Document Type : Research Paper
Authors
1
Master of Science, Department of Civil Engineering, Faculty of Engineering, Ferdowsi University of Mashhad, Iran
2
Department of Civil Engineering, Engineering Faculty,Ferdowsi University of Mashhad, Mashhad, Iran And School of Civil Engineering, College of Engineering, University of Tehran, Tehran, Iran
3
PhD student in the Safety and Traffic Department, University of Police Sciences, Amin, Tehran, Iran
Abstract
Accurate prediction of soil CBR (California Bearing Ratio) is of great importance in pavement design, but traditional CBR testing is time-consuming, expensive, and requires laboratory equipment. In this study, a hybrid approach based on Artificial Neural Network (ANN) and the Copula method is presented for estimating soil CBR. For this purpose, a database consisting of 1011 real samples of geotechnical parameters (including sand percentage, fine content, plastic limit, plasticity index, maximum dry density, and optimum moisture content) was collected and preprocessed. Then, a Multilayer Perceptron (MLP) neural network was designed and implemented using the Python programming language. Initial results showed that the base model achieved coefficient of determination (R²) values of 0.630 and 0.606 for training and testing data, respectively. In the following, to improve model performance, synthetic data were generated using the Copula method. Kolmogorov-Smirnov (KS) and Anderson-Darling (AD) statistical tests, along with comparison of probability density function (PDF) curves, demonstrated that the generated synthetic data had very high statistical similarity to the real data. After adding 50,000 synthetic data points to the training set, the neural network was retrained. The results indicated a significant improvement in model performance; the coefficient of determination (R²) for test data increased from 0.606 to 0.672, representing an approximately 11% improvement. The Mean Absolute Percentage Error (MAPE) decreased from 7.08% to 6.38%. Other evaluation metrics including MAE, RMSE, IOA, and IOS also showed meaningful improvements. The findings of this study confirm that the Copula method can serve as an effective strategy for synthetic data augmentation of geotechnical data and enhancing the accuracy of machine learning models. The proposed approach also offers significant economic and environmental benefits, including cost reduction, time savings, and decreased natural resource waste.
Keywords
Main Subjects
)