Reconstruction of road defects and road roughness classification using vehicle responses with artificial neural networks simulation

Abstract

The road damage assessment methodology in this paper utilizes an artificial neural network that reconstructs road surface profiles from measured vehicle accelerations.The paper numerically demonstrates the capabilities of such a methodology in the presence of noise, changing vehicle mass, changing vehicle speeds and road defects. In order to avoid crowding out understanding of the methodology, a simple linear pitch plane model is employed. Initially, road profiles from known roughness classes were applied to a physical model to calculate vehicle responses. The calculated responses and road profiles were used to train an artificial neural network. In this way, the network renders corresponding road profiles on the availability of fresh data on model responses.The results show that the road profiles and associated defects can be reconstructed to within a 20% error at a minimum correlation value of 94%.

Reconstruction of road defects and road roughness classification using Artificial Neural Networks simulation and vehicle dynamic responses: Application to experimental data

“…The architecture of the static feed-forward neural network employed in this study is shown in Figure 5. This is a departure from the dynamic layer-recurrent NARX used in the previous paper [2]. The reason for the change is that the NARX network requires feedforward from true road profiles which unfortunately may not be available in real test situations.…”

“…In the previous study [2], half-sine waves were used to represent road bumps and random road profiles were generated in a purely mathematical way using the wellknown one-dimensional random profile function [2,13]. In this study, two kinds of road profiles are used: trapezoidal-shaped bump profiles and Belgian paving.…”

“…In the previous paper [2], an eight degree-of-freedom (8-DOF) linear pitch-plane model with four degrees of freedom representing vehicle motions and additional four degrees for seat and driver vibrations, was employed for the investigation. Eight road roughness classes having well-known displacement spectral densities (DSDs) on the International Organization for Standardization power spectral density (ISO PSD) classification of road roughness [4] were applied to the model to calculate corresponding sprung mass and axle accelerations.…”

“…Eight road roughness classes having well-known displacement spectral densities (DSDs) on the International Organization for Standardization power spectral density (ISO PSD) classification of road roughness [4] were applied to the model to calculate corresponding sprung mass and axle accelerations. The accelerations, serving as inputs, and the corresponding road profiles, serving as targets, were applied to a Non-linear Auto-Regressive with eXogenous inputs (NARX) network having 20 hidden neurons and one output neuron [2]. 4 The vehicle model was subsequently simulated with different sets of road profiles which were generated by a random road surface function to calculate corresponding sprung mass and axle accelerations.…”

“…The simulated road profiles were compared with the actual road profiles. These evaluations were performed for different road roughness conditions (based on the ISO PSD roughness scale [4]), different vehicle speeds, growing road defects, noisy conditions, and different vehicle payload conditions [2]. It was noted that the methodology was able to reconstruct the road profiles to within an error margin of 20 % with a minimum correlation of 94 % [2].…”