span or midspan, are investigated. Results show that the developed model can well capture the TR, which contributes to 63%–93% of the total variation of cable forces. By eliminating the TR, the entropy can be reduced by 19%–43% and obtain 0.7–1.4 bits. It is also shown that the introduction of GAIN makes the sensing of TF robust to sensor missing, even when the missing ratio reaches 28.6%, and thus makes the TR inference robust. It is expected that this study can bring more reliable and robust SHM for critical infrastructures.