An Innovative Pavement Performance Prediction Method Based on Few-Shot Learning

Abstract

Pavement deterioration is a critical road maintenance issue, particularly in countries such as China, which has a vast road network. In the coming decades, many countries will need to address a significant number of road maintenance problems. Deep learning methods are widely used in the engineering field. However, the excellent performance of these deep learning methods requires sufficient data for model training. Therefore, an approach that can address the data shortages in pavement performance databases must be explored. This study proposes an innovative few-shot learning pavement performance prediction model based on time-series generative adversarial network (TimeGAN) data augmentation and transfer learning techniques. The proposed method integrates generative adversarial networks, advanced deep learning techniques, and transfer learning strategies to improve the accuracy of pavement performance predictions. The results show that the proposed method attains the highest accuracy compared with traditional machine learning and deep learning methods. In addition, a significant agreement was observed when comparing the predictors with the measured data, with an average coefficient of determination (R2) of 0.758. This study innovatively demonstrates great potential for developing an approach combining generative adversarial networks, deep learning, and transfer learning strategies for pavement performance prediction with insufficient data.