Enhancing bridge inspection data quality using machine learning

Bridge condition assessment is often compromised by errors in inspection data, limiting reliable maintenance and management decisions. This paper investigates how to enhance inspection data quality by automatically identifying and correcting the inaccurate assessment of structural conditions. A model that integrates textual and quantitative features is proposed to identify defect and condition ratings through defect descriptions, with corresponding dynamic partitioning strategy to detect ambiguous data, and a down-sampling and bagging ensemble to address class imbalance. Validated with ten years of real inspection data from 464 bridges, results show 98 % accuracy in correcting condition scores and 100 % accuracy in condition-level identification. These findings underscore the method's potential to improve the reliability of condition assessment and strengthen bridge management decision-making. Future research can focus on refining condition level identification algorithms for severely deteriorated structures.