Application of Machine Learning Techniques for the Analysis of National Bridge Inventory and Bridge Element Data

Abstract

The MAP-21 Act requires information on bridge assets to be at the element level for management operations in the U.S.A. This approach has the objective of improving future predictions of the performance of bridge assets for a more precise evaluation of condition and correct allocation of management funds to keep bridges in a good state of repair. Although bridge element conditions were introduced in the 1990s, the application of such data had never been mandatory for bridge asset management until 2014, therefore, the amount of historical data on bridge element (BE) condition is still limited. On the other hand, National Bridge Inventory (NBI) ratings have been collected since the 1970s and a wide range of data are available. Therefore, it is natural to ask whether BE condition can be predicted using NBI data. In the past, researchers statistically related BE and NBI data, but little has been done to revert NBI to BE. This paper addresses both challenges of mapping BE–NBI condition data using several machine learning techniques. The results of the analysis of these techniques applied to a sample of about 9,000 bridges from northeastern states of the U.S.A. shows that between 79.8% and 100% of the NBI ratings for deck, superstructure, and substructure can be predicted within a rating error of ± 1. The back-mapping operation of NBI time-dependent ratings to BE deterioration profiles for deck, superstructure, and substructure can also be predicted accurately with a probability greater than 50% at the 95% confidence level.