CSI has collaborated with notable entities in the Tunneling & Underground Construction industry: Sizense, Enzan Koubou, and Kajima to implement AI/ML techniques to improve TBM tunneling.
1) Supervised geologic interpretation based on TBM data
This research aims to develop a supervised system for real-time interpretation of geologic conditions by connecting continuous tunnel boring machine (TBM) operation data with geologic information from sparsely distributed boreholes. The study involved multiple tunnel datasets from various locations worldwide. The Random Forest algorithm was utilized to identify patterns between geologic classes and TBM feature interactions. Feature importance measures were employed to evaluate the contribution of different features to the classification model. The findings demonstrate that the model successfully inferred geologic transitions and captured geologic features. Moreover, the results emphasize the significance of feature interactions in capturing geologic information, even when certain features have higher weights for classification.
2) Unsupervised geologic anomaly detection based on TBM operation data
This research aims to develop an unsupervised anomaly detection system for real-time inference of changing ground conditions during tunneling. The approach combines dimensionality reduction and density-based outlier detection techniques using features from an earth pressure balance tunnel boring machine (EPBM), such as cutter, thrust, and ground conditioning systems. Principal component analysis (PCA) is employed to extract ground condition information and reduce dimensionality, while the local outlier factor (LOF) measures the anomaly degree of projected data points. The study, based on data from the State Route 99 (SR99) tunnel construction in Seattle, WA, demonstrates that PCA effectively clusters EPBM data based on ground conditions, and LOF proves to be a reliable measure for detecting changes in ground conditions.
3) Exploring interactions of EPB TBM features utilizing Bayesian network structure learning
This research aims to utilize a Bayesian network (BN) and a structure learning algorithm to investigate the interactions and causal relationships among excavation features of an earth pressure balance tunnel boring machine (EPBM). The study utilized a dataset from the State Route 99 (SR99) tunnel construction in Seattle, WA. A score-based structure learning algorithm was employed to construct the BN graphs, and the influence of explicit geologic condition information on the interactions was assessed. The findings demonstrate that BN graphs have the potential to model the interactions among EPBM features in a concise and interpretable manner. The score-based algorithm successfully captured several meaningful mechanisms of feature interactions based on the data, and their dependencies may provide insights into how EPBM operators controlled the machine.
4) Real time estimation of tunneling-induced ground movement based on TBM operation data
This research aims to develop an artificial intelligence (AI) system that connects tunnel boring machine (TBM) operation data to ground monitoring data to estimate tunnelling-induced ground movements in real-time. The study utilized a dataset from the State Route 99 (SR99) highway tunnel in Seattle, USA. The AI system, based on a dynamic sequential scheme and employing Random Forest as the prediction method, successfully estimated ground movements along the tunnel chainage and accounted for changing TBM control parameters. The system fills gaps in ground movement information from discrete monitoring instruments. This research demonstrates the potential of the system as a tool for controlling TBM and reducing induced ground movements in urban tunnelling.
5) Modeling TBM steering control decisions using machine learning
This research aims to develop an artificial intelligence (AI) system for steering control decisions in tunnel boring machines (TBM) that adapt to changing trajectories and ground conditions. The study utilized a dataset from the State Route 99 (SR99) highway tunnel in Seattle, USA. The AI system, based on a multivariate Random Forest algorithm, successfully mimicked human operator decisions by assigning appropriate steering control parameters based on the current and next TBM positions. The study highlights the importance of a sequential training scheme and incremental learning to address variations in ground conditions and trajectories. The results suggest the potential for developing a self-driving TBM system using an end-to-end learning framework.
Collaborators: Sixense, Enzan Koubou, Kajima
Researchers: Dayu Apoji, Kenichi Soga (UC Berkeley), Yuji Fujita (Enzan Koubou), Zhangwei Ning (Sixense)
Domains: Lifeline Underground Infrastructure
Capabilities: Data Analytics & Machine Learning
Publications:
Apoji, Dayu & Soga, Kenichi. (2023). Soil Clustering and Anomaly Detection Based on EPBM Data Using Principal Component Analysis and Local Outlier Factor. 1-11. 10.1061/9780784484982.001.
Apoji, Dayu & Ning, Zhangwei & Soga, Kenichi. (2023). Connecting EPBM Data to Ground Movement Data Using Machine Learning. 181-194. 10.1061/9780784484708.017.
Apoji, Dayu & Ning, Zhangwei & Soga, Kenichi. (2022). From Sensing to Machine Learning in Geoengineering. GEOSTRATA Magazine. 26. 44-51. 10.1061/geosek.0000453.
Apoji, Dayu & Fujita, Yuji & Soga, Kenichi. (2022). Soil Classification and Feature Importance of EPBM Data Using Random Forests. 520-528. 10.1061/9780784484029.052.
Apoji, Dayu & Fujita, Yuji & Soga, Kenichi. (2022). Exploring Interactions among EPBM Features using Bayesian Networks. 10.31224/2141.