Neural Network-Based Modeling of Ground Surface Subsidence in Underground Mining of Potassium-Magnesium Salts
A. I. Manevich1,2, A. D. Gvishiani1,3, A. A. Baryakh4, V. N. Tatarinov1,2,3, A. A. Kamaev1,2, A. V. Evseev4
1Geophysical Center, Russian Academy of Sciences, Moscow, Russia
2National University of Science and Technology-NUST MISIS, Moscow, Russia
3Schmidt Institute of Physics of the Earth, Russian Academy of Sciences, Moscow, Russia
4Mining Institute, Ural Branch, Russian Academy of Sciences, Perm, Russia
Keywords: Neural network-based modeling, ground surface subsidence, selection of attributes, regression, potassium-magnesium salt deposit, impermeable strata, Verkhnekamsk deposit
Abstract
The authors describe application attribute-based interpretable neural network models which allow optimization of risk management in underground mining of potassium-magnesium salts. The proposed procedure makes it possible: to classify initial natural data and manmade factors with the priority of transformation to a continuous representation, and with introduction of quality classes of sets; to build dimensionless indexes-attributes which ensure physical interpretability of models; to perform integrated selection of attributes on the basis of the single-factor linear analysis, correlation filtration, multiple regression and nonlinear analysis of a learned artificial neural network; to learn and validate neural network models with interpretation of contributions of attributes. For the transition from subsidence prediction to risk-oriented prediction, an integrated index of subsidence, inclination, curvature and accumulated damage parameter, with data mapping is proposed. The procedure is tested at Berezniki-1 Mine at the Verkhnekamsk deposit of potassium-magnesium salts.
|
