Intelligent identification model of mine water inrush sources under the condition of unbalanced samples in complex hydrogeological settings

Rapidly and accurately determining the sources of mine water inrush is vital for hazard control. Nevertheless, the common problem of class imbalance in water databases and the practical challenges in quickly gathering supplementary groundwater samples pose a major obstacle to acquiring the balanced training data needed for effective model development. This study proposes a Fisher water inrush sources identification algorithm optimized via hydrochemical indices derivation and dynamic feature selection to overcome the limitations of traditional models. Additionally, the Borderline Synthetic Minority Oversampling Technique (Borderline-SMOTE) is integrated into the framework to address class imbalance, effectively overcoming traditional limitations such as restricted discriminant indices, reliance on fixed ion combinations, and non-adaptive features. This optimization expands discriminant indices dimensionality and enables automatic screening of key indices based on mining area hydrogeology, enhancing model flexibility and adaptability. Applied to Malan Coal Mine, 8 basic hydrochemical indicators from 91 water samples were expanded to 20, with Borderline-SMOTE augmenting the training set and dynamic selection identifying 11 optimal indices. The performance evaluation of the optimized model in comparison with other models demonstrated that the improved model achieved an average accuracy of 93.08% in 5-fold cross-validation on the training set, a resubstitution accuracy of 93.97%, and a discrimination accuracy as high as 86.96% for 23 groups of water samples in the test set, which is significantly higher than those of conventional Fisher models. Under three data splitting ratios (3:1, 7:3, and 8:2), the model maintained stable discriminatory performance, indicating that the K‑BS‑D‑Fisher model—optimized through the derivation and expansion of hydrochemical indices, Borderline‑SMOTE, and dynamic feature selection algorithms—retains strong generalization capability even under complex hydrogeological conditions. The findings offer a new perspective for the accurate identification of mine water inrush sources.