Particle peak velocity prediction based on risk-oriented hybrid ensemble learning.

Area of Science:

• Engineering
• Machine Learning
• Risk Assessment

Background:

• Accurate peak particle velocity (PPV) prediction is critical for structural safety in blasting engineering.
• Symmetric loss functions in machine learning inadequately address the risks of underestimation in safety-critical applications.

Purpose of the Study:

• To develop a risk-oriented hybrid ensemble model for enhanced PPV prediction accuracy and safety.
• To implement an asymmetric safety assessment system to mitigate hazardous underestimations.

Main Methods:

• Utilized a stacking ensemble framework integrating LightGBM, XGBoost, and CatBoost gradient-boosting models.
• Employed Bayesian Optimisation (BO), Grey Wolf Optimiser (GWO), and Particle Swarm Optimisation (PSO) for hyperparameter tuning.
• Introduced an asymmetric weighted mean squared error (W-MSE) and hazardous low-estimation rate (HLR) for performance evaluation.

Main Results:

• The proposed hybrid ensemble model demonstrated strong overall prediction performance.
• The model effectively suppressed hazardous underestimations, significantly enhancing safety and reliability.
• The integrated model showed clear advantages in PPV prediction compared to traditional methods.

Conclusions:

• The developed model offers a reliable solution for PPV prediction in blasting engineering.
• It provides a reusable paradigm for incorporating engineering safety constraints into machine learning.
• The approach offers valuable technical support for safety planning and risk minimization in blasting projects.