Using Gaussian process machine learning to predict dynamic road wear of a rigid heavy vehicle

Abstract

The paved road network is a critical asset to any economy. South Africa has a paved road network that has an estimated value above R2 trillion. This asset is however under threat as there was a backlog in maintenance of more than R416.6 billion in 2018. Heavy vehicles are primarily responsible for road wear, and overloaded vehicles can cause more than 60% of road wear. Most road wear assessments use static axle loads that are assumed to be symmetrical on either side to calculate the road wear caused by a heavy vehicle. Previous work has shown that the effect of crossfall (CF) cannot be ignored when considering the dynamic road wear of heavy vehicles. This paper expands on previous work through the development of a novel Gaussian process machine learning (GPML) model that can predict the dynamic road wear of a rigid heavy vehicle given 15 input parameters. The road wear criteria considered are the first (1st) and fourth (4th) power aggregate forces on the left and right sides using the 95th and 99th percentile conditions. The results show that the model is very accurate and requires comparatively few inputs to train an accurate model. For interpolated results, the average absolute error is less than 1% and for extrapolated results, the average absolute error is less than 3%. The results also include the standard deviation associated with the result which is important for future research to minimise training examples. Using machine learning models to predict dynamic road wear allows for rapid calculation and testing and also does not require expensive multibody dynamics software tools to calculate. This would be very advantageous to the industry, especially when developed for the Smart Truck Pilot Project.