An Optimum Neural Network Model for Road Damage Classification

This work focuses on developing a simple artificial neural network (ANN) model for automatic classification of road anomalies. The model is particularly important to facilitate the development of the road condition monitoring system. We limit the discussion to the system that utilizes vehicle vibration data to predict the associated road condition. The vibration data are obtained from a test vehicle moving through four road types: a road in a good condition, a road containing a pothole, a road containing a speed bump, and a road containing an expansion joint. The data are then used to extract the vehicle maximum accelerations in its three directions: longitudinal, lateral, and vertical. This study reports the aspects of the size of the training data, the effects of the number of the neurons in the hidden layer, and the level of the achievable classification accuracy. The results are of the following: the current ANN model is able to predict the road condition with the level of accuracy of more than 80\%; the model using three neurons in the hidden layer is the most optimum; and the training data size of more than xxx data is required to produce the most accurate model.

Fergyanto E Gunawan Dr Eng and Yudi Purnama