| description abstract | A simple but robust faulting prediction model is essential for jointed concrete pavement design and timely maintenance and rehabilitation activities placement. This study utilized machine learning algorithms, including linear model, support vector regression (SVR), k-nearest neighbor (KNN), decision tree, random forest (RF), and neural network (NN), to develop faulting prediction models based on the comprehensive Long-Term Pavement Performance (LTPP) data. The RF model turned out to be the most suitable model with the highest prediction accuracy. The most influential variables were selected to ensure the robustness of the model. The hyperparameters in the model were also finely tuned to improve its prediction performance. Moreover, the RF model was evaluated from various aspects. First, the variables were ranked by their importance, and the three most important variables are intense precipitation, pavement age, and dowel diameter, which are in good agreement with the faulting causes (i.e., moisture infiltration, traffic repetitions, and load transfer efficiency). Second, by comparing with the full model, the reduced RF model can still achieve a decent prediction accuracy (R2=0.848) while retaining robustness. Third, the confidence interval of model accuracy (R2) was constructed via bootstrapping to quantify the uncertainty. The result indicates a 95% chance that the R2 value falls between 0.643 and 0.854, which implies the model has satisfactory adaptability to other data sets. | |
