Deck, Superstructure, and Substructure Deterioration Prediction for Bridges Using Deep Artificial Neural Networks

Around 13.1% or 3,195 bridges in Missouri are structurally deficient. In addition, 4,800 bridges need repairs with an estimated total cost of 4.2 billion. Assessment of the deterioration conditions of bridges is therefore necessary to develop and optimize future rehabilitation plans for bridges and to properly allocate the available funds. Researchers have used several methods to develop models that estimate and predict the deterioration conditions of bridges, including statistical techniques, Markov chains, neural networks, and fuzzy logic. Nevertheless, most of the existing studies focused on the prediction of a single bridge element. To this end, there exists a need for research that assesses the conditions of different structural elements and that compares between different predictions models. As such, the goal of this research is to investigate the use of deep artificial neural networks (DANN) and regression models for bridge deterioration prediction. Accordingly, a DANN model is developed using various numbers of hidden layers and neurons. In parallel, a linear regression (LR) model is developed as to act as a baseline. Data on long span bridges in the state of Missouri is used to train, test, and fit the models. Finally, the DANN and LR models are validated and compared based on the obtained mean squared errors. It was found that both models have generally similar accuracy with that of DANN being slightly higher. In addition, it was concluded that the accuracy of the DANN is highly dependent on its configuration. The outcomes of this research is models that can be used to predict the deterioration conditions of bridges in the state of Missouri and to assist in the development of rehabilitation plans. The developed models can be modified to accustom bridges in other states and to attempt various configurations of the DANN.