Geological hazards such as collapses, water ingress, and landslides pose serious threats to tunnel construction, potentially leading to delays, cost overruns, and safety incidents. The challenge is compounded in the early stages of excavation, when detailed subsurface information is limited and conventional geological survey methods—whether invasive drilling or non-invasive geophysical techniques—offer only partial, sometimes inconsistent, insights.
A research team from Huazhong University of Science and Technology and Nanyang Technological University has developed a novel online hidden Markov model (OHMM) to tackle this uncertainty. By merging online learning capabilities with the probabilistic framework of hidden Markov models, the OHMM can continuously update its geological risk predictions as new in situ observations arrive, without waiting for lengthy data collection cycles.
One of the major innovations of the approach is an observation extension mechanism designed for the data-scarce early excavation stage. This mechanism integrates pre-construction borehole samples—often the only reliable geological data available before tunneling begins—into the OHMM framework. By intelligently extending short observation sequences to the length of a complete excavation dataset, the model preserves predictive accuracy even when historical data is minimal.
The researchers validated the method in a tunnel excavation project in Singapore, where geological conditions vary along the alignment. The OHMM produced high-resolution, ring-by-ring forecasts of geological risks ahead of the tunnel boring machine (TBM). Compared with traditional approaches—including standard hidden Markov models, neural networks, long short-term memory networks, and support vector machines—the OHMM consistently delivered superior accuracy, especially for forward predictions into yet-to-be-excavated regions.
The study revealed key operational benefits. First, the continuous model updating allowed for timely adaptation to changing geological patterns, enabling earlier hazard alerts. Second, by maximizing the utility of sparse borehole data, the method bridged the information gap between pre-construction investigations and ongoing excavation. Finally, the framework’s ability to operate effectively with minimal historical data made it particularly valuable for early-stage risk prediction, where proactive measures can prevent costly and dangerous incidents.
Beyond immediate tunneling applications, the researchers note that the OHMM framework could be extended to other infrastructure projects where geological uncertainty is high and observational data arrives incrementally—such as mining operations, slope stability monitoring, and underground storage development. The integration of machine learning with probabilistic risk modeling also opens the door for future enhancements, including coupling with real-time sensor networks and incorporating physical process models for improved interpretability.
By providing a dynamic, data-efficient, and forward-looking tool for geological risk prediction, the OHMM offers engineers and project managers a new way to safeguard tunneling operations, optimize construction schedules, and improve safety outcomes in complex subsurface environments.
See the article:
Geological risk prediction under uncertainty in tunnel excavation using online learning and hidden Markov model
https://doi.org/10.1007/s42524-024-0082-1
DOI:
10.1007/s42524-024-0082-1