A Machine-Learning-Based Method for Identifying the Failure Risk State of Fissured Sandstone under Water-Rock Interaction

Jinbo Qu, Cheng Song, Jinwen Bai, Guorui Feng, Xudong Shi, Junbiao Ma2024 · DOI

TL;DR

This paper introduces a machine learning approach to predict the instability risk of fissured sandstone under water-rock interaction, a critical issue in underground mining. By analyzing acoustic emission (AE) data, the study developed a Random Forest model that achieved an impressive 97.87% accuracy in identifying different risk states, offering a more efficient and accurate alternative to traditional methods.

What Problem Does This Paper Address?

Underground mining operations frequently encounter fissured rock masses, where the interaction with water can significantly degrade rock strength and lead to instability. Identifying precursor information for this instability is crucial for preventing accidents and ensuring safety. While acoustic emission (AE) monitoring is a valuable non-destructive technique for detecting rock deterioration, existing studies often focus on limited AE parameters or lack integration with advanced analytical methods for complex scenarios like water-rock interaction. This paper aims to bridge this gap by developing a robust machine learning model that leverages multiple AE parameters to accurately assess the instability risk of fissured sandstone under both natural and water-saturated conditions.

Methods

The study employed a multi-pronged approach:

  • Acoustic Emission (AE) Data Acquisition: Uniaxial compression tests were conducted on fissured sandstone specimens under both natural and water-saturated conditions. AE signals were collected throughout the entire damage process.
  • Feature Engineering: Several AE parameters were extracted and analyzed, including ringing count, energy, centroid frequency, peak frequency, Rise Angle (RA), Average Frequency (AF), and the b-value. The "state" of the sandstone (natural or saturated) was also included as a feature.
  • Machine Learning Models: Three popular machine learning algorithms were chosen for their complementary strengths:
  • Random Forest (RF): An ensemble method known for its robustness against overfitting and ability to handle high-dimensional data.
  • Multilayer Perceptron (MLP): A deep learning neural network capable of capturing complex non-linear relationships.
  • AdaBoost: An iterative ensemble method that combines weak learners to improve overall accuracy.
  • Model Training and Evaluation: A 10-fold cross-validation strategy was used to train and validate the models. Performance was evaluated using metrics like accuracy, precision, recall, and F1-score, with confusion matrices providing a detailed breakdown of classification results. Feature importance analysis was conducted to identify the most influential parameters.

Datasets

The study utilized a custom-built dataset comprising 585 acoustic emission data points collected from uniaxial compression tests on fissured sandstone specimens. This dataset was divided into:

  • 291 data points from natural condition specimens.
  • 294 data points from water-saturated condition specimens.

The data was categorized into three instability risk states:

  • Label 0: No instability risk (stages I and II of failure process).
  • Label 1: Medium instability risk (stage III).
  • Label 2: High instability risk (stages IV and V).

This dataset was split into an 80% training set and a 20% test set, with a further subdivision of the training set for validation.

Key Insights and Ideas

  • Water's Impact is Significant: The study highlights that water saturation leads to distinct changes in AE parameters like the b-value, indicating its crucial role in altering sandstone behavior and making it a vital feature for instability prediction.
  • b-Value as a Strong Indicator: The b-value, which reflects microfracture scale, emerged as a highly significant predictor of instability, particularly its sharp decline indicating accelerated crack propagation.
  • State Matters: The "state" of the sandstone (natural vs. saturated) was found to be one of the most influential features, underscoring the importance of considering environmental conditions.
  • Random Forest Dominance: The Random Forest model consistently outperformed MLP and AdaBoost, demonstrating its superior ability to handle the complexity and nuances of AE data for this specific task.
  • Feature Importance Reveals Key Drivers: The analysis revealed that the "state" and "b-value" together accounted for over 60% of the classification outcome, providing clear guidance on which parameters are most critical for accurate prediction.

Key Results

  • The Random Forest (RF) model achieved an accuracy of 97.87% in identifying the instability risk state of fissured sandstone.
  • The RF model also demonstrated high performance across different risk states, particularly in identifying medium and high instability risks.
  • Feature importance analysis indicated that "state" and "b-value" were the most significant predictors, contributing over 60% to the classification results.
  • Spearman correlation analysis showed that most selected features had relatively low correlations with each other, suggesting a reasonable feature selection.

Limitations and Future Work

  • Limited Parameter Scope: While the study analyzed several key AE parameters, the authors suggest that mining more AE statistical parameters through further feature engineering could enhance precursor information for instability prediction.
  • Model Complexity and Speed: For real-world applications, more advanced fusion algorithms could be explored to improve the speed of machine learning model training and prediction performance.
  • Generalizability: While the RF model showed excellent performance on the developed dataset, further validation on a broader range of fissured sandstone types and geological conditions would be beneficial to assess its generalizability.
  • Feature Engineering: The authors suggest that future research could focus on more effective feature engineering methods to reduce correlations between features with high correlation, potentially further improving model performance.

Summary updated

BibTeX
@misc{qu2024machine,
  author = {Jinbo Qu and Cheng Song and Jinwen Bai and Guorui Feng and Xudong Shi and Junbiao Ma},
  title = {{A Machine-Learning-Based Method for Identifying the Failure Risk State of Fissured Sandstone under Water-Rock Interaction}},
  year = {2024},
  doi = {10.3390/s24175752},
}
Figure