Autoencoder-based method to assess bridge health monitoring data quality

The data quality determines the reliability of big data-based bridge condition assessments. However, rapidly discerning data conditions and identifying low-quality data segments pose considerable challenges. This study introduces a transformer-based autoencoder neural network for rapid data quality assessment in bridge health monitoring. The average Euclidean distance was used to quantify the dispersion of multiple hidden variables, and the overall quality of the multiple data fragments was quantitatively evaluated. Moreover, a method was introduced to calculate the adaptive thresholds based on the Euclidean distance. The application to the acceleration data of a cable-stayed bridge demonstrates that the proposed method can extract hidden variables from acceleration data segments of length 3000. The network achieves a high compression rate of 1/1500, and the extracted hidden variables retain pertinent information regarding data quality. The proposed approach is data-driven and exhibits significant advantages in efficiency, accuracy, and user-friendliness.