Machine learning methods for ore/waste classification in the mining industry

In a recent article published in the journal Ore Geology ReviewThe Spanish researchers are focusing on developing a comprehensive methodology to identify lithology in underground mines by integrating chemical analysis of drill tips and measurement-while-drilling (MWD) data.

This integration is essential to assess the chemical composition of the borehole zone at high resolution, a challenge addressed in the illuMINEation project. The aim of this study is to develop a site-specific predictive model for defining the ore/waste boundary utilizing machine learning techniques and to validate an offline automated drilling system combined with MWD data.

background

Drilling plays a vital role in the mining industry by providing key insights into rock mass properties, including ore grade and boundary definition. Core drilling operations involve the extraction of samples for geomechanical and chemical characterization essential for resource estimation and mine planning. Traditional drilling exploration campaigns are conducted to guide mining method selection and continually update the geological model during mine operations.

Underground mining increases the challenges associated with sampling and analyzing drilled chips due to limited space, harsh working conditions, and the need for adaptability to different drilling rig configurations. Innovative systems such as “autosamplers” have been developed to address these challenges by collecting secondary samples for offline X-ray fluorescence (XRF) analysis. Additionally, the use of water as a flushing fluid in underground drilling produces wet or semi-wet chips, necessitating the adaptation of analytical techniques such as XRF and laser-induced breakdown spectroscopy (LIBS).

Current Research

The methodology adopted in this study was to use a modified drilling rig with a semi-automated control system to monitor drilling data from a complex fluorspar deposit. The database for analysis consisted of more than 840 chemical samples and 23,000 MWD records accompanied by geological and geotechnical data. Two main approaches were used: a K-means clustering heuristic algorithm and a machine learning (ML) ensemble model.

A k-means clustering algorithm was applied to define rock classes based on physicochemical characteristics. The algorithm helped classify rock classes by identifying similarities and differences in drilling parameters. By dividing the data into distinct clusters, the algorithm helped classify ores and wastes based on their unique properties.

The ML ensemble model was trained using the defined rock classes to distinguish between ore and waste based on drilling parameters. Oversampling and undersampling techniques were implemented to address the imbalance in rock class sizes. The prediction success of the model, measured by F1 score, was approximately 70% for ore with high silica content and 86% for waste.

Continuous drill tip sampling was critical to reduce operator bias and improve the F1 scores for all classes using various ML techniques. The study also investigated the correlation between rock mineralogical properties and mechanical response during drilling. This analysis highlighted the benefits of using MWD-based information for analysis-while-drilling (AWD) in the ore grading process.

Results and discussion

This study revealed important insights into classifying rock classes based on drilling parameters and distinguishing between ore and waste in underground mining operations. Machine learning techniques, especially the k-means clustering algorithm and ML ensemble model, proved to be effective in obtaining accurate classification results.

Unsupervised ML techniques such as the k-means algorithm were crucial in classifying rock classes based on the variance of drilling parameters. This approach helped identify key parameters such as penetration rate (PR) and feed pressure (FP), which contributed significantly to the classification process. By splitting the data into distinct clusters, the algorithm provided a structured framework for analyzing the relationship between drilling parameters and rock types.

The hybrid ML model, built using resampling methods, ensemble decision tree classifiers, and hyperparameter tuning techniques, demonstrated the ability to handle imbalanced classes and complex relationships between drilling parameters and rock types. The model achieved high F1 scores, over 70% for ores with high silica content and about 86% for wastes, demonstrating its effectiveness in accurately distinguishing ores and wastes based on drilling data.

Continuous sampling data played a key role in improving the classification performance of the ML model. The model achieved higher F1 scores for both ore and waste categories by reducing operator bias and increasing data resolution. This study highlighted the importance of considering continuous sampling data and employing advanced ML techniques such as Gaussian SVM to improve ore/waste recognition accuracy in underground mining operations.

Conclusion

Integrating MWD data with chemical analysis of drill tips utilizing machine learning techniques provides a promising approach for accurate identification of lithology in underground mining operations. The findings of this study highlight the effectiveness of ML models to classify rock classes based on drilling parameters, leading to improved efficiency and accuracy in ore/waste recognition. This study contributes to bridging the gap in underground working environments by combining state-of-the-art MWD data analytics with AI techniques.