Modeling deterioration of concrete bridge decks using neural networks

Markovian deterioration model is the most common method applied by the bridge management systems (BMSs) in the States. However, the basic assumptions of Markovian deterioration model have been questioned due to its reliability in modeling bridges' deterioration.; This research is aimed to prove the two underlying assumptions, the age independency and the maintenance-history independency, of Markovian deterioration model is not consistent with bridges' behaviors. And this research is also aimed to develop an accurate and reliable application model for predicting the concrete decks' deteriorations.; Throughout researching all of the historical data recorded from the concrete bridge decks' inspections and maintenances allover the State of Wisconsin, it was found that the age and maintenance history are two important factors influencing decks' deterioration. Besides these two factors, the other eight more factors were identified significant through the analyzing of the bridges' inventory data recorded in the State of Wisconsin. These eight factors include district, deck's length, deck's area, average daily traffic, environment, degrees of skew, and number of spans. In researching the accurate and reliable deterioration model, Neural Network (NN) was implemented, and two classifiers, Back-Propagation (BP) approach Multi-Layer Perceptron (MLP) classifier and K-Nearest Neighbor (KNN) classifier, were applied and evaluated. With all the factors identified significant as the inputs, the model with the KNN classifier could significantly reach a better classification level than the one with BP approach MLP does. Therefore, the KNN classifier was selected for executing the deterioration model.; As the conclusion of this research, the strong evidences have shown that the age and maintenance history are two important factors for deck's deterioration. Consequently, the Markovian deterioration model was proved to be not consistent with observed deterioration processes through this research. As an achievement for creating the better modeling, the NN prediction model has been demonstrated with better performance than what Markovian deterioration model has stated. With the very positive research result, the new deterioration model developed in this research is not only to yield a strong classification level from the project-level point of view, but also outstandingly to provide the better network-level information than Markovian determination model does.