Application of Machine Learning for Data Analysis and Prediction in Tunnelling

In modern TBM tunnelling, vast amounts of data, like penetration, torque, thrust force and gripper force, are produced. Additionally, geophysical investigations and drilling programs are conducted to explore the geology beyond the tunnel face. Although this abundance of data offers many possibilities to analyze the encountered geology, utilizing it to its full extent is a difficult task for construction site personal and researchers as well.

Machine learning is a field of computer science, which applies algorithms to recognize patterns and structure in data. Recent breakthroughs of this discipline are on the one hand based on today's powerful computational capabilities, on the other hand on the massive amount of data that is available. Supervised learning, unsupervised learning and reinforcement learning are the three subdisciplines that concern classification and outcome prediction, finding hidden data structures or learning a series of action from the data.

The research goal of this project is to explore the applicability of various machine learning techniques on the analysis of advance investigations from recent major tunnelling projects. Strong focus on the practical use of algorithms is seen as key to optimize data interpretation workflows and efficiency of predictions. Ideally the findings can be used for correlating recorded data to observed geology, for enhancing the quality of predictions ahead of the tunnel face and to transfer learned knowledge from existing to future tunnel excavations.



Excavation in fault zones

Fault zone – an expression which has a grave meaning in tunnelling. Neither a geologist nor a geotechnical engineer needs a more detailed description to know which geotechnical situations are connected with this term. Although many fault zones have already been excavated more or less successfully, the gained knowledge is seldom documented or shared. Thus, it is not surprising that major problems occur consistently and evidences for upcoming fault zones are ignored or misinterpreted. Not at least because fault zones are one of the most challenging ground conditions in tunnelling.

What possibilities does the engineer have to predict an upcoming fault zone at site and in time? What measures must be applied to cross the fault zone as favourable as possible (concerning the utilization of the lining and the costs)? What chosen measures by the engineer have more than unfavourable consequences? To answer these questions, former excavations in fault zones shall be investigated to draw geotechnical as well as constructional relevant conclusions, validated by theoretical and scientific considerations. Consecutively, general guidelines for excavations in fault zones shall be elaborated, focussing on their practicability.


Predicting the behavior of rock mass FFG Bridge project ROCKBURST

Predicting the behaviour of rock mass is very important in rock engineering. The Institute of Rock Mechanics and Tunnelling (TUG), the Veitsch Radex GmbH & Co OG, the Institute of Applied Geosciences (TUG) and the Department of Geoinformatics - Z_GIS (University of Salzburg), work together in a project called "Devastating micro cracks: researching spontaneous rock failure with rock mechanical testing, µCT, OBIA and geostatistics (ROCKBURST)”.

This 3-year FFG-project presents an interdisciplinary approach to research the influence of micromechanical and microstructural characteristics and their individual effect on rock failure. State-of-the-art rock mechanical testing, data acquisition and simulation methods will be nested in an innovative work-flow: Following initial mineralogical investigations, artificial and natural rock samples will each be subject to destruction-free microstructure analysis and compression testing. From each cycle, micro computed tomography will provide 3D data on the sample's micro-fabric before a successive step of uniaxial testing controlled by acoustic emission testing is carried out. From the resulting series of µCT data, object-based image analysis will assemble a process model of microcrack evolution. Geostatistical simulations will be used to upscale the results from sample to excavation scale. New insights into relevant micromechanical processes in highly-stressed brittle rock and their quantitative comprehension are expected.


Rock Mechanical Aspects of the Bischofsmütze

© Land Salzburg - Geologischer Dienst

The Große Bischofsmütze is one of the famous peaks in the Salzburg mountain region. However, this peak slowly crumbles. Huge parts of the peak detach during major rock fall events. The mass movements can reach up to 50,000 m³. The latest two events happened in the 1990s and in 2007 at the southern and eastern mountainside.

These two events led to an intense monitoring programme and new discontinuities close to the peak indicate a new rock fall event. For this reason, the stability of the Große Bischofsmütze shall be investigated. Following questions shall be answered:

  • Can a new rock fall at the Große Bischofsmütze be expected?
  • What would be the volume of this new rock fall event?
  • What are the sensitive parameters for the stability of the endangered area?
  • What risk proposes a novel land slide to the adjacent infrastructure?

This thesis is part of a study to determine the volumetric block-size distribution and rock mass characteristics with the information of only one outcrop by using remote sensing. Tools for the elaboration of this thesis are statistical analysis, numerical simulations and close-range terrestrial digital photogrammetry.

Remote Characterization of Rock Masses

© Buyer – TU Graz/FMT

Discontinuities dominantly influence the mechanical behaviour of rock masses. Thus, it is of crucial importance in rock mechanics to have a profound knowledge about the discontinuity network. Traditional measuring techniques provide a rough knowledge about a discontinuity network but are also limited within reachability, geological/geotechnical knowledge, time and scale. The results are subjective and not reproductive.

The application of remote sensing techniques like photogrammetry and laser-scanning, in geotechnics helped to collect data about the discontinuity network and reduce the bias, which is introduced by manual mapping. Together with numerical models, reliable predictions about the stability of the rock face and the discontinuous rock are nowadays possible. But there is still missing an automatic data collection, as well as an automatic implementation of the collected data in a numerical rock model.

In his research, Andreas Buyer focusses on the automatic mapping of discontinuities aided by remote sensing techniques using ShapeMetriX3D (3GSM GmbH). This means:

  • Extraction of discontinuity features from digital images and digital surface models;
  • Automatic coupling of mapping data with numerical software like UDEC or 3DEC (Itasca Inc.);
  • Numerical simulation of discontinuous rock masses for stability analysis.

Contact: Andreas Buyer

Common interests
image/svg+xml
Students, who are interested in the research topics described are very welcome to contribute to the research progress by writing a scientific thesis. Engineers around the world are invited to share thoughts and experiences in order to bundle strengths.