Unsupervised machine learning for data-driven rock mass classification: addressing limitations in existing systems using drilling data

TL;DR

This paper presents an unsupervised machine learning approach using drilling data to improve rock mass classification, overcoming limitations of traditional systems by leveraging high-resolution data and advanced clustering techniques.

Contribution

It introduces a novel data-driven classification system based on MWD data, utilizing dimensionality reduction and clustering algorithms with domain knowledge integration.

Findings

Well-defined rock mass clusters can be formed from MWD data

Clustering results correlate with physical rock properties

Data-driven approach reduces human bias in classification

Abstract

Rock mass classification systems are crucial for assessing stability and risk in underground construction globally and guiding support and excavation design. However, these systems, developed primarily in the 1970s, lack access to modern high-resolution data and advanced statistical techniques, limiting their effectiveness as decision-support systems. We outline these limitations and describe how a data-driven system, based on drilling data, can overcome them. Using statistical information extracted from thousands of MWD-data values in one-meter sections of a tunnel profile, acting as a signature of the rock mass, we demonstrate that well-defined clusters can form a foundational basis for various classification systems. Representation learning was used to reduce the dimensionality of 48-value vectors via a nonlinear manifold learning technique (UMAP) and linear principal component analysis (PCA) to enhance clustering. Unsupervised machine learning methods (HDBSCAN, Agglomerative Clustering, K-means) clustered the data, with hyperparameters optimised through multi-objective Bayesian optimisation. Domain knowledge improved clustering by adding extra features to core MWD-data clusters. We structured and correlated these clusters with physical rock properties, including rock type and quality, and analysed cumulative distributions of key MWD-parameters to determine if clusters meaningfully differentiate rock masses. The ability of MWD data to form distinct rock mass clusters suggests substantial potential for future classification systems using this objective, data-driven methodology, minimising human bias.