Data-Based postural prediction of shield tunneling via machine learning with physical information

When applying machine learning algorithms to predict the posture of shield machines during tunneling, the generalization performance of these models might not be as good as expected when applied to different projects. In this study, the Transformer method was improved by incorporating physical information from finite element method (FEM) numerical simulation to enhance the generalization performance of the models, achieving accurate prediction of shield machine posture during the construction of new projects. A refined FEM numerical simulation of the shield machine construction process is established, and the FEM surrogate model is trained through the calculation of a large number of cases. Simultaneously, a data-driven machine learning model (Transformer) is developed based on historical shield machine construction data. The FEM surrogate model is then introduced into the data-driven Transformer model as physical information, resulting in the creation of a Transformer model with physical information. By comparing the prediction accuracy of the FEM surrogate model, the data-driven Transformer model and the Transformer model with physical information from a new project, the results show that the Transformer model with physical information significantly improves the prediction accuracy.