Road Surface Monitoring System Using Semi-Supervised Machine Learning Ensemble Models

This study investigates a road surface monitoring system that integrates machine learning (ML) classification ensemble models to classify road surface elements based on raw acceleration and angular velocity data. The models aim to categorize roads into five distinct types, including speed bumps, patches, potholes-manholes, and both even and uneven road surfaces. To improve the performance of the models, a semi-supervised training methodology is employed, utilizing pseudo-labels obtained from the predictions of the original system for model retraining. The effectiveness of the ensemble models is evaluated using the F1-score metric, with the binary ensemble achieving an F1-score of 0.967, and the multiclass model obtaining an F1-score of 0.854. Testing on independent data confirms an overall F1-score of 0.877 for the classification of the five surface classes, representing an improvement over the original F1-score of 0.827. Further research is warranted to automate the semi-supervised training process, taking into account confidence thresholds and efficient retraining methods. This study contributes to the field of road surface monitoring by demonstrating the potential of semi-supervised training to enhance the accuracy and performance of ML ensemble models in classifying road surface conditions.