Prediction of water inrush from coal seam floor based on machine learning with small sample data

With the development of computer technology, machine learning method has become an important technology for the prediction of water inrush in coal seam floor. However, the prediction accuracy of many machine learning algorithms requires a high number of samples, which restricts the practical application. In this paper, by using the nearest neighbor algorithm (KNN) and the combination algorithm of gradient boosting decision tree (GBDT) and logistic regression (LR), a water inrush prediction model was established based on the sample data of six indexes, including water pressure, mining height, water-barrier thickness, fault drop, coal seam inclination, and fault distance from the working face. The influence rule of sample number on prediction accuracy was discussed, and the comparison study was conducted with the commonly used particle swarm, support vector machine, BP neural network, random forest and convolutional neural network. The results show that when the number of samples reaches 18, the prediction accuracy of KNN and GBDT+LR remains stable. The prediction accuracy of KNN and GBDT+LR is higher than that of conventional models under small sample conditions. The predicted results of the model agree with the actual situation.