The complex behavior of asphalt binders makes it difficult to accurately predict their dynamic viscoelastic properties. However, the dynamic viscoelastic properties of asphalt binders are determined by their compositions. The main aim of this study was to model the relationship between compositions and dynamic viscoelastic properties in asphalt binders with machine learning algorithms. In this paper, the dataset of dynamic viscoelastic properties parameters was built using DSR test, and the dataset of composition parameters was established using elemental analyzer and infrared spectrometer tests. The effects of asphalt types, oil sources and aging effect on the macro and micro properties of asphalt binder were investigated. The principal component analysis (PCA) and distance correlation coefficient (DCC) methods were adopted to determine the input parameters of the model, while three machine learning algorithms, i.e., SVM, relevance vector machine (RVM) and random forest (RF) were used to predict the dynamic viscoelastic properties of asphalt binders. Results indicate that the content of element C was similar for different samples. The absorption peak at 1700 cm−1 was the most sensitive to oil sources. Nine variables were selected as input factors, namely, A1700, A870, A810, A745, A720, I1600, N, S and C/N. The modeling results show that the SVM performed better in predicting the dynamic viscoelastic properties than RVM and RF. Mean absolute percentage error (MAPE) of all fitting parameters were less than 10 % in SVM. The study enriches the prediction method of asphalt binder dynamic viscoelastic properties and provides technical support for realizing the on-demand design of asphalt material.
Read full abstract