Session: 01-10-01: Machine Learning, Artificial Intelligence, and Deep Learning in Dynamics, Vibrations, and Control

Paper Number: 167248

Machine Learning and Sensory Integration for Real-Time Road Surface Assessment 

Road safety and vehicle efficiency are important concerns in our society and are heavily influenced by the condition of the road surface. The National Highway Traffic Safety Administration indicates that 22% of vehicular accidents are weather-related, resulting in substantial economic and human losses. Slippery roads due to rain, snow, and ice, can affect a vehicle's functionality and the safety of its occupants. Knowing the type of road surface (rough, smooth, etc) the car is traveling on is important for emergency braking and stability control. In this investigation we address these issues by effectively detecting and analyzing various road conditions in real-time, providing valuable information to the vehicle's control systems, which is essential for enhancing vehicle control systems, including braking and stability controls as well as better range prediction of electrical vehicles.
We developed a system that utilizes a combination of machine learning algorithms and hardware to deliver road surface assessments. The system classifies road conditions such as rough or slippery surfaces through tire sound and vehicle vibrations. Our methodology includes developing machine learning models that are optimized for processing multimodal data from microphones and accelerometers, resulting in accurate and timely road condition classification. The system consists of hardware embedded into existing automotive frameworks, notably via a Lincoln MKZ platform.The test data was collected using a data acquisition system consisting of a custom-made National Instruments LabVIEW interface connected via four channels to the following sensors: an accelerometer attached to the wheel well under the front hood, and three microphones—one next to the front wheel, one inside the vehicle, and one near the back wheel of the vehicle. Real-time signals were collected while driving on various highways and roads in Michigan that have different road conditions, and at a different range of speeds.
Given the raw waveform of one of the 4 sensors (accelerometer or microphones), we train a baseline machine learning model on the collected dataset to classify road type. Our architecture is inspired by human speech processing: it combines several convolutional encoders to learn feature representations and a transformer to classify the compressed sequence. Our preliminary results indicate that the model can distinguish between arterial and dirt roads with 84% accuracy.
Embedding this technology into vehicles will enhance the vehicle's control systems, facilitate V2X communication for traffic control and improved safety, and provide both the onboard systems and drivers with crucial road condition information for better decision-making and automated responses.

Presenting Author: Mihai Burzo University of Michigan

Presenting Author Biography: Mihai Burzo is an Associate Professor of Mechanical Engineering and Frances Willson Thompson Fellow at the University of Michigan, Flint, MI. His research interests include use of machine learning for computational fluid mechanics and heat trasfer, heat transfer in microelectronics and nanostructures, thermal properties of thin films, and multimodal sensing of human behavior. He is the recipient of several awards, including FWT Fellowhsip (2021) the 2006 Harvey Rosten Award For Excellence for “outstanding work in the field of thermal analysis of electronic equipment”, the best paper award at the PETRA conference in 2016 and the Semitherm conference in both 2013 and 2006, the Young Engineer of the Year from the North Texas Section of ASME (2006), a Leadership Award from SMU (2002), and a Valedictorian Award (1995).

Authors:

Artem Abzaliev University of Michigan
Rutchanon Hatasen University of Michigan
Hussein Kokash University of Michigan
Linda Zhu University of Michigan
Jun Chen Oakland University
Mihai Burzo University of Michigan