Automated Construction of Bridge Condition Inventory Using Natural Language Processing and Historical Inspection Reports

The aging bridge infrastructure network is in critical need of maintenance, rehabilitation or replacement (MR&R) as nearly half of this inventory is approaching the end their design service lives. Agencies responsible for managing this network have limited resources that are insufficient for the scale of the problem, highlighting the need for smart, system-level decision-making strategies that can be integrated with current practice. A large amount of rich information on element-level condition descriptions are buried in bridge inspection reports, but this local information is seldom used holistically to infer system performance. Current decision-making strategies are constrained by limitations in bridge deterioration prediction models, which lack comprehensive and well-structured databases needed for automation of processes associated with high resolution forecasting. How to draw meaningful information from the details of these localized reports to assist system-level bridge condition comparison and maintenance prioritization still remains unclear and warrants further study. To bridge this gap, this paper proposes a Natural Language Processing framework to extract information from the raw textual data in bridge inspection reports. This raw data provides a source for capturing the experience-driven metric inherent to the bridge inspection process. The proposed framework constructs an innovative bi-directional Long-short Term Memory neural network that automatically reads inspection reports into different condition categories and achieves 96.2% accuracy when examined on inspection reports collected by Virginia Department of Transportation. The extracted information forms a well-structured bridge condition inventory that contains rich historical and local condition information, and hence enables smart, system-level bridge MR&R decision-making.