Long-term forecasting of shield tunnel position and attitude deviation using the 1DCNN-informer method

Abstract

Accurate prediction of shield machine position and attitude is crucial for ensuring the quality of tunnel construction. However, current machine learning models for predicting the position and attitude deviations of shield machines encounter significant challenges in achieving reliable long-term forecasting during shield tunneling. This study introduces a novel deep learning model, termed 1DCNN-Informer, which integrates the one-dimensional convolutional neural network (1DCNN) and the Informer model. The model was trained and validated using datasets from the Nanjing Metro shield tunnel project in China. Furthermore, the 1DCNN-Informer model was transferred to datasets from both similar and different geological conditions using the domain adversarial neural network (DANN) transfer learning method. The importance of input features was analyzed using the Shapley additive explanations (SHAP) method, complemented by experiments with various input parameter combinations. Results demonstrate that the 1DCNN-Informer model achieves superior performance compared to the Informer model and surpasses other comparative models, such as PatchTST, iTransformer, and Dlinear, in the majority of input sequence length and prediction sequence length combinations. Additionally, the DANN transfer learning method significantly enhances the 1DCNN-Informer model’s performance in the target domains dataset. The cutterhead rotation speed, advance speed, and chamber pressure are of critical importance in the prediction of shield position and attitude deviation. The proposed model not only represents a significant advancement in intelligent shield tunneling but also holds potential for broader application in automated equipment operations and multi-domain transfer learning studies in the field of engineering.