Prediction of Probability and Time of Ground Subsidence above Mines Using Neural Network
V. A. Driban, N. A. Dubrova, V. V. Trofimov, A. I. Yarkova, A. V. Antipenko
Republican Academic R&D Institute of Mining Geology, Geomechanics, Geophysics and Surveying-RANIMI, Donetsk, Russia
Keywords: Ground surface subsidence, mine workings, survival analysis, neural networks, Deep Survival Machnines, censored data, event occurrence time, risk prediction, genetic algorithms
Abstract
The authors focus on probability and time prediction of ground subsidence in the influence zones of underground mining. The relevance of the research is conditioned by limitedness of classical empirical and numerical methods for the assessment of mining safety in complex geological conditions, by deficiency of initial data and by the need of accounting for censored observations, i.e. events of zero ground subsidence in the observation periods. The main tool was the Deep Survival Machines model. Hyperparameters of the model were selected using genetic algorithms. The proposed model allows interpolation of ground subsidence time beyond the limits of a learning sampling and demonstrates an acceptable prediction accuracy (the average absolute error is 10-12 years). The time of occurrence of an event is systematically overestimated due to disbalance of data. two approaches are proposed for eliminating the overestimation: quantile-based interpretation of the survival function and forecast correction based on residual analysis. Modifications of the model allow a substantial reduction of the overestimation risk in the occurrence time of a ground subsidence event at the preserved acceptable accuracy. The approach can be used in estimation of underground mine stability and for decision-making on industrial safety.
|
