Sparse Data Transformation for Unsupervised Clustering for the Exploration Ahead of Tunnel Face
A. Sapronova, P.J. Unterlas, J. Hecht-Méndez, T. Dickmann, Thomas Marcher · NSG2021 27th European Meeting of Environmental and Engineering Geophysics · 2021
Summary In collaboration with Amberg Technology, 1Institute of Rock Mechanics and Tunnelling at Graz University of Technology is developing a model to predict geological conditions ahead of the tunnel face. The model employs a cascade of unsupervised and supervised machine learning algorithms and uses the seismic data and available geological documentation at the underground construction site. At first, we use unsupervised methods to cluster the entire dataset into an arbitrarily defined number of clusters so that each cluster contains a unique label (representing geological conditions). We then train supervised classifiers to predict the label(s) for each cluster. In this work, we elaborate on developing two critical parts of this model: pre-processing and unsupervised clustering of the dataset. We tested several methods for sparse data de-noising and clustering to answer the following question: ""given a sparse structured mesh or a graph (unstructured), what method can be used for pre-processing and unsupervised clustering, to reveal major characteristics (features) of the dataset at the low dimensional space."" Four methods that showed the most robust results were selected and used further in the model’s versioning.