Artifical Intelligence and Roadway Friction Modeling

Date and Time: Tuesday, May 9, 2023: 10:00 AM - 12:00 PM
Location: Keck 100

 

Lead Presenter:  Thomas Brummet, Software Engineer
Affiliation:  National Center for Atmospheric Research (NCAR)
Social Media Handle: 

Lead Presenter Biography

Tom Brummet is a software engineer who has spent a large part of his career working with machine learning (ML) and artificial intelligence (AI). Tom has experience using machine learning in many different areas of research, from clean energy (such as wind and solar power generation) to maintenance and decision support systems, and has experience working with various state DOTs. Tom has also worked to develop and install real-time ML models which were used at Minneapolis-St. Paul International Airport to provide runway friction forecasts in order to assist with scheduling runway plowing and maintenance. Tom has grown into one of the primary engineers behind the Pikalert road forecasting and alert system. Pikalert provides a graphical display of state highways and roadways, and alerts state DOTs to potential upcoming weather events that could impact driving conditions. The Pikalert system is currently operational for Wyoming, Iowa, Oregon, and Alaska, and is in development for Colorado. 

Co-Authors

Gerry Wiener, Senior Software Engineer, National Center for Atmospheric Research;
Seth Linden, Software Engineer, National Center for Atmospheric Research;
Laura Fay, Senior Research Scientist, Western Transportation Institute

Presentation Description 

The objective of this research was to determine the relationship between weather conditions and friction measurements, determine if it is possible to standardize friction measurements from multiple friction sensors for identical weather conditions and pavement types, and to leverage artificial intelligence and machine learning to produce models that can accurately predict roadway friction coefficient using weather conditions at sites where friction measurements may not be available. Additionally, utilizing the notion of transfer learning, could a model trained on RWIS data from one state still be reasonably accurate when applied to RWIS data from a different state. Such a model could be applied at any RWIS station that records standard RWIS measurements. This research shows that ML models can be trained to accurately predict friction using just a small set of reliable measurements from RWIS sensors, along with proving that an ML model can be developed and used for transfer learning without a significant degradation in skill. This research also highlights the importance of variable selection in transfer learning models, and how specific variables, such as water layer depth and road state, can improve model performance, but also severely hurt the model when applied to a new domain.

Extended Summary

Extended Abstract

Download Extended Abstract PDF 

Presentation Video