Deep learning model based on big data for water source discrimination in an underground multiaquifer coal mine

In view of the difficulty and low efficiency of models established with big data, a deep feedforward network-based mine water inrush discriminant model is established. The stochastic gradient descent (SGD) algorithm is employed to optimize model parameter training. The cross-entropy loss and accuracy of the training and test samples are considered to characterize the model training effect. To prevent overfitting during model training, the dropout optimization method is incorporated into the hidden layers of the model. The four main water-filled aquifers in the Huainan Panxie mining area are chosen as discrimination objects, 1952 sets of water samples are applied as modeling data, and a total of nine indicators, including K+  + Na+, Ca2+, Mg2+, Cl−, SO42−, HCO3−, CO32−, pH, and total dissolved solids (TDS), are selected as discriminating factors. Through model training, the dropout rate, epoch number, batch size, and learning rate are set as 0.1, 90, 25, and 10−2, respectively, and the discriminant accuracy rates for the training and test samples are 93.34% and 96.68%, respectively. The trained model is applied to discriminate 30 groups of water samples, and the discrimination results for 28 groups of water samples are consistent with observations. The results indicate that the deep learning discriminant model based on big data achieves high accuracy, good applicability, and a suitable discrimination ability, and provides a certain guiding significance in the prevention and control of water inrush hazards and related field work.