Risk assessment of water inrush from coal floor based on enhanced samples with class distribution

In the risk assessment of water inrush from coal floors, the amount of measured data obtained through on-site testing is small and random, which limits the prediction accuracy and generalizability of a model based on measured data. Using the distribution characteristics of the measured data and mega-trend diffusion theory, we propose a virtual sample enhancement method based on class distribution mega-trend diffusion technology (CDMTD) and introduce constraints on the class distribution of influencing factors. This method was used to generate virtual samples and enhance the measured database. A prediction model of the water inrush risk for the coal seam floor was established using a coupled algorithm of extreme learning machines, self-adaptive differential evolution, and CDMTD (PCA-CDMTD-SaDE-ELM) and was used to evaluate the water inrush risk in the 19,105 working face of the Yunjialing Mine. The CDMTD method could effectively solve the problem of virtual sample distribution variation in the overall trend diffusion theory and enhance the measured database, reducing the impact of small sample sizes. Compared to other optimization models, our model showed the best prediction performance, with an error reduction of 42.95-51.27% and results biased towards safety. Our results support safe and efficient coal mining above Ordovician limestone-confined water.