Optimizing bridge modal property estimation via quality-driven crowdsourced smartphone data selection

Recently, crowdsourced smartphone-vehicle trip (SVT) data has enabled cost-effective estimation of bridge modal frequencies and absolute mode shapes. SVT data includes acceleration and GPS measurements collected from smartphones within vehicles as they cross the bridge. The SVT data is inherently contaminated mostly with sensor noise, vehicle dynamics, and road profile uncertainties. These factors cause variability in the amount of embedded bridge dynamic information, thereby affecting the overall data quality. This study presents a novel method and metric to quantify the SVT data’s quality based on each trip’s impact on the identified aggregated mode shape. Then, a data-driven model is used to detect the quality parameter automatically using the convolutional neural network. The model is trained and tested on over 900 asynchronous SVT data collected from smartphones over the Cadore viaduct bridge in Italy. The results demonstrated that this method could improve the quality of an identified mode shape, increasing the modal assurance criterion of 0.8 in blind aggregation to 0.97 with model sorting and eliminating low-quality trips. Ensuring the quality control of crowdsourced data is crucial due to multiple noise sources, and discarding erroneous datasets can significantly improve dynamic characterization identification of the bridge. • Proposing a quality metric for individual smart-phone acceleration data. • Quality metrics directly reflect each trip’s contribution to modal properties. • Methodology works successfully on the field collected data from real-life case studies. • Designing a machine learning model to predict quality score for individual trips. • Proposing computationally efficient solutions to handle large crowdsourced datasets.