基于Piper-PCA-MLP神经网络的矿井涌水水源识别方法研究

Study on Mine Water Inrush Source Discrimination Method Based on Piper-PCA-MLP Neural Network

为了快速准确地识别矿井涌水水源,提出一种基于Piper-PCA-MLP的矿井涌水水源判别模型。以乌东煤矿为例,通过Piper三线图揭示了矿区不同地下水含水层的水化学特征,筛选了来自矿区不同涌水含水层的23个水样,获得了22个代表涌水含水层的典型水样。选用Na^(+)+K^(+)、Ca^(2+)、Mg^(2+)、Cl^(-)、SO_(4)^(2-)、HCO_(3)^(-)、NO_(3)^(-)和TDS八种水化学指标作为判别指标,考虑到乌东煤矿含水层水质边界不明显和水化学组分差异性不大的特点,选择MLP神经网络作为模型识别的基础,为了解决MLP神经网络在面对高维数据时训练收敛较慢,建模时间较长等缺点,使用主成分分析算法对数据进行降维,提取三种主要指标作为判别因子。以22组水样数据中17组作为训练样本,5组作为测试样本建立PCA-MLP模型,为了凸显模型的可靠性和准确性,选取同一样本集对水样数据分别建立PCA-层次聚类模型、MLP、PCA-RBF模型,并对其判别结果进行分析比较。结果表明:经过PCA降维的MLP神经网络能够有效消除水样水化学指标间的相互影响,对第四系孔隙潜水,地表水判别正确率均为100%,对基岩裂隙孔隙水的判别正确率为92.3%,明显优于其它方法,可为矿井涌水水源的识别提供一种新方法。

To fast and accurately discriminate mine water inrush source,a kind of mine inrush water source discrimination model based on Piper-PCA-MLP has been proposed.Taking the Wudong coalmine as an example,through Piper trilinear diagram has revealed mine area different groundwater aquifer hydrochemical features.23 water samples from mine area different water inrush aquifers have been selected and acquired 22 typical water samples representing water inrush aquifers.Using Na^(+)+K^(+)、Ca^(2+)、Mg^(2+)、Cl^(-)、SO_(4)^(2-)、HCO_(3)^(-)、NO_(3)^(-) and TDS 8 hydrochemical indices as discrimination indicators.Considering aquifers in Wudong coalmine have characteristics of obscure water quality confines and hydrochemical components otherness not large have selected MLP neural network as the basis of model recognition.To solve MLP neural network slower training convergence when facing high dimensional data,modeling time longer etc shortcomings,the principal component analysis algorithm has been used to carry out data dimension reduction,extract 3 main indicators as discrimination factors.To take 17 groups from 22 groups water sample data as training samples,5 groups as testing samples have established PCA-MLP model.To highlight model reliability and accuracy can select same sample set to establish water sample data PCA-hierarchical clustering,MLP,PCA-RBF models respectively,and analytically compared their discrimination results.The result has shown that MLP neural network after PCA dimension reduction can eliminate interaction between water sample hydrochemical indices effectively.To Quaternary pore phreatic water and surface water,their discrimination accuracies are all 100%,while to bedrock fissure-pore water is 92.3%,thus obviously better than other methods.The study has provided a new method for mine water inrush source discrimination.