Predictive performance of machine learning algorithms for ore reserve estimation in sparse and imprecise data.

摘要

Traditional geostatistical estimation techniques have been used predominantly in the mining industry for the purpose of ore reserve estimation. Determination of mineral reserve has always posed considerable challenge to mining engineers due to geological complexities that are generally associated with the phenomenon of ore body formation. Considerable research over the years has resulted in the development of a number of state-of-the-art methods for the task of predictive spatial mapping such as ore reserve estimation. Recent advances in the use of the machine learning algorithms (MLA) have provided a new approach to solve the age-old problem. Therefore, this thesis is focused on the use of two MLA, viz. the neural network (NN) and support vector machine (SVM), for the purpose of ore reserve estimation. Application of the MLA have been elaborated with two complex drill hole datasets. The first dataset is a placer gold drill hole data characterized by high degree of spatial variability, sparseness and noise while the second dataset is obtained from a continuous lode deposit.; The application and success of the models developed using these MLA for the purpose of ore reserve estimation depends to a large extent on the data subsets on which they are trained and subsequently on the selection of the appropriate model parameters. The model data subsets obtained by random data division are not desirable in sparse data conditions as it usually results in statistically dissimilar subsets, thereby reducing their applicability. Therefore, an ideal technique for data subdivision has been suggested in the thesis. Additionally, issues pertaining to the optimum model development have also been discussed.; To investigate the accuracy and the applicability of the MLA for ore reserve estimation, their generalization ability was compared with the geostatistical ordinary kriging (OK) method. The analysis of Mean Square Error (MSE), Mean Absolute Error (MAE), Mean Error (ME) and the coefficient of determination (R2) as the indices of the model performance indicated that they may significantly improve the predictive ability and thereby reduce the inherent risk in ore reserve estimation.