For details, please refer to the paper
The smartphone-collected vehicle response data is being used to estimate the International Roughness Index (IRI). Among the existing smartphone-based methods, the data-driven approach, which involves training a machine learning model, is drawing more attention. However, surveying the IRI using conventional methods is expensive and there is limited labelled response data. In contrast, there exists a wealth of unlabelled response data from road users. This scenario presents opportunities for exploring semi-supervised learning (SSL) algorithms, which have been insufficiently researched in this domain. This study addresses this gap by applying an SSL framework to refine an IRI estimation model. Our results show that the SSL-trained model achieves a lower RMSE than the fully supervised trained model.
Research framework
Machine Learning Integration: Our model applies a multilayer perceptron (MLP) structure trained using real-world vehicle responses. We introduced the VIME ( Variational Information Maximizing Exploration) method in our SSL framework to leverage unlabelled data and improve the model’s accuracy over fully supervised models.
VIME algorithm
We engaged real participants, collecting data from different vehicle models and smartphone mounting configurations to ensure our model accounts for variations in real-world scenarios. Our model demonstrated significant improvement, achieving a lower Root Mean Squared Error (RMSE) compared to fully-supervised models, showcasing the effectiveness of integrating SSL into public vehicle-based pavement monitoring systems.
RMSE of the fully-supervised and semi-supervised models in different label rates