An Optimum Prediction Model for the Strength Index of Unclassified Tailings Filling Body

Jian Yao, Shenghua Yin, Dongmei Tian, Yi Chen, Jinglin Xu, Leiming WangProcesses 2025 · DOI

TL;DR

This paper introduces an optimized Support Vector Machine (SVM) model for predicting the strength of unclassified tailings filling bodies, a critical factor in deep mining operations. The SVM model significantly outperforms traditional Backpropagation (BP) neural networks, achieving a mean prediction error of just 1.01%, leading to potential reductions in cement consumption and operational costs while ensuring stope stability.

What Problem Does This Paper Address?

As accessible mineral resources dwindle, mining operations are increasingly shifting to deeper environments. This necessitates robust backfill materials to ensure stope stability, especially in methods like non-pillar backfill mining where the backfill might be exposed during subsequent mining phases. Accurately predicting the required strength of these backfill bodies is crucial. Traditional methods like empirical formulas and physical tests are often time-consuming, costly, and may not adequately capture the complex, multi-dimensional factors influencing backfill strength. Existing predictive models, including some AI-based approaches, have limitations in accuracy, especially with the often limited datasets available in this domain. This paper aims to address this by developing a more accurate and efficient prediction model for backfill strength.

Methods

The core of this research lies in developing an optimized Support Vector Machine (SVM) model. The authors selected eight key factors influencing backfill strength as input attributes, including geological conditions (buried depth, thickness, strike length, ore body stability), material properties (tailings fractal dimension), backfill body stability (reliability index), and operational parameters (height and exposed area of the backfill body). The decision attribute was the backfill strength itself.

The SVM model utilized a Radial Basis Function (RBF) as its kernel function, chosen for its theoretical soundness, ease of implementation, and ability to handle nonlinear data. Parameter optimization was performed using a grid search method to find the optimal combination of the kernel function's parameters (g and C). The model was trained on 72 groups of data and validated on 6 additional groups. For comparison, a Backpropagation (BP) neural network model was also developed and evaluated.

Datasets

The study utilized a dataset comprising 78 groups of domestic and overseas backfill strength design data. Of these, 72 groups were used for training the SVM and BP models, and 6 groups were reserved for calibration and testing. The paper provides a partial table (Table 1) showcasing some of these data points, with eight condition attributes and one decision attribute (strength in T/MPa). The data was normalized using a maximum-minimum normalization method to ensure comparability across different attribute scales.

Key Insights and Ideas

  • SVM's Superiority in Small-Sample Learning: The paper strongly advocates for SVMs in scenarios with limited data, highlighting their ability to achieve good generalization even with fewer training samples compared to other methods like BP neural networks.
  • RBF Kernel's Effectiveness: The choice of the RBF kernel function proved effective in capturing the complex, nonlinear relationships between the input factors and the backfill strength.
  • Grid Search for Robust Optimization: Employing a grid search method for parameter tuning (g and C) allowed for a systematic exploration of the parameter space, leading to a more robust and optimized model.
  • Economic Implications: Beyond mere prediction, the study emphasizes the practical economic benefits of an accurate prediction model, specifically the potential to reduce cement consumption and overall backfill costs while maintaining safety.

Key Results

The developed SVM model demonstrated superior performance compared to the BP neural network:

  • Mean Prediction Error: The SVM model achieved a mean error of 1.01% for both regression prediction and verified predictions, significantly lower than the BP model's 5.26%.
  • Maximum Error: The SVM model's maximum error was 4.97%, while the BP model's was 10.17%.
  • Minimum Error: The SVM model achieved a minimum error of 0.12%, substantially better than the BP model's 1.49%.
  • Calibration Results: In calibration tests, the SVM model showed a maximum error of 3.546% and an average error of 2.232%, indicating high prediction accuracy and stability across diverse samples.
  • Practical Application: A case study in a gold mine demonstrated the model's ability to predict required backfill strengths (0.81 MPa and 0.97 MPa) that met specifications, providing reliable guidance for mine backfill design.

Limitations and Future Work

The authors acknowledge that while the SVM model shows promise, further research is needed to expand the dataset scope to include a wider range of geological conditions and mining environments. This expansion is expected to further enhance the model's accuracy and generalization capabilities.

Future work directions suggested include:

  • Incorporating More Data: Expanding the dataset to cover more diverse mining scenarios.
  • Advanced Machine Learning: Exploring deep learning and ensemble learning methods for even greater predictive power.
  • Algorithm Integration: Developing hybrid models that combine the strengths of multiple algorithms for more intelligent and efficient predictive systems.

Summary updated

Abstract

In order to improve the poor prediction effect of current filling body strength design, a support vector machine (SVM) and Lib Toolbox were used to build an optimal match model or strength index of unclassified tailings filling body. Eight main factors were analyzed and screened as condition attributes, and backfill strength as a decision attribute. Next, we selected 72 groups of training samples and 6 groups of calibration samples. Our model adopts a radial basis function (RBF) as the kernel function and uses a grid search method to optimize parameters; it then tests the combination of optimal parameters by cross-validation. Results show that the mean error of regression prediction and verified predictions made by the SVM match model were 1.01%, which were more accurate than the BP neural network model’s predictions. On the premise that stope stability is ensured, the SVM match model may decrease cement consumption and the cost of backfill more effectively, and improve economic efficiency.
BibTeX
@inproceedings{yao2025optimum,
  author = {Jian Yao and Shenghua Yin and Dongmei Tian and Yi Chen and Jinglin Xu and Leiming Wang},
  title = {{An Optimum Prediction Model for the Strength Index of Unclassified Tailings Filling Body}},
  booktitle = {Processes},
  year = {2025},
  doi = {10.3390/pr13082395},
}
Figure