Prediction of Geological Conditions Ahead of the Tunnel Face: Comparing the Accuracy of Machine Learning Models Trained on Real and Synthetic Data

Abstract

Risk assessment during the construction of an underground structure (e.g., tunnel) should rely on accurate information of the rock mass that will be excavated. Advances in engineering equipment allowed vast data collection and thus opened a possibility for data-driven prediction of geological conditions. For an accurate prediction, the integration of various data sources is required, which makes the predictions complex and time-consuming for specialists. The application of machine learning (ML) methods provides a shortcut in analyzing complex data to predict geological conditions.

In this work, the authors present a new approach for predicting the geological information ahead of the tunnel face by using an ensemble of ML methods combined with an oversampling technique.

Implicit dimensionality reduction is introduced to deal with nonlinear coupling between seismic data: unsupervised machine learning methods are used to cluster the seismic data from two underground construction sites. Obtained information on clusters is then integrated with various seismic and geological variables, and a supervised machine learning model is trained to predict the rock mass class and/or the rock type. The data is over-sampled to avoid biased results when the training datasets are imbalanced and a mix of real and synthesized data is used in training, while an accuracy check is performed only on real data.

Our results show that the proposed ML ensemble model has high accuracy in predicting geological conditions. Furthermore, the application of the oversampling technique helps to improve the accuracy of the ML predictors further.