Deep Neural Network Model for Accurate Relocation of Mining-Induced Seismic Events

By Dr. Noopur JainReviewed by Laura ThomsonAug 14 2024

In a recent article published in the journal Applied Sciences, researchers presented the development of a deep neural network model designed to enhance the accuracy of seismic event relocation, particularly in the context of mining-induced seismicity. The research aims to create a model that predicts the locations of seismic events and estimates wave velocities based on P-wave arrival times recorded by seismometers.

Background

The accurate relocation of seismic events is crucial for understanding subsurface dynamics, particularly in regions affected by mining activities. As mining operations progress deeper underground, the potential for induced seismicity increases, posing significant safety and operational efficiency risks.

Traditional methods of seismic event relocation often rely on fixed velocity models, which can lead to inaccuracies due to the complex and heterogeneous nature of geological formations. These inaccuracies can hinder effective hazard assessment and risk management in mining environments. Recent advancements in machine learning, particularly deep learning techniques, offer promising solutions to these challenges.

The Current Study

The methodology employed in this study consists of three primary phases: data generation, model training, and model testing. Initially, synthetic data is generated using an assumed velocity model, which serves as the foundation for the training process. The authors created a layered velocity model with a depth range of 3000 meters, where each layer is assigned a constant velocity that increases linearly from the bottom to the top, ranging from 2000 m/s to 3000 m/s. Random variations of 50 m/s are introduced within each layer to simulate natural heterogeneity.

The model's training uses a deep neural network architecture to process the generated data. The authors employ the Adam optimizer and the Mean Squared Error (MSE) as the loss metric to evaluate model performance.

The learning rate is dynamically adjusted between 0.001 and 0.0001 to ensure efficient convergence during the training process. Each model undergoes 10,000 iterations, during which data is subjected to a forward and backward propagation pass. This iterative training approach allows the model to learn from the data progressively, with each iteration building upon the previous one.

A vital aspect of the methodology is the flexibility to incorporate new field recording data as it becomes available. This continuous training process enables the model to adapt to changing conditions and improve its predictive capabilities. The authors emphasize that each iteration of the model is independent, allowing for the application of diverse methodologies at each step.

Results and Discussion

The study demonstrates the effectiveness of the proposed model in accurately predicting seismic event locations and wave velocities. Various performance metrics, including error plots and histograms, were used to illustrate the distribution of prediction errors compared to actual source coordinates. The findings indicate that the model consistently outperforms previous iterations, showcasing the benefits of the iterative training approach.

The study compared four models using linear data and four models using scattered data, highlighting the robustness of their methodology. The results reveal that the model's performance improves with each iteration, regardless of the initial training conditions. This iterative enhancement is attributed to the scientifically sound methodologies employed at each step, contributing to the model's overall accuracy.

Furthermore, the study discusses the implications of real-time velocity and source inversion capabilities. While the training process may be time-consuming, the actual prediction phase is significantly faster, allowing for efficient monitoring of seismic activities. The authors emphasize the importance of incorporating accurate data collected during engineering and monitoring processes, as this real-world information enhances the model's predictive accuracy. The continuous training framework enables the model to evolve and adapt, resulting in a highly capable earthquake source relocation model tailored for specific sites.

Conclusion

In conclusion, the study presents a significant advancement in seismic monitoring by developing a deep neural network model for mining-induced seismicity.

The authors successfully demonstrate the feasibility of an iterative training approach for continuous model performance improvement. The model achieves high precision in predicting seismic event locations and wave velocities by leveraging synthetic data generation and flexible training methodologies.

This research extends beyond mining applications, offering valuable insights for engineering projects requiring accurate seismic monitoring. The proposed model's ability to adapt to changing conditions and incorporate real-time data positions it as a powerful tool for enhancing safety and efficiency in underground operations. Overall, this study contributes to the ongoing efforts to refine seismic monitoring techniques and underscores the potential of deep learning in geophysical research.