基于QDA-RF-MLP的矿井突水水源判识模型

Water Inrush Source Identification Model Based on QDA-RF-MLP

摘要: 受采动破坏与断裂构造影响，深部煤矿突水水源的水化学特征难以解释，具有混合性和重叠性，导致传统单一模型识别精度受限。本文以安徽祁南煤矿为研究背景，提出一种基于二次判别分析(QDA)、随机森林(RF)与多层感知机(MLP)的堆叠集成水源判识模型。在系统解析矿井主要含水层水化学演化规律的基础上，利用SMOTE技术重构非平衡数据集，并引入水化学离子比率等高阶特征。通过五折交叉验证优选基学习器参数，构建双层异构融合模型。实验结果表明：该集成模型在独立测试集上的识别准确率达95.83%，F1分数达0.96，显著优于单一最优基模型RF (92.7%)，有效克服了小样本条件下的过拟合问题，对复杂混合水源具有较强的鲁棒性。

Abstract: Under the influence of mining disturbances and geological fault structures, the hydrochemical signatures of water inrush sources in deep coal mines often exhibit complex nonlinear mixing and overlapping characteristics. Consequently, traditional single-model approaches frequently suffer from limited identification accuracy. Taking the Qinan Coal Mine in Anhui Province as a case study, this paper proposes a stacking ensemble learning model for water source identification, integrating Quadratic Discriminant Analysis (QDA), Random Forest (RF), and Multilayer Perceptron (MLP). First, the hydrochemical evolution mechanisms of the primary aquifers were systematically analyzed. Subsequently, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to address dataset imbalance, and high-order features—specifically hydrochemical ion ratios, were introduced to enhance discriminative capability. A two-layer heterogeneous fusion model was then constructed, with base learner parameters optimized via five-fold cross-validation. Experimental results demonstrate that the proposed ensemble model achieved an accuracy of 95.83% and an F1-score of 0.96 on an independent test set. This performance significantly outperforms the best single base model (RF, 92.7%). The study confirms that the proposed method effectively mitigates overfitting risks associated with small sample sizes and exhibits strong robustness in identifying complex, mixed water sources.