Оценка качества технической воды в горно-промышленном производстве на основе методов машинного обучения

Abstract

Introduction. Mining requires water treatment and wastewater processing, abstraction and discharge during mining increases consumption several times. Since water consumption in mining and processing is usually associated with domestic, industrial and technical needs, the need for water supply systems required for water treatment increases. Water from different sources can be used for treatment: incoming water, process and reused water, and wastewater. But the water obtained from any of the sources must meet all the norms and requirements. Water quality is determined by physical, chemical and bacteriological properties. The main directions for improving water consumption by mining enterprises are to reduce the consumption of drinking water from rivers, lakes and municipal water supply, as well as to expand the use of mine and quarry water for domestic and technical needs. Materials and methods. As training data for training the neural network, a dataset that includes water quality data obtained from fresh water sources was selected for the methods work, and using machine learning, develops a model that predicts whether the water is suitable for technical use in mines. This dataset includes 2293 values (samples) as well as 9 attributes. Correlation, neural network, and decision tree methods were used to build the models in this study. Results. Various machine learning methods (neural network and decision trees) were used to build a predictive model to assess the quality of water that would be suitable for use in the mining industry for technical purposes. With the help of the built models were processed data obtained from public sources, when analyzing which it was found that the method of decision trees was more accurate. The constructed model, for determining dependencies, thus, has high accuracy (small error). To increase the practical significance of the study, a number of transformations of the initial data set were carried out, in particular, an experiment with the division of attributes into groups of importance, in relation to the data, taking into account the subject area. The results obtained made it clear that checking only for hazardous impurities does not guarantee the suitability of water, but almost completely excludes (low significance factor) samples with impurities that do not meet the requirements, and the model can have practical significance. Allocation of the group for rapid quality determination, showed that for the express test, in an emergency situation or under time constraints, the possibility of practical use of the obtained model, has a justification, due to the small error. In general, the conducted experiments have shown that when taking into account the costs (total) for data collection, it makes sense to use models, taking into account the reduction of collected data, on the parameters (factors) of technical water. Discussion. In general, on the basis of the conducted research, we can talk about the successful application of machine learning methods in determining the suitability of technical water in the mining industry. During the experiments, the decision tree method performed particularly well, with the lowest error values. In addition, further work can be carried out to reduce the error in the models, in particular, by possibly increasing the number of attributes, as well as more fine-tuning of the applied machine learning methods. Conclusions. The authors conclude that machine learning techniques can be successfully integrated to determine the quality and suitability of process water in the mining industry in today’s world. Resume. The paper compares machine learning methods such as decision trees and neural network method. The comparative analysis of these methods and their quality of information processing is shown on the example of a set of data on water quality in the mining industry. With the help of built models were processed data obtained from open sources, when analyzing which it was found that the method of decision trees was more accurate. The constructed model for determining dependencies has high accuracy (small error). Suggestions for practical applications and future research directions. This study can form the basis for research in this or related fields to conduct further studies on the reliability and accuracy of using machine learning to predict the quality of water used in the mining industry. Continued work in the above direction may be the rationale for wider use of the above methods to improve various meaningful production performance in this or related areas.