Image-based prediction for enclosure structure deformation in pipe-roof tunnel construction using a physical-guided and generative deep learning method

Abstract

This paper proposes an image-based enclosure structure deformation prediction model called the physical-guided and generative deep learning (PG-GDL) method for pre-support tunnel construction, filling critical gaps in physical-guided image-based datasets and image-to-image prediction of structure deformations. The PG-GDL method establishes reliable correlations between real-time construction information and tunnel deformation patterns, enabling the generation of high-accuracy deformation predictions derived from on-site construction information. It comprises a physical-guided data visualization method to convert one-dimensional datasets into intuitive visual representations and employs a hybrid Convolutional Autoencoder Wasserstein Generative Adversarial Networks (CAE-WGANs) framework to achieve high-accuracy predictions of numerical values and planar images. For validation, the proposed PG-GDL is applied in a pipe-roof project in Shanghai, China, generating precise maximum deformation predictions and high-quality planar deformation images. The PG-GDL method provides engineers with a reliable and multi-perspective tool for deformation prediction and improves decision-making in pre-support tunnel construction projects.