One method of generating synthetic data to assess the upper limit of machine learning algorithms performance. Issue 1 (1st January 2020)

• Record Type: Journal Article
• Authors: Kuchin, Yan I.; Mukhamediev, Ravil I.; Yakunin, Kirill O.
• Editors: Pham, Duc
• Is Part Of: Cogent engineering. Volume 7:Issue 1(2020)
• Publication Date: 2020-01-01
• Subjects: Machine learning (ML) -- Supervised machine learning (SML) -- simulated borehole -- accuracy -- precision -- recall -- artificial neural network (ANN) -- SVM -- XGBOOST -- LSTM; Engineering -- Periodicals; Technology -- Periodicals; Engineering; Technology; Periodicals; 620
• DOI: 10.1080/23311916.2020.1718821
• Languages: English
• ISSNs: 2331-1916

Abstract

Abstract: Based on statistics from the World Nuclear Association, Kazakhstan has the highest uranium production in the world. Most of the uranium in the country is mined via in-situ leaching and the accurate classification of lithologic composition using electric logging data is economically crucial for this type of mining. In general, this classification is done manually, which is both inefficient and erroneous. Information technology tools, such as predictive analytics with Supervised Machine Learning (SML) algorithms and Artificial Neural Networks (ANN) models, are nowadays widely used to automate geophysical processes, but little is known about their application for uranium mines. Previous experiments showed an ANN accuracy of about 60% in the task of lithological interpretation of logging data. To determine the upper limit of the accuracy of machine learning algorithms in such task and for indirect assessment of the experts' influence, a digital borehole model was developed. This made it possible to generate a complete set of data avoiding subjective expert assessments. Using these data, the work of various ML algorithms, both simple (kNN) and deep learning models (LSTM), was evaluated.