Abstract: Bridges are one of the most important transportation infrastructures as they guarantee the flow of people and goods, thus it is crucial for monitoring bridge safety. However, due to their intrinsic construction load as well as the extrinsic traffic load and environmental temperature, bridge deflection varies constantly. Moreover, the deflection outlier will cause a huge safety risk for bridges. The present detection methods for bridge deflection outlier still exist some limitations in the lack of synthetically considering both the extrinsic impact factors and the intrinsic variation features. Therefore, the paper proposes the detection method for bridge deflection outlier by fusing multi-sourced surveying data. It calculates and fuses the multi-sourced features based on temperature, bridge traffic load, and bridge deflection data. Besides, it utilizes the random forest model to judge whether the deflection is the outlier. The experimental results illustrate that the proposed method could get a good performance of the accuracy of 88.18%. In addition, the method's performance exceeds other classical machine learning models. In summary, the proposed method could help the bridge administrators detect the bridge deflection outlier to eliminate the safety risks, and further promote their maintenance and administration levels.

Key words: bridge deflection, outlier detection, data fusion, random forest