Managing Production Blasting Vibrations: A Machine Learning Based Prediction of PPV in Chilean Copper Mines

This work has addressed the problem of creating a predictive model of the maximum particle velocities in the rock mass (PPV). The model designed specifically can model the subsequent vibrations of blasting based on geometric parameters, explosive loading, and the blasting process itself. Models based on artificial intelligence (AI) algorithms can find useful non-linear patterns to solve classification, regression, clustering, among other problems. They are flexible mathematical transformations adaptable to various objectives and types of data. With their different levels of abstraction, they can find a function that relates the input variables to the sought-after response. These models have been successfully used in similar contexts (Prashanth, (2018)) and have achieved state-of-the-art results in the field (Lawal, (2021)). The main challenge is to adapt them to the context of Chilean mines, in addition to including the geometric information of the blasting and the heuristic to obtain the set of optimal parameters, it is proposed to follow the work of (Simonyan, (2013)) and obtain "saliency maps" adapting it to tabular variables. We implemented an Artificial Neural Network (ANN) with a large number of neurons in the hidden layers. Models based on neural networks can find useful non-linear patterns to solve classification, regression, clustering, among other problems. They are flexible mathematical transformations adaptable to various objectives and types of data. With their different levels of abstraction, they can find a function that relates the input variables to the sought-after response. The literature shows that, for the problem of rock type identification, basic neural networks perform well (Simonyan, (2013)), but lack uncertainty estimates. This is a data-driven model, and in this sense, the model training was carried out using data from multiple operations in the large copper mining industry in Chile. In these databases, data consistency and the balance of the various independent variables that have the greatest impact on the model were ensured. With the above, the optimal hyperparameters of the model were determined, obtaining a predictive R-squared of 0.94.