Domain Knowledge Guided Deep Neural Networks (DKG-DNN) for Prediction of Diaphragm Wall Deformation Induced by Excavation

Prediction of diaphragm wall deformation is vital to intelligent construction and environment safety in excavation engineering. This paper presents a novel prediction model entitled domain knowledge guided deep neural networks (DKG-DNN) which is tailored for the particular problem of diaphragm wall modeling. In the proposed model, deep neural network (DNN) is being trained to improve the prediction accuracy, whereas domain knowledge (DK) enables that the deformation predictions translate to the relationships between other physical variables, namely inclination location and vertical depth. To illustrate the effectiveness of DKG-DNN, experiments are conducted based on datasets from a real-world excavation project. The results show that DKG-DNN is competent to extract the complex spatio-temporal correlations of the wall deformations and to present satisfactory prediction accuracy. With the incorporation of the DK module, the robustness of the DNN model is improved to a large extent. Furthermore, collaboration between domain knowledge and deep learning in terms of framework design and learning objectives provides a new research paradigm which is highly possible to be generalized to other prediction problems in geotechnical engineering.