Loading...
Study of tunnel-induced ground settlement using machine learning and remote sensing techniques
Liu, Linan
Liu, Linan
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2022
Date Submitted
Keywords
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
2023-11-04
Abstract
Tunnel excavation in urban settings inevitably disturbs the ground and leads to surface settlement. Such tunnel-induced ground subsidence should be limited to a tolerable threshold to avoid affecting aboveground structures. In this research, Machine Learning (ML) techniques are applied to predict tunnel-induced ground subsidence based on in-situ monitoring data collected from the Yuji tunnel project in China. Interferometric Synthetic Aperture Radar (InSAR), a remote sensing technique, is applied to map uneven ground subsidence along the tunnel alignment in the context of twin tunnels in downtown Los Angeles, USA.
Urban tunnel projects are typically constructed within a limited period, leading to small ground monitoring datasets ranging from 10 s to 100s being available to access. The effectiveness of ML application with a small dataset is discussed. Seven widely applied ML models in worldwide tunnel projects are selected and compared, including multiple linear regression, decision tree, random forest, gradient boosting, support vector regression, back-propagation neural network, and permutation importance-based back-propagation neural network models. Results demonstrate the non-linear relation between ground subsidence and correlated parameters. Our study reveals that the RF algorithm outperforms other models with the highest model prediction accuracy (0.9) and best stability/variance (3.02 × 10–27).
ML techniques provide new insight into tunneling excavation from a data-driven perspective. However, engineers desire to produce physically consistent results from ML models. This study proposed a physic-informed machine learning (PIML) framework to fulfill the gap between physics theories and ML architectures, taking advantage of the generalizability of physics-based models and the robustness of data-driven ML models. Physics models and theories were applied to guide input selection and architecture design of a developed voting regressor model formed by three identified well-performing base algorithms, including random forest, gradient boosting and decision tree models. The proposed PIML framework aims to improve ML predicting performance and interpretability.
Datasets for ML analysis are collected from in-situ measurements, which are monitored section to section along the tunnel alignment. In order to have a synoptic view of ground subsidence along the entire tunnel alignment, the InSAR technique is applied to map tunneling-induced ground subsidence that occurred concurrently with the construction of twin tunnels in downtown Los Angeles, USA. The average ground subsidence rate map indicates that the ground settlements along tunnel alignment are unevenly distributed. ML-based parametric analysis reveals that such an uneven subsidence field is geologically sensitive, dominated by the thickness of artificial fill and alluviums above tunnel alignment.
Together, these insights may improve the application of ML and InSAR techniques to tunnel-induced ground settlements. Also, this study can serve as a reference for geotechnical engineers to apply ML and InSAR techniques in practical tunnel projects.
Associated Publications
Rights
Copyright of the original work is retained by the author.