Machine learning-based method enhances accuracy of measuring dip angles and directions in rock facets

Machine learning (ML) algorithms are constantly finding new applications in all scientific fields, and geological engineering is no exception. Over the last decade, researchers have developed various ML-based techniques to determine geological features more effortlessly in rocks, such as the dip angle (the angle at which a planar feature is inclined to the horizontal plane) and direction of rock facets in tunnels. Understanding these characteristics is essential for large construction projects as they help ensure structural stability and safety, preventing potential failures or collapses.

Although powerful, most ML models still struggle to differentiate between joint bands and joint embedment points in rock. To clarify, joint bands are broader, less distinct areas within the rock that may include multiple parallel fractures, while joint embedment points are more localized features representing the actual intersections of rock layers.

As direct indicators of surface orientation, joint embedment points enable a more accurate measurement of dip angle and direction than joint bands. Thus, methods that can eliminate joint bands from input data can increase the accuracy of ML-based techniques, leading to more precise geological assessments.

To fulfill this challenge, a research team led by Professor Hyungjoon Seo of Seoul National University of Science and Technology (SEOULTECH) developed the Roughness-CANUPO-Dip-Facet (R-C-D-F) method. This ML-powered, multistep approach combines many filtration techniques to remove joint bands while preserving most joint embedment points in the data, leading to excellent accuracy when measuring dip angle and direction. Their paper was published in the journal Tunnelling and Underground Space Technology on December 1, 2024.

The first step of the filtration process consists of a roughness analysis on an input 3D point cloud, taken directly from a rock surface. This step removes minor surface irregularities and noise from the data, preserving continuous lines on the surface but removing joint lines.

The second filtration step uses the CANUPO algorithm, which classifies points based on their geometric characteristics and isolates key features, removing even more joint lines. The third filtration step eliminates connecting rock segments based on dip angles, isolating distinct rock formations. Finally, the measurement stage consists of facet segmentation to obtain the dip angle and direction of each section of the rock sample.

The researchers tested the R-C-D-F method on various real tunnel face images, achieving remarkable accuracy rates ranging from 97% to 99.4%. Notably, 100% of joint bands were successfully removed while still preserving 81% of joint embedment points. But the most attractive aspect of this technique was its fully autonomous nature, requiring no human intervention.

“By automating the process of filtering and segmenting rock features, it reduces human error and computational inefficiencies, making it ideal for modern infrastructure projects that demand high accuracy and reliability,” highlights Prof. Seo.

Overall, the proposed approach could find promising applications across many disciplines of structural and geological engineering.

“The R-C-D-F method’s integration of ML and deep learning ensures reliable and accurate geological data processing, which can directly improve the safety of large-scale engineering projects like tunnels and underground structures,” notes Prof. Seo. “It could also enable the development of smarter and faster geological analysis tools, reducing costs and improving efficiency in industries reliant on subsurface exploration and infrastructure development.”

The innovative approach thus holds great promise for paving the way for safer and more efficient geological engineering solutions.

Provided by
Seoul National University of Science & Technology