1DCNN-based prediction methods for subsequent settlement of subgrade with limited monitoring data

Deep learning has attracted considerable attention in studies on soil deformation behaviour. However, its training process requires a large amount of data, while real engineering data often suffer from issues such as insufficient scale and irregular structure. This study proposes a subgrade settlement prediction method for reclamation airports in coastal areas with high advance prediction capability and precision. The method employs a one-dimensional variant of the convolutional neural network (1DCNN). To overcome the challenge of limited and irregularly engineering data, the model is trained on a high-fidelity synthetic dataset generated from ABAQUS. The effectiveness and dependability of the approach are assessed by predicting settlement in real-world projects. Furthermore, the study conducts an analysis of the internal prediction mechanism and the generalisation performance of the 1DCNN-based models. The results indicate that the proposed 1DCNN-based model offers higher prediction accuracy and superior long-term settlement forecasting capability compared to the Asaoka method. Additionally, 1DCNN outperforms the other two DL methods (BiLSTM and ConvLSTM) in terms of prediction accuracy. As the input data of pre-monitored settlement are processed, the 1DCNN-based models learn abstract features that transition into output labels. The learning rate emerges as the most critical factor influencing the reliability of prediction and should be adjusted as a priority to achieve optimal performance. Overall, this study provides a potential methodology for accurate advance prediction of subsequent settlement development in subgrade under staged loading conditions, utilising a small amount of pre-monitoring data.