Uso de aprendizado de máquina interpretável para avaliação da deformação permanente em misturas asfálticas

• Ano de defesa: 2023
• Autor(a) principal: Mariano, Antonio Lucas Gabriel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Aprendizado de computador; Deformação permanente; Redes neurais (Computação); XGBoost; Shapley additive explanations (SHAP); Asfalto - Deformação; CNPQ::ENGENHARIAS::ENGENHARIA DE TRANSPORTES
• Link de acesso: http://repositorio.ufc.br/handle/riufc/75193

Permanent deformation or rutting is one of the main distresses observed in asphalt pavements. The new Brazilian pavement design method, called MeDiNa, considers permanent deformation of the surface course through the flow number (FN), indicated to classify asphalt mixtures in accordance with the solicited traffic. This method treats permanent deformation as a defect to be mitigated during the mixture design stage. Therefore, the existence of a prediction model for FN could be of great use in guiding the design methodology. Several works have used machine learning (ML) in different areas, including road infrastructure, and the technique presents itself as an alternative to make predictions about the behavior of asphalt mixtures in relation to permanent deformation. Many machine learning systems are essentially considered “black boxes”, because of the difficulty in understanding how the code works. In several applications, understanding the model’s prediction can be as relevant as the accuracy of that prediction. The work presented herein proposed the development of a modeling for classification of asphalt mixtures in relation to permanent deformation according to the corresponding traffic, using the explainable machine learning tool, SHAPley Additive exPlanations (SHAP). Artificial neural networks (ANN) and eXtreme Gradient Boosting (XGBoost) were used. The constructed database has information from 251 asphalt mixtures. Three configurations were proposed in relation to input variables (C1, C2 and C3). The model using XGBoost presented an accuracy of 84% for the predictions of the traffic class corresponding to the mixture using cross validation in configuration C3. Moreover, an analysis with the SHAP values has shown how the variables considered in the research affect the mixture behavior with respect to permanent deformation and consequently to the traffic to which the mixture is fit, therefore increasing the understanding and the way to treat this relevant distress for the asphalt pavement industry.