With the rapid development of tunnel engineering, geological dangers such as collapse during tunnel excavation, water intrusions and landslides pose serious threats to construction safety, resulting in delays in projects, cost overruns and even casualties. Accurate prediction of geological risks is important to prevent geohazards and to ensure an efficient and safe construction process. However, traditional geological exploration techniques have limitations. Borehole logging provides relatively accurate but sparsely sampled information, but non-invasive methods like seismic techniques have high spatial resolution, but not sufficient accuracy. On the other hand, traditional machine learning methods for geological risk prediction combat limited early stage structural data and often fail to adaptively update new streaming data.
Therefore, Limao Zhang, Ying Wang, Xianlei Fu, Xianlei Fu, and Penghui Lin of Huazhong University of Science and Technology in China and Nanyang Technology University in Singapore, conducted jointly with “Geological Risk Prediction under Uncertainty in Tunnel Excavation Using Online Learning and Hidden Markov Models.”
This study proposes an online hidden Markov model (OHMM) that combines online learning with hidden Markov models to estimate geological risk. OHMM is tailored to the continuity nature of the observational data, allowing for adaptive updates with each new data. To address the limited data challenges in the early stages of the structure, pre-construction borehole samples are used as additional data, and the observational extension mechanism is designed to extend the short sequence of observational data to the length of the complete sequence. The effectiveness of OHMM using this observational extension mechanism has been demonstrated through case studies on a tunnel excavation project in Singapore, including a 915 tunnel ring.
The findings show that OHMM outweighs traditional methods such as hidden Markov models, long-term short-term memory networks, neural networks, and vector machines for predicting geological risks before tunnel boring machines. In particular, with the help of observational augmentation mechanisms, OHMM can accurately predict geological risks in areas that have not yet been constructed using limited observational and site survey data. This achieves a forward accuracy of 0.968 when using 600 observation rings and 300 observation rings. It maintains high sensitivity to high-risk geological conditions and provides reliable early warnings for high-risk areas under construction. Furthermore, OHMM can effectively predict geological risks up to 100 rings. Also, a distance of 30 rings is recommended to balance prediction accuracy and stability. This study advances geological risk prediction models by providing online update capabilities for tunnel drilling projects, enabling early stage risk prediction and long-term forecasts with minimal historical data requirements, and maximizing the use of site survey data.
“Geological risk prediction under uncertainty in tunnel excavation using online learning and hidden Markov models,” written by Limao Zhang, Ying Wang, Xianlei Fu, Xieqing Song, and Penghui Lin. Full paper: https://doi.org/10.1007/S42524-024-0082-1.
