Deep mining operations are increasingly challenged by rockburst events, which pose significant safety risks to personnel and equipment. This study focuses on enhancing rockburst prediction capabilities through the integration of microseismic (MS) monitoring data and advanced machine learning techniques.
Introduction to Rockburst and MS Monitoring
Rockburst is a sudden and violent failure in hard rocks under high-stress conditions, often referred to as the ‘cancer’ of deep mining due to its unpredictable nature and potential for catastrophic consequences. MS monitoring has emerged as a critical tool for investigating stress evolution and early warning systems for geological hazards like rockbursts.
Methodology
The research methodology involved several key steps:
- Data Collection and Preprocessing: MS monitoring data was collected from the Dahongshan Copper Mine, with a total of 3729 events recorded between December 2022 and April 2023. The data was preprocessed to identify relevant parameters for MS event classification.
- Addressing Class Imbalance: The dataset suffered from severe class imbalance, with MS events being significantly outnumbered by other events. The SMOTE-ENN algorithm was employed to balance the dataset, enhancing the minority class representation.
- Ensemble Learning for MS Event Identification: Nine machine learning algorithms were evaluated as base learners, with stacking and voting ensemble techniques used to create more robust models. The V-S model, combining stacking and voting ensemble learning, demonstrated superior performance with an F1 score of 0.9819 and an AUC value of 0.9989.
- Rockburst Early Warning Model: A rockburst early warning model (V-soft) was developed using ensemble learning strategies, incorporating six MS monitoring parameters. The model achieved peak accuracy and F1 scores of 0.9394 and 0.9173, respectively, outperforming conventional machine learning algorithms.
Results and Discussion
The study’s findings indicate that the proposed ensemble learning framework significantly enhances the accuracy and stability of rockburst prediction. The V-soft model demonstrated exceptional performance in predicting strong rockbursts, achieving a 100% accuracy rate for this critical category.
Conclusion
This research effectively addresses the challenges of MS event identification and rockburst prediction through the integration of SMOTE-ENN and ensemble learning techniques. The developed short-term risk assessment system enables automated 24-hour risk quantification, demonstrating strong engineering applicability and scalability. Future research directions include advancing full-chain automation in data preprocessing and refining composite sample balancing frameworks.
