Strength prediction of asphalt mixture under interactive conditions based on BPNN and SVM

Abstract

Predicting the strength of asphalt mixtures with different specifications under various conditions was a highly challenging task. The standard strength test lacked consideration of multiple factors, resulting in an inability to accurately characterize the properties of the pavement. This paper proposed a strength prediction approach based on influence factor analysis using the back-propagation neural network (BPNN) and support vector machine (SVM). The strength dataset was processed to realize the physical analysis of factors influencing asphalt mixture strength. Stress state (Direct tensile with$250mm\times 50mm\times 50mm$, Uniaxial compression with $100mm\times \Phi 100mm$, Indirect tensile with$63.5mm\times \Phi 100mm$,Four-point bending with $380mm\times 63.5mm\times 50mm$), temperatures (35 ̊C, 25 ̊C, 15 ̊C, 0 ̊C, −15 ̊C, −25 ̊C), and load rates (0.02 MPa/s, 0.05 MPa/s, 0.1 MPa/s, 0.5 MPa/s) were selected as input features to train the BPNN and SVM. The strength prediction model for asphalt mixture under complex conditions was established by optimizing the parameters of algorithms. The performance of the BPNN and SVM was evaluated and compared by the root mean square error, determination coefficient, and mean absolute percentage deviation. The results show that the asphalt mixture specimen with different specifications under various stress states presents significant discrepancies. The maximum compressive strength is followed by the bending strength, then comes the indirect tensile strength, and the smallest is the direct tensile strength. The difference in the role of asphalt or aggregate is the main reason for the diversity in strength. The increase in temperature leads to asphalt softening, which reduces the strength of the asphalt mixture. The increased loading rate meant the loading time was short cause the strength increased. In addition, the predictive value of the strength under various conditions was consistent with the results of the experiments. The hidden neurons in the BPNN were set to 9, achieving the prediction accuracy is high ($R^2=0.99$). The penalty coefficient of the SVM was set to 500 and the kernel function parameter was set to 300, resulting in the error within 0.02 %. When comparing the performance metrics of BPNN and SVM, it becomes evident that SVM outperforms BPNN in terms of prediction accuracy. Specifically, SVM exhibits a coefficient of determination of 0.9983, a root mean square error of 0.208, and a mean absolute percentage deviation of 0.145, whereas BPNN demonstrates respective values of 0.9979, 0.233, and 0.067. This study lays a theoretical foundation for the digital and intelligent road construction.