Interpreting time series roughness progression rates and identifying outlier types with MML inference.

A useful measure of performance comparison between sections of pavement is the roughness progression rate (RPR) which describes increase in roughness per time increment. This rate can be used to predict future levels of pavement roughness and a comparison of deterioration rates for different pavements within a network. Selecting appropriate regression functions to describe the RPR for each pavement interval presents two major problems. Roughness time-series can include roughness data with error, appearing to act independent of the observed time series trend. Including likely error values will bias the calculated roughness progression rate. The problem of identifying likely error is made more difficult with the possibility of maintenance intervention which may reduce the roughness level and/or progression rate. A criterion to select RPR based upon minimum message length (MML) inference is introduced in this paper and is referred to herein as MML RPR. We propose a method to learn RPR which is the combination of two parts. A segmentation model is used to learn whether any maintenance has caused a change in roughness progression. Secondly, a classified mixture model is used to identify likely error. We perform simulated comparisons comparing common segmentation criteria using unclassified mixture models and conclude that MML RPR is the preferred criterion. (a) For the covering entry of this conference, please see ITRD abstract no. E217099.