Water Inrush Source Identification In Panxie Mining Area Based On Deep Learning

The hydrogeological condition of coal mine in China is very complicated,and the water gush disaster is frequent.Once the mine water inrush disaster occurs,not only the production efficiency will be reduced,but even the well will be flooded,which will cause significant economic loss and affect the life safety of the mine workers.It is of great significance for mine water prevention and control to identify the source of water burst accurately.This paper takes the main water-filled aquifers in Huainan Panxie mining area as the research object.Through water quality data collection,sample collection and test analysis,the water chemical characteristics of single aquifers such as Cenozoic loose layer water samples、Permian Sandstone water samples、Carboniferous Taiyuan formation limestone water samples、Ordovician limestone water are studied.The main findings of the paper are as follows:（1）The water samples in the study area have the highest cation content in Na⁺+K⁺,and most of the anion content in the water samples are Cl^-.The distribution trend of various indicators in the aquifer is generally similar.Permian Sandstone water shows the characteristics of high Na⁺+K⁺and HCO₃^-content.Most of the water chemistry types of Cenozoic loose layer water and Ordovician limestone water are Cl-Na,Cl·SO₄-Na;The highγNa⁺/γCl^-in Permian sandstone water may be affected by human activities.Theγ(Ca²⁺+Mg²⁺)/γHCO₃^-value of Cenozoic loose layer and Permian sandstone water is relatively low;theγ(Ca²⁺+Mg²⁺)/γHCO₃^-value of the Ordovician limestone water is relatively close to 1,indicating that the source of the cations Ca²⁺and Mg²⁺in the water sample may be related to the dissolution of salts and minerals.Theγ(Ca²⁺+Mg²⁺/HCO₃^-)/γ(SO₄^2-/HCO₃^-)of the all aquifers are close to 1,indicating that carbonic acid participates more in the dissolution process of carbonate minerals.（2）Established a deep learning water source discrimination model.Based on the Keras framework,the Re LU function is selected as the model activation function,the Softmax classifier and the cross-entropy loss function to build the depth model,the stochastic descent gradient algorithm is used for model training,and the Dropout method is used to prevent the model from overfitting.（3）The computationally determined model includes 3 hidden layers,20 hidden layer nodes,and 5 output layer nodes.Comparing the output results of the model,it is concluded that when Dropout is set to 0.5,epoch is set to 90,batch＿size is set to 2⁵,and the learning rate is set to 10^-2,the output of the model is optimal.The combined weight-improved grey correlation theory model and the PCA-Fisher water source recognition model are established,and the recognition results of the two models are compared with the deep learning water source recognition results.The results show that the deep learning model has the highest recognition accuracy,with an accuracy rate of 83.33%;the PCA-Fisher water source recognition model with an accuracy rate of 53.33%;and the combination weight-improved grey correlation theory model has the lowest accuracy with 36.67%.