Existing bridge management systems have adopted Markov chain models to predict the condition of different bridge components for network-level analysis. These models assume that the future condition depends only on the present condition and not on the past condition (i.e., state independence). Moreover, these models do not explicitly account for the effect of governing deterioration parameters, such as average daily traffic, percentage of trucks, and environmental impacts, on the predicted condition. Machine learning approaches have been proposed by many researchers as successful tools for modeling infrastructure deterioration. A work in progress evaluates the prediction accuracy of decision tree algorithms--one of the most common techniques of machine learning--against the prediction accuracy of Markov chain models. Field data of concrete bridge decks were obtained from the Ministere des Transports du Quebec, Canada, database to develop and evaluate the performance of decision tree algorithms in modeling bridge deck deterioration. Evaluation results have indicated a slight increase in the performance of decision trees over existing Markov chain models when the past condition is considered or governing deterioration parameters are incorporated.
Abstract