Data-driven analysis on the subbase strain prediction: a deep data augmentation-based study

The service quality of the subbase may affect the overall road performance during its service life. Thus, monitoring and prediction of subbase strain development are of great importance for civil engineers. In this paper, a method based on the time-series augmentation was employed to predict the subbase strain development. The time-series generative adversarial network (TimeGAN) model was implemented to perform the augmentation of time-series data based on the original monitored data. The augmented data was trained through deep learning network to learn the feature correlation of the subbase strain. The effectiveness of TimeGAN on the prediction accuracy was evaluated through the Attention-Sequence to Sequence (Attention-Seq2seq) model, and temporal convolution network-adaptively parametric rectifier linear units (TCN-APReLU) model. Results indicated that the TimeGAN network could capture sufficient information from the time-series monitored data of subbase strain development so that the corresponding augmented data matches well with the original data, which improves the prediction accuracy. It is also discovered that the combination of TimeGAN and TCN-APReLU appropriately predict the subbase strain development based on the original monitored data