Bridge Infrastructure Management System: Autoencoder Approach for Predicting Bridge Condition Ratings

Abstract

Being an essential component of our economy, bridges play a vital role in facilitating transportation of people and goods all over the world. Bridge infrastructure systems help in developing an effective way to monitor and protect structures in all aspects. However, the bridges are exposed to various kinds of damages due to aging, heavy load of traffic, quality of construction, and so on. Hence, determining the condition of such bridges through a proper infrastructure management system is important for understanding the potential loss in the longevity of the structures. This research paper presents the development and evaluation of an autoencoder-random forest (AE-RF) model to predict the condition rating of bridges using the National Bridge Inventory (NBI) database. To demonstrate the proposed model, a case study using bridges present in the US state of Florida has been performed using historical NBI data (2011–2020). Through this research, it was identified that when one of the deep learning models, named autoencoder, is combined with random forest (RF), it results in an efficient model for determining the condition ratings of the bridge components with fewer input parameters. The developed model was about 90% accurate in predicting the bridge’s deck condition by using other rating values as input variables and about 79% accurate without the use of any other rating factors, thereby addressing the existing research gap in determining the condition rating without the use of historic condition rating. On the other hand, the model was 78% and 77% accurate in determining the condition ratings for superstructure and substructure without using historic ratings and evaluation parameters. Hence, the proposed model would be more reliable in the evaluation of condition of existing bridges across the nation.