Theory-guided machine learning for predicting and minimising surface settlement caused by the excavation of twin tunnels / Chia Yu Huat

Chia , Yu Huat (2024) Theory-guided machine learning for predicting and minimising surface settlement caused by the excavation of twin tunnels / Chia Yu Huat. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (143Kb)
	PDF (Thesis PhD) Restricted to Repository staff only until 31 December 2026. Download (4Mb)

Abstract

In response to worsening urban traffic congestion, metro tunnels have emerged as a solution to ease pressure on road networks. Shield machines, like earth pressure balance and slurry machines, are pivotal in modern tunnel construction. However, twin tunnel construction in urban areas commonly faces surface settlement (SS) issues, which threaten nearby structures. Traditional empirical formulas for SS estimation are limited to specific soil types and lack consideration of other factors. To overcome these limitations, this study introduces a comprehensive approach that combines 3D numerical analysis and machine learning to predict SS during twin tunnel excavation. The 3D numerical analysis factors in construction stages, tunnel geometry, and operational parameters while incorporating in-situ and lab test results to establish engineering soil parameters. Validation against field measurements yields R2 values of 0.94 and 0.96 for the first and second bored tunnels. While 3D numerical analysis provides accurate SS estimates, it is time-consuming. To enhance prediction efficiency, validated numerical models serve as the foundation for data generation. This dataset, alongside key parameters like cover-to-depth ratio, pillar width, soil stiffness, cohesion, friction angle, and overburden-to-face pressure ratio, integrates into a machine learning framework using a theory-guided approach. Conditional Tabular Generative Adversarial Networks (CTGAN) generate additional data from 20% of the 3D numerical analysis results. The study primarily focuses on tree-based techniques, including Random Forest (RF), Adaptive Boost (ADABoost), Gradient Boosting Tree (GBT), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting (LGBoost), and Categorical Gradient Boosting (CatBoost). Comparative analyses highlight CatBoost as the most accurate SS predictor among all machine learning (ML) models. Besides, in comparison with the CTGAN data generated for the ML analysis, data generated from the finite element model used in the ML analysis has outperform the prediction than the CTGAN of synthetic and hybrid data. This is due to the data generated from the numerical model possess the pattern for the ML algorithm ease of prediction. In addition, Coati Optimization algorithm, Particle Swarm Opimisation (PSO) and Bayesian Optimsiation (BO) are integrated to identify optimal parameters and minimize settlement during twin tunnel excavation and GBT with the optimisation algorithm has shown consistent capability identifying the least SS induced by twin tunnels Keyword:

Actions (For repository staff only : Login required)