Abstract:
This paper reports the performance of an Artificial Neural Network based road condition monitoring methodology on measured data
obtained from a Land Rover Defender 110 which was driven over discrete obstacles and Belgian paving. In a previous study it was
demonstrated, using data calculated from a numerical model, that the neural network was able to reconstruct road profiles and their
associated defects within good levels of fitting accuracy and correlation. A nonlinear autoregressive network with exogenous inputs
was trained in a series–parallel framework. When compared to the parallel framework, the series–parallel framework offered the advantage
of fast training but had a shortcoming in that it required feed-forward of true road profiles. In this study, the true profiles are not
available and the test data are obtained from field measurements. Training data are numerically generated by making minor adjustments
to the real measured profiles and applying them to a full vehicle model of the Land Rover. This is done to avoid using the same road
profile and acceleration data for training and testing or validating the neural network. A static feed-forward neural network is trained
and consequently tested on the real measured data. The results show very good correlations over both the discrete obstacles and the
Belgian paving. The random nature of the Belgian paving necessitated correlations to be made using their displacement spectral densities
as well as evaluations of RMS error percent values of the raw road profiles. The use of displacement spectral densities is considered to be
of much more practical value than the road profiles since they can easily be interpreted into road roughness measures by plotting them
over an internationally recognized standard roughness scale.
