Deep learning application to roughness classification of road surface conditions through an e-scooter’s ride quality

Abstract

This study integrates e-scooter vibrational data with smartphone sensors, employing machine learning to evaluate road surfaces. The goal is to classify the road surface roughness level(s) equivalent to the high cycle fatigue threshold(s) experienced by the e-scooter. This information is fundamentally critical in determining the remaining service life prior to repairing or reconditioning the e-scooter. Three machine learning models—Random Forest Classifier, Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) with k-means clustering—were tested using various hyperparameter tuning, post-processing, and data splitting strategies. The models achieved high accuracies above 95%, with the SVM and k-means clustering model consistently reaching up to 100% accuracy and processing times under 700 ms, indicating potential for real-time applications. Despite challenges in data collection and preprocessing, the top SVM configuration using 5-fold cross-validation demonstrated substantial promise. An 80/20 data split initially resulted in lower accuracies due to inappropriate sequencing, which was rectified by adjusting data handling methods. The most successful model has promising applications in monitoring rider comfort and support preventative maintenance for e-scooters. For instance, a sudden change in classification outputs (e.g. derived from large ampitude vibrations) of an e-scooter could indicate maintenance needs, enabling timely interventions. This approach aligns with data collection efforts by companies such as Beryl and could be integrated into existing infrastructures. Future research could expand on these findings by examining a wider variety of surfaces and speeds and incorporating regression analysis to advance the models from classification to predictive analytics.