Project on creating a classifier of lithological types for uranium deposits in Kazakhstan

Abstract

A program that provides the task of classifying electric log data using machine learning methods is described. The program, after the training phase, allows one to automatically determine the composition of the rocks along the wellbore, which is necessary to ensure the technological process of uranium mining. The applied methods of data preprocessing, methods of forming a "floating window" of data and some results are briefly described. When using multilayer perceptron, the obtained average values of precision, recall, f1-score with recognition of 7 rock classes are 49%.At the same time, the developed program allows one to apply a wide range of classification algorithms, including Nearest-Neighbor (k-NN)), Logistic regression, Decision Tree, Support Vector Classifier, artificial neural network, busting, LSTM, etc. For example, when using XGBoost on same data with a change in the size of the floating window and the weight of individual classes (rocks), the indicated accuracy metric was up to 54%. The results of a comparative analysis of the mentioned methods on an extended data set will be presented in the next article of the authors.