MnROAD Data Mining, Evaluation and Quantification -- Phase I

A data filtering system for the MnROAD temperature database was designed and implemented. Fourteen inter-dependent quantitative tests were developed to identify and flag erroneous, questionable, or exceptional data. Four of the tests identify missing and intermittent data streams. Three of the tests analyze the time series from individual sensors and identify outliers. Three of the tests compare data streams of similar sensors; "similar" implies identical pavement type, general location, and sensor depth. The remaining four tests are summary tests that identify periods of unreliable data.

The specific analysis and quantitative results are based upon the 471,178,324 data records from 1,313 thermocouple sensors in 48 MnROAD test cells collected from 1 January 1996 through October 2007. The considered test cells include both hot mix asphalt and Portland cement concrete sections from both the Mainline and Low Volume Road.

The majority of the sensors performed very well: 714 of the 1,282 operational sensors produced reliable data more than 99 percent of the time. Only 18 of 1,282 operational sensors produce reliable data less than 50 percent of the time. Only 31 of the original 1,313 sensor were wholly non-operational. A wide variety of statistical tables and graphical representation were produced in a digital format for the considered data.

Although this project focuses on a particular set of data, the concepts and tools developed in this project are designed to be extensible to accommodate the filtering of the ongoing and future data collection efforts at MnROAD.