ABSTRACT
Recent studies have demonstrated that important bridge dynamic properties can be extracted from crowdsourced smartphone data collected during vehicle trips. The key benefit of crowdsourced smartphone vehicle trip (SVT) data is the potential ease in producing large volumes of useful data at very low costs compared to modern SHM networks. Crowdsourcing inherently involves trade-offs in data quality for data scale; however, the extent to which data quality varies with respect to trip metadata, such as vehicle speed or smartphone model, etc., are not well understood. This paper evaluates the sensitivities of bridge dynamic property estimates with respect to individual trips or groups of trips, i.e., subsets, based on about 700 SVT datasets collected on a real bridge. Overall, this paper demonstrates an application of classification models in identifying “low quality” datasets, i.e., those that corresponded to less accurate mode shape estimates. This tool shows potential to quickly flag low-quality datasets during preprocessing which could help optimize the accuracy of the SHM information produced.
