Deterioration models for prediction of remaining useful life of timber and concrete bridges: A review

Bridge deterioration models are used for prioritization and maintenance of bridges. These models can be broadly classified as deterministic and stochastic models. There are mechanistic models (or physical models) as well as Artificial Intelligence (AI)-based models, each of which can be stochastic or deterministic in nature. Even though there are several existing deterioration models, state-based stochastic Markov chain-based model is widely employed in bridge management programs. This paper presents a critical review of different bridge deterioration models highlighting the advantages and limitations of each model. The models are applied to some case studies of timber superstructure and concrete bridge decks. Examples are illustrated for arriving at bridge deterioration models using deterministic, stochastic and Artificial Neural Network (ANN)-based models based on National Bridge Inventory (NBI) data. The first example is based on deterministic model and the second on stochastic model. The deterministic model uses the NBI records for the years 1992–2012, while the stochastic model uses the NBI records for one year (2011–2012). The stochastic model is state-based Markov chain model developed using Transition Probability Matrix (TPM) obtained by Percentage Prediction Method (PPM). The two deterioration models (i.e., deterministic and stochastic models) are applied to timber highway bridge superstructure using NBI condition data for bridges in Florida, Georgia, South Carolina and North Carolina. The illustrated examples show that the deterministic model provides higher accuracy in the predicted condition value than the stochastic Markov chain-based model. If the model is developed based on average of transition probabilities considering the data for the period 1992 to 2012, the prediction accuracy of stochastic model will improve. Proper data filtering of condition records aids in improving the accuracy of the deterministic models. The third example illustrates the ANN-based deterioration model for reinforced concrete bridge decks in Florida based on the NBI condition data for the years 1992–2012. The training set accuracy and testing set accuracy in the ANN model are found to be 91% and 88% respectively. The trained model is utilized to generate missing condition data to fill the gaps due to irregular inspections of concrete bridges. This paper also discusses scope for future research on bridge deterioration modeling.