An empirical-driven machine learning (EDML) approach to predict PPV caused by quarry blasting

Abstract

Blasting in mining and quarrying serves multiple purposes but poses environmental challenges, notably generating shockwaves and vibrations through peak particle velocity (PPV) from explosions. Previous efforts to predict PPV values have relied on empirical equations using parameters such as maximum charge per delay (MC) and distance from the blast face (D). Numerous attempts have employed machine learning (ML) to estimate PPV with the same input parameters. This study introduces a novel approach called empirical-driven ML (EDML), which integrates empirical equations and their outcomes as inputs for PPV prediction. EDML leverages existing knowledge to enhance model performance, interpretability, and generalization. For the EDML approach, four empirical equations, namely USBM, CMRI, General Predictor, and Ambraseys-Hendron have been chosen based on prior research. These four empirical equations were selected based on their good performance as reported in the literature. Using these equations’ PPV values as inputs, three advanced tree-based techniques (random forest, deep forest, and extreme gradient boosting) have been employed for model training. Comparison with the conventional ML approach (using only maximum charge per delay and distance from the blast face) reveals EDML’s superior predictive capacity for PPV estimation. Note that the inputs of these databases were directly and indirectly extracted from MC and D with the same PPV values. The proposed EDML approach effectively integrates data-driven insights with domain expertise, improving accuracy and interpretability through the inclusion of PPV and blasting observations.