UC Berkeley

This dissertation presents a development of an integrated framework of artificial intelligence (AI) systems for earth pressure balance (EPB) tunnel boring machine (TBM) tunneling. The framework is constructed based on the feedback loop control system. The AI systems are developed using machine learning algorithms and structured to follow the human cognitive model, i.e., sensing, perceiving, and decision-making. The development of the systems is conducted in three parts.

The first part discusses the characteristics of TBM data and the effects of data preparation on data-driven models. This is achieved by (i) proposing a knowledge-based EPB TBM feature taxonomy and (ii) investigating the effects of data aggregation and feature selection on prediction models. The investigation shows that models developed using different data aggregation levels produce comparable prediction trends and similar feature importance rank in the conditions of sufficient observations and predictors. However, models with a coarser aggregation level may appreciate higher prediction performance due to the lower variance. The investigation also shows that the knowledge-guided TBM feature selection offers benefits over embedded machine learning-based feature selections. The developed model can produce relatively consistent feature importance in different tunneling cases, indicating better generalizability of the model.

The second part proposes AI systems that perceive tunneling environments in real-time based on the streams of sensor data during tunneling operation. This is achieved by developing (i) a supervised AI system to interpret the encountered geologic conditions, (ii) an unsupervised AI system to detect the encountered geologic anomalies, and (iii) a supervised AI system that connects TBM data to the ground monitoring data and estimates tunneling-induced ground movements. The proposed geologic interpretation system uses either Random Forests (RF) classification or regression algorithms to infer the geologic transitions along the tunnel alignment. The proposed geologic anomaly detection system combines Principal Component Analysis (PCA) to project the data into a lower dimension space and Local Outlier Factor (LOF) to measure the degree of the anomaly of the projected data points. The proposed tunneling-induced ground movement estimation system uses RF regression to approximate any shape of ground movements solely based on TBM operation data and tunnel spatial geometries without prior assumptions on the ground movement shape, geologic material parameters, and the expected ground loss.

The third part proposes AI systems that model tunneling and its decision-making processes. This is achieved by introducing (i) a combination of probabilistic graph modeling and structure learning algorithms as a tool to systematically explore the causal effect interactions contained in TBM operation data and (ii) a multi-output supervised AI system to determine multiple steering control parameters simultaneously and steer the TBM along the tunnel alignment. The study shows that Bayesian Network Structure Learning (BNSL) can potentially be used to model the interactions of human operator decisions, TBM behaviors, and ground conditions in an integrated representation. The study also shows that the multivariate Random Forest (MRF) algorithm can concurrently make multiple decisions on TBM steering control parameters during driving.

This dissertation has demonstrated that TBM data contain valuable information that can be extracted to benefit tunneling operations. Due to the complex relationships within the data, the nonlinear and nonparametric machine learning models offer advantages over the conventional linear and parametric models. This development is envisioned to be a building block that advances the development of autonomous TBM technology.