Research on Piper-PCA-Bayes-LOOCV discrimination model of water inrush source in mines

Abstract

The occurrence of mine water inrush constantly threatens the safety production of coal mines and causes serious financial losses in China. Existing water inrush source identification methods do not consider the mixing effect of aquifers and the complexity of discrimination indexes. To identify water inrush rapidly and accurately, an identification model that combines water chemistry and multivariate statistical methods is proposed. The Piper trilinear diagram was used to screen 48 water samples taken from a water inrush aquifer in a mining area. Forty-two typical water samples that represent the water inrush aquifers were obtained. They were selected as training samples with discrimination indexes (Ca2+, Mg2+, Na++K+, HCO3−, SO42−, Cl−). Principal component analysis (PCA) was used to extract three principal indexes. Then, with these indexes taken as factors of Bayesian discrimination, we established a model for determining the source of water bursting in Pingdingshan Mine. Finally, the prejudgment classification stability of the constructed model was evaluated by leave-one-out cross-validation (LOOCV), which reports 95.2% overall classification accuracy. The constructed model was used to obtain a prognosis of 10 samples collected from Pingdingshan Mine, and it reported one misjudgment on one sample. In addition to comparing the prediction results with fuzzy comprehensive evaluation and gray relational degree model, results indicate that the constructed Piper-PCA-Bayes-LOOCV discrimination model of the water inrush source in mines can increase the recognition accuracy effectively, thus guaranteeing the safety production of mines.