Improving Reliability of Markovian-based Bridge Deterioration Model Using Artificial Neural Network

Bridge Management Systems (BMSs) as a Decision Support System (DSS), have been developed since the early 1990’s to reliably manage a bridge network. Forecasting long-term performance of bridge by deterioration model is a crucial component in a BMS. Markovian-based models are one of the most typical methods to predict long-term bridge performance. It has been used by a number of BMS software including the popularly used PONTIS, BRIDGIT and OBMS. The Markovian- based model is based on transition matrix obtained from overall condition rating of bridges in a network. The change in condition ratings with time provides typical deterioration rates, which can normally be determined from a non-linear regression analysis. Reliable regression analysis requires either large bridge network or sufficient historical condition ratings to obtain accurate transition probability for bridges. Markovian-based model prediction is a simple way to forecast long term performance of individual bridge. However, most bridge agencies do not have adequate condition rating records. This has become a major shortcoming in deterioration modelling.

In order to minimise the abovementioned problem, this paper presents modified Markovian method using previously developed Backward Prediction Model (BPM). Based on Artificial Neural Network (ANN) technique, the BPM is able to generate missing historical condition ratings thereby providing more historical trend of condition depreciation. In this study, BPM-generated condition ratings are used for regression analysis to obtain reliable transition probability required by the Markovian-based model. The results of the proposed study are compared with those of a typical Markovian-based model to identify the advantage of BPM and limitations for further development.