Joseph N. Tack, Doctoral Student (jtack@eng.utoledo.edu)

and

Eddie Chou

University of Toledo

Department of Civil Engineering

Toledo, OH 43606

An artificial neural network (ANN) is applied to the problem of predicting the condition of pavement in subsequent years. The ability to accurately model how a pavement deteriorates is of utmost importance to many pavement management decisions regarding maintenance, including type, timing, and cost of repair.

Pavement deterioration is a function of three groups of variables that have complex interactions. The three groups of variables may be summarized as climate, materials, and traffic. Five years of pavement condition data was obtained from the Ohio Department of Transportation for this project. The data included pavement distress information, annual traffic volumes, roughness, pavement type, and location.

Neural Networks are typically used for the purposes of classification, prediction, and pattern recognition. They are often labeled black boxes, because there needs to be no assumption a priori about the relationships among variables. Essentially, a number of input variables are sent through an ANN, which produces a number of outputs that correlate to the desired results. The ANN generates its outputs by using layers of interconnected neurons. Each neuron is connected to every neuron in both previous and post layers. Signals are transmitted from one neuron to another according to the weight of the connection. The incoming signals to a neuron determine the magnitude of the signal it generates. Therefore the strength of each signal sent or received by a neuron is a function of the connection weight and sending neuron’s signal magnitude.

The neural network is trained using an error back-propagation algorithm. Multiple configurations of the number of layers and the number of neurons per layer were tested to check for the best possible convergence. The available data was randomly divided into three groups: training set, validation set, and test set. The training data is used to adjust the weights between neurons in the neural network. The validation set is used to monitor the convergence of the training, and ensure that the neural net can generalize for data which it was not trained on. The test set is used to test the neural network’s ability to predict pavement performance on new data. One half of the data was used for the training set and one quarter was used for each of the validation and test sets.

The final neural network model produces satisfactory results for predictions. There does exist the possibility of even better results being achieved once more data becomes available. Ten additional years of condition data is currently being acquired, as is more data in regards to maintenance practices, weather, and materials. It is believed that the addition of this new data is going to enhance the neural network’s ability to generalize, and therefore increase its accuracy of prediction.