Keywords: pavement, condition, data mining, minimum message length, network, cleaning.

Road agencies often collect large databases describing the networks condition with respect to both length and time. From this information the current network condition can be inferred. Without good knowledge of the current condition, predicting future levels of deterioration is difficult. Regardless of deterioration models accuracy, the wrong starting point will lead to the wrong predicted value. To indentify current condition the historical trends in the data must be identified. These historic trends or rates of change are referred to as progression rates (PR).

From this historic data an agency may collect a range of condition data; roughness, rutting, strength, texture etc. The current condition along the network for each condition must be identified before deterioration models can predict future levels. This data is however often prone to errors, missing values, unrecorded maintenance interventions, changes in measuring devices etc. All these combine to make identifying the progression rates and current conditions very difficult.

This paper highlights the importance of identifying the current condition and how it may in fact lead to better overall predictions of deterioration than focussing on deterioration models. It introduces a new method to model a networks historical database based on a two dimensional matrix. It then builds on this to present a new data mining criterion based on minimum message length (MML) [1]. This new criterion has been designed from first principles as an inference tool to simultaneously identify PR, homogeneous segments [2], and maintenance interventions from practical datasets owned by road agencies. This is the first algorithm than identifies both PR and HS simultaneously, sharing information.

The new MML criterion referred to as the MML 2DS introduces a new method to identify noise or error in the data. It proposes the use of a mixture model which has been shown in Byrne [3] to correctly identify multiple types of outliers evident in the pavement condition data. This is combined with the PR and HS segmenting to describe the total progression of a complicated road agency. An example is presented whereby the MML 2DS is correctly able to identify a complicated network with multiple homogenous segments, maintenance interventions, and a large additional noise component.

1

C.S. Wallace, "Statistical and Inductive Inference by Minimum Message Length", Springer, Berlin, New York, 2005.

2

M. Byrne, S. Sanjayan, D. Albrecht, "Identifying Error and Maintenance Intervention of Pavement Roughness Time Series with MML Inference", Journal of Pavement Engineering, 2009.

3

M. Byrne, "A Data Mining Investigation into Pavement Roughness Using Minimum Message Length Inference", PhD Thesis, Monash University, Australia, 2007.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £140 +P&P)