National Technical University of Athens (NTUA)
Πρόγνωση έντασης ψαθυρής θραύσης σε βαθιές εκσκαφές με χρήση τεχνικών μηχανικής μάθησης
Abstract
dc:descriptionThe rising need for mineral resources, combined with the gradual exploitation of surface reserves, is driving underground mining to deeper levels. The increasing extraction depth leads underground excavations to increasing stresses and hence increased instability risks. Concurrently, rocks at great depths, which usually exhibit higher geomechanical properties and brittle behaviour, are capable of storing large amounts of elastic energy, which they release violently in the event of failure, causing severe damage to excavations and endangering the safety of workers and the functionality of construction sites. As one of the most serious geological hazards, rock bursts have killed hundreds of miners and injured many more. It is therefore necessary to understand the conditions and mechanisms that underlie the phenomenon, to improve risk assessment methods and adopt appropriate preventive measures and techniques. Approaches regarding the prediction of the intensity and timing of a rockburst event can be divided into short-term and long-term, based on the stage of the project development. Long term forecasting is utilized in the design stage of a project primarily through the combination of empirical criteria and numerical modelling, while short term forecasts serve the construction stage of the project primarily relying on the use of geotechnical monitoring. Focusing on the long-term rockburst forecasting, this research aims at four objectives. Firstly, the performance of the most widely used empirical criteria regarding the prediction and classification of the phenomenon is examined to a dataset containing rockburst cases collected from the literature. Accordingly, the second objective is to investigate the corresponding performance achieved by a random forest algorithm, while at the same time examining the effect on the performance of the same algorithm due to the introduction of synthetic incidents derived from the SMOTE oversampling technique. The third objective concerns the development of a new oversampling methodology through the generation of synthetic instances derived from numerical simulations. The last objective concerns the preparation of the framework for a future import of new features into the dataset (e.g. support measures) or the addition of autonomous synthetic instances through the use of a more appropriate simulation code of brittle failure. The three introductory chapters provide the theoretical background regarding supervised and unsupervised machine learning in the field of data mining, the issue of class imbalance that is related mostly to small databases and the theoretical background of brittle failure and strainbursting emphasizing on its long-term prognosis. The theoretical background of the semi-empirical methodology Burst Hazard Potential is also presented. Chapter 4 concerns the development, analysis and pre-processing of the dataset consisting brittle failure incidents of various intensities that occurred in underground mines and in construction projects. In addition, a presentation of the class imbalance issue is presented, which hinders the development of reliable predictive models. Furthermore, the empirical criteria for predicting the intensity of brittle failure are evaluated so as to clarify their predictive potential. Chapter 5 concentrates on exploring the effect caused when introducing synthetic instances through the SMOTE technique on the predictive ability of a random forest algorithm. The influence of the SMOTE technique in terms of different attributes input numbers was examined, while the optimal number of neighbours required to optimise the results of the technique was explored. Finally, the overall performance range of the algorithm was assessed as well as the advantages and drawbacks of the technique. Chapter 6 proposes an alternative methodology for dealing with imbalanced datasets in which both the oversampling and undersampling techniques are exploited. The core of the methodology is the substitution of clusters containing rockburst instances among similar materials by representative synthetic instances derived from numerical modelling on the basis of the BHP index. The models are tuned and calibrated so that their mechanical properties are aligned with the clusters’ centroids and mimic their nearest neighbours in terms of stresses and rockburst intensities. Subsequently, they are subjected to a variety of stresses to fill all rockburst classes. As a result, a balanced synthetic dataset in terms of material and class participation is formed, which is used to train a Random Forest algorithm to predict rockburst. The results attained indicate the potential of the proposed methodology in enhancing the learning process and thus provide reliable long-term rockburst predictions.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Papadopoulos, Dimitrios
- Παπαδόπουλος, Δημήτριος
Subjects
dc:subject × 10Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/58699