Damage Prediction on Bridge Decks considering Environmental Effects with the Application of Deep Neural Networks

Abstract

Due to limited budgets and professional manpower, predicting possible damage in advance is essential for supporting on-site bridge inspections. This study aims to predict the severity of damage on bridge decks considering the effects of traffic and weather. First, the authors obtained identification, structural, environmental, and inspection data of pre-stressed concrete I-type bridges from the Korean Bridge Management System, and the final dataset of 16,728 tuples and 53 variables was prepared. Next, correlation analysis was performed to remove redundant variables, and random forest identified important factors that caused the more serious condition of damage to the deck. A total of 32 variables were finally used to develop Deep Neural Networks to predict different types of deck damage. The developed model successfully predicted the occurrences of seven different types of damage to bridge decks, that is, linear cracking, map cracking, scaling, breakage, leakage, efflorescence, and corrosion of exposed rebar, with the average weighted F1 score of 91%. Environmental effects on prediction were also determined; for example, traffic, temperature, and precipitation increased the F1 score of linear cracking by 4%. This research was a pioneering attempt to develop a model that enables specific damage-level prediction using both statistics and artificial intelligence techniques.