A two-stage temperature-driven method for detection of fault sensors and abnormal stress using structural health monitoring data

Abstract

Steel box girders’ stress is a crucial monitoring item for large-span bridges, as abnormal stress can indicate performance degradation or damage. Establishing the correlation model between stress and temperature is an effective method for detecting abnormal stress, given that stress and ambient temperature are highly correlated under normal operational conditions. In this study, a two-stage temperature-driven method was used to identify sensor failures and abnormal stress. In the first stage, a convolutional neural network (CNN) was used to learn features from Gramian Angular Field images, which were converted from the temperature-induced stress time series. Then, the well-trained CNN can automatically classify monitoring data, thereby extracting high-quality data used for subsequent correlation model training. Based on the classification results, channels that deviate from the collective behavior of the sensor network can be attributed to sensor failures or severe local damage. In the second stage, a Gaussian process (GP) model was used to model the correlation between the stress and the temperature. The performance of the GP method in correlation modelling was quantitatively evaluated by the residuals between the GP model predictions and test data. Finally, an early warning framework based on the residuals was proposed. The warning method effectively identified the simulated abnormal stress of varying severity, demonstrating the effectiveness of the proposed method in detecting minor structural damage.