Characterisation of blastability variation within drill holes using a recurrent neural network with reduced supervision,generative transformer assisted training

Abstract

With the increasing availability of computational resources, machine learning (ML) has become a significant and rapidly growing technology. By leveraging geological uncertainties and machine learning techniques, drilling and blasting can be re-focussed from a bulk mining operation to more selective, precise and efficient extraction for ore preconcentration techniques and mine to mill optimisation in critical metal mining and/or lump-to-fine optimisation for iron ore extraction. While ML can have a notable impact on open-pit drilling and blasting, training ML models with small exploration datasets or noisy production data is challenging. To leverage the often limited and noisy data with the aim of improving the accuracy of ML models, this paper presents a generative transformer (i.e. TTS-CGAN)-assisted recurrent neural network (RNN) methodology to better understand the variability of blastability index (BI) extracted from the fusion of measurement while drilling (MWD) data (strength and fracture percentage) with rock density from assay information, on a bench scale. An implementation of the proposed method at a platinum mine and an iron ore deposit shows that mixing a small amount of augmented data with real data is beneficial for RNN performance. However, an optimal point exists as the addition of too much synthetic data may introduce additional noise.