Optimizing Flyrock Forecasting in Open-Pit Blasting Using Hybrid Machine Learning Models

doi:10.1007/s00603-025-04730-2

Optimizing Flyrock Forecasting in Open-Pit Blasting Using Hybrid Machine Learning Models

2025-07-15

0

PRE

AI

Danial Jahed Armaghani

J

Jian Zhou *

DOI：10.1007/s00603-025-04730-2

Original

Original request for help

Save

Abstract

En 中文

Flyrock occurrences pose significant environmental and safety risks in open-pit blasting, threatening equipment, personnel, and potentially causing fatal accidents. This study intends to reduce the risk of flyrock by developing multiple machine learning approaches to estimate its throw range, utilizing Random Forest (RF) as the primary model. Two optimization algorithms, the Light Spectrum Optimizer (LSO) and the Puma Optimizer Algorithm (POA), were employed to fine-tune the hyperparameters of the RF model. Researchers conducted field surveys of blasting events in quarries in Malaysia and the Ayazaga area of Istanbul to compile two databases. Based on these datasets, the reliability and practicality of the LSO-RF and POA-RF models were examined. For comparison, an unoptimized RF model was also created to estimate the flyrock throw distance. The performance of the models was assessed using four metrics: root mean square error (RMSE), mean absolute percentage error (MAPE), coefficient of determination (R2), and variance accounted for (VAF). Results show that LSO-RF and POA-RF outperform the unoptimized RF model across all metrics. For the Malaysian dataset, LSO-RF achieves training/testing R2 of 0.9661/0.9597, MAPE of 4.445/6.711, and RMSE of 12.081/16.091; POA-RF yields training/testing R2 of 0.9662/0.9596, MAPE of 4.443/6.717, and RMSE of 12.075/16.094. For the Istanbul dataset, LSO-RF achieves training/testing R2 of 0.9706/0.9698, MAPE of 5.369/6.512, and RMSE of 6.128/6.537; POA-RF yields training/testing R2 of 0.9725/0.9680, MAPE of 6.006/7.554, and RMSE of 6.087/7.335. Performance differences between hybrid models are within 0.2% across datasets, indicating consistent reliability. Results from the engineering case validation demonstrate that the developed hybrid model exhibits strong generalization performance, with R2 of 0.8778. SHAP-based sensitivity analysis identifies key parameters: hole diameter (HD) and maximum charge per delay (CD) for Malaysia, specific charge (q) and spacing-to-burden ratio (S/B) for Istanbul, all positively correlated with FD. An intelligent software tool integrating these models supports data-driven blasting parameter optimization, enhancing safety and efficiency by minimizing flyrock risks.

Keywords:

Flyrock distance prediction

Open-pit blasting optimization

Light spectrum optimizer

Puma optimizer algorithm

Random forest

Interpretability analysis

Intelligent feedback system

Optimizing Flyrock Forecasting in Open-Pit Blasting Using Hybrid Machine Learning Models

Abstract

Journal

Organization