Weighing trucks in motion using Gaussian-based neural networks

The authors describe the application of neural networks to the problem of weighing trucks in motion using strain spectra measured on beams supporting a highway bridge. The learning processes of both a sigmoidal network with the generalized delta rule and a Gaussian-based network with its own training procedure are evaluated. This application requires 95 input neurons, 3 output neurons, and 2304 training patterns. The Gaussian-based network exhibits a much faster rate of convergence than that of the sigmoidal network and achieves a much higher degree of accuracy. Both networks are tested on 1000 random patterns not used during training. The Gaussian-based network shows a significantly superior performance. Overall, the Gaussian-based approach demonstrates the feasibility of using neural networks to determine track axle loads from strain data