ABSTRACT
North Carolina was one of the original states to develop and implement a bridge management system (BMS) shortly after formal bridge inspections were mandated by federal legislation in the United States. However, the general form of the deterioration models currently in use by the North Carolina Department of Transportation to predict future changes in condition ratings of bridges has remained unchanged since original development. These models are deterministic in nature, piecewise-linear, are based on average condition rating durations calculated from historical bridge inspection records, and have relied on pre-classification to account for the effects of explanatory factors on deterioration rates. To overcome limitations in the predictive fidelity of these legacy models, a proportional hazards-based probabilistic deterioration model was developed and implemented on the historical database of over 35 years of deck, superstructure, and substructure condition ratings for the North Carolina state bridge inventory. These probabilistic models are multivariable, duration-based, and non-stationary with a non-homogeneous Markov chain prediction process. A simplified stationary approach was also developed for use within a conventional homogeneous Markov chain. This paper presents a comparative analysis of the predictive fidelity of updated deterministic models and the newly developed probabilistic models. In the absence of future response data, a methodology developed for a quantitative assessment of the predictive fidelity based on existing bridge records is described. The analysis includes over 6000 concrete deck records and a prediction period of 15 years. Probability distributions of prediction errors demonstrate that both the nonstationary and stationary implementations of the probabilistic models achieve a much higher degree of accuracy and precision in predicting bridge condition ratings than the deterministic models. Reasons for the improved predictive fidelity of these probabilistic models are discussed briefly while making a case for adoption of these models in place of deterministic models.
