Monitoring Road Quality Using Real-Vehicle Sensor Data and Machine Learning Models

Abstract

Urban road maintenance is challenging, especially for scalable, real-time systems. This research proposes a sensor-based method using affordable hardware to monitor road quality. The system integrates a Raspberry Pi with GPS modules (Libelium and Quectel L80) and MPU6050 IMU to capture vibrations and geolocation data. Testing spanned two setups: a custom Donkey Car running on Raspberry Pi on controlled tracks for repeatability, and a production vehicle on public roads for real-world validity. This dual approach combines precision with practicality. Real-road tests produced up to 100,000 timestamped records per session, totaling 438,629 points, including lat/long, timestamps, speed, three-axis accelerations, and gyroscopic data. Statistical analysis involved distributions, correlations, collinearity checks, and terrain segmentation. Vertical acceleration and speed stood out as key roughness indicators, aligning with physical intuition. From there, a chunk-based method was used to pull out features and build a range of regression models, from simple linear ones to more complex polynomial versions, to predict road roughness scores straight from the raw sensor feeds. These models performed really well, hitting desired R² and RMSE scores, which means they're solid for spotting potholes and bumps reliably. A strong link was also noticed between the patterns in gravitational acceleration and the actual physical traits of the road surface. To make the results easier to grasp, mapping was used and turned onto interactive GPS visuals using folium and open street libraries. These let to zoom around city maps and see road conditions briefly. The research emphasizes cost-effective, modular hardware bridging data science, embedded systems, and urban planning. It paves the way for predictive maintenance, smart mobility, and efficient infrastructure management, integrating Machine learning into city operations.