Abstract: Due to the interconnection of different aquifers in mines through water-conducting faults and fracture networks, water samples with similar chemical characteristics become mixed and overlap. Additionally, they are interfered with by irrelevant redundant data from the environment, which easily leads to misjudgment of water sample feature types and reduces the accuracy of identifying the source of water inrush. To address this issue, a method for identifying the source of water inrush in aquifer mines based on the IPFA algorithm and mean impact value （MIV） is proposed. The K-nearest neighbor （k-NN） method is employed to expand the acquisition of high-dimensional mutual information entropy among the chemical composition features of aquifer water sources, enabling a more comprehensive reflection of the correlations between features. As a result, features truly relevant to the identification of water inrush sources are selected, and the interference of redundant features is reduced. The IPFA algorithm is utilized to optimize the parameters of the Extreme Learning Machine （ELM）, thereby enhancing the model’s generalization ability and recognition accuracy. This approach prevents ELM from getting trapped in local optima and reduces misjudgment caused by the mixed and overlapping similar chemical characteristics of water samples. The mean impact value （MIV） method is applied to calculate the MIV values for various types of features, allowing for the selection of features with high average impact contribution rates. This facilitates a deeper understanding of the role each feature plays in water source identification. An aquifer mine water inrush source identification model based on IPFA-ELM-MIV is constructed, and water source identification is accomplished through this model. Experimental results indicate that the proposed method can elevate the high-dimensional mutual information entropy of sample features to above 0.9 through feature selection. In nine identification processes, the probability of misjudgment using the proposed method is 0. The chemical characteristic types of water inrush in aquifer mines are accurately identified, improving the accuracy of identifying water inrush sources from different aquifers in mines. Moreover, the R₁ value for identifying water inrush sources in aquifer mines under different sampling environments exceeds 0.95, demonstrating stronger recognition adaptability. This method is of great significance for preventing and managing mine water inrush disasters.