A Gaussian process machine learning model for cemented rockfill strength prediction at a diamond mine

Abstract

As a main strategy of backfilling in mining operations, cemented rockfill (CRF) is extensively used because of its high strength and mine waste disposal convenience. The CRF strength has a direct bearing on ground support performance in backfill mining, which necessitates investigating CRF strength determination. This study employed a Gaussian process (GP) machine learning model to reflect the relationship between CRF compressive strength and material components as well as curing age. More than one thousand data from a public database were used to train the GP model with an automatic hyperparameter optimization. A series of laboratory tests prepared eight test samples for our predicting as well as the true values for model validation. The GP model achieved a predicting accuracy with the r2 value 0.90 and the MSE value 7.78 based on CRF true values we obtained in the laboratory. In addition, seven test samples’ true values resided inside the 95% confidence interval of the GP prediction. We also constructed three other machine learning models to conduct the same work as the GP model did. The results showed that the GP model performed the best of four models, which demonstrated that the GP model was effective and robust in dealing with time series predicting task.