Neural networks estimation of truck static weights by fusing weight-in-motion data

This paper presents a complete neural networks approach developed to improve the accuracy of multiple-sensor Weigh-In-Motion system. This system consists in fusing the dynamical measurements of individual sensors installed in the road, into one improved estimate of static gross or axle weight. This task is complex due to the difficulty to inverse the model that describes the dynamical vehicle-pavement interactions. The sensors are also difficult to calibrate and remain sensitive to the environmental conditions. We chose to model the data with neural networks named 'general feedforward neural networks'. This class of neural nets includes in particular the famous Multilayer Perceptron model used in many applications. In order to increase the accuracy and the generalization capacity of this neural model, we also use an automatic method of model selection based on genetic algorithms. Simulated traffic data, computed with a realistic model based on vehicle-pavement interaction, indicated that a neural network can produce results with a higher degree of accuracy than the simple linear regression method. Moreover, our results show that the fusion of several sensors significantly improved the estimation of the static weight, much more in the frame of non linear modeling than in the frame of linear one. In our simulation, the optimal number of sensors to fuse stands around 7