A Hybrid-Statistical Machine Learning Model for Blast Vibration Prediction in Surface Mines

Abstract

Blast-induced ground vibration is a major environmental and operational constraint in surface mining, and Peak Particle Velocity (PPV) is the principal response used for vibration assessment. This study develops a physically informed data-driven framework for PPV prediction using controlled-blast records, conventional statistical modelling, and supervised machine learning. The response was the maximum PPV recorded from three orthogonal vibration components. Raw predictors represented blast geometry, explosive loading, initiation timing, monitoring distance, and rock-breakage performance; additional ratios and scaled distance were introduced through feature engineering. After data cleaning and exploratory analysis, the observations were divided into 75% training and 25% independent testing subsets, with cross-validation applied within the training data. Linear Regression, Random Forest, and an Artificial Neural Network (ANN) were compared using R², mean absolute error, and root mean squared error. The ANN produced the strongest testing performance (R² = 0.9337, MAE = 0.5650 mm/s, RMSE = 0.7328 mm/s), outperforming Linear Regression and Random Forest. Correlation and feature-importance results identified monitoring distance as the dominant vibration-control variable, with total explosive quantity, powder factor, and engineered distance-related features also contributing. The results support ANN-based site-specific PPV estimation while retaining interpretable statistical and ensemble models for engineering checks.