谱聚类结合LIF在矿井突水水源类型识别中的应用

突水事故威胁井下人员的生命安全和造成财产损失，因此准确检测出突水水源类型具有重大意义。使用水化学分析法检测水源类型耗时长、过程复杂。激光诱导荧光（LIF）技术具有快速、灵敏、干扰小等优点，将LIF技术结合智能算法建立突水水源识别模型可以准确检测出突水水源的类型。目前这类模型一般需要对荧光光谱进行去噪、降维、波段选取等处理，过程繁琐，并且模型都是在均匀分组的突水水源荧光光谱上建立的，并没有讨论不均匀分组对模型的影响，也没有针对不均匀分组建立模型。在实际工程应用中，采集的样本数量是有很大概率呈现不均匀的，因此本文提出一种飞蛾扑火（MFO）算法结合谱聚类（SC）的方法实现对不均匀分组的突水水源荧光光谱的识别。实验中，首先从淮南煤矿获取5种实验水样，使用激光诱导荧光实验设备采集所有水样的荧光光谱，五种水样的组数分别为75，80，80，30和135。其次，建立MFO-SC水样识别模型，通过对比后标签映射方式选择K-Means、相似矩阵的计算方式选择高斯核函数和划分准则选择ncut，用MFO对高斯核函数的参数寻优得到σ的值为1.745并且固定模型的初始聚类中心。随后，分别建立K-Means，SVM和MFO-SVM3种水样识别模型。对比MFO-SC模型与K-Means模型，得到MFO-SC模型的最优准确率为100%且平均准确率也为100%，K-Means模型的最优准确率为99.75%，而平均准确率为79.57%；再分别计算SVM模型和MFO-SVM模型的训练集准确率和测试集准确率，SVM模型训练集准确率为80%，测试集准确率为80%；MFO-SVM模型训练集准确率为100%，测试集准确率为95.625%。最后，使用4种模型对其他三个不均匀分组的突水水源荧光光谱进行识别，研究结果表明将MFO-SC算法用于突水水源类型的识别上是有效的，可以准确地检测出突水水源的类型，对煤矿生产安全有重要意义。

Application of Spectral Clustering and LIF in Recognition of Mine Water Inrush Source Types

Water inrush accidents threaten the lives of people and cause property damage.Therefore,it has great significance in accurately detecting the type of water inrush.Hydrochemical analysis method takes a long time and has a complicated process to detect the type of water inrush.Laser-induced fluorescence(LIF)technique has the advantages of fastness,high sensitivity,and low interference.Water inrush source recognition model building with LIF technique and intelligent algorithms can accurately detect the type of water inrush.At present,such models generally require de-noising,dimension reduction,and band selection on the fluorescence spectra,and this process is complicated.The models are built on the fluorescence spectra of the water inrush source which is evenly grouped.The influence of the uneven grouping on the model is not discussed,and the model is not built for the uneven grouping.In practical engineering applications,the number of samples collected is highly likely to be uneven,so Moth-flame optimization(MFO)algorithm combined with spectral clustering(SC)is proposed to realize the uneven grouping of water inrush fluorescence spectrain this paper.In the experiment,firstly,five kinds of experimental water samples were obtained from Huainan coal mine.Laser-induced fluorescence experimental equipment was used to collect fluorescence spectra of all water samples.The number of groups of five water samples is 75,80,80,30 and 135.Secondly,build MFO-SC water sample recognition model.After comparison,K-Means is selected for the label mapping method,the Gaussian kernel function is selected for the calculation method of the similarity matrix,and the ncut is selected for the partition criterion.The parameters of the Gaussian kernel function were optimized by using MFO to obtain the parameter value of 1.745,and the initial clustering center of the model was fixed.Subsequently,build three water sample recognition models of K-Means,SVM and MFO-SVM,respectively.Comparing the MFO-SC model with the K-Means model,the optimal accuracy of the MFO-SC model is 100%,and the average accuracy is 100%.The optimal accuracy of the K-Means model is 99.75%,and the average accuracy is 79.57%.Then calculate the training set accuracy and test set accuracy of the SVM model and MFO-SVM model respectively.The accuracy of the training set of the SVM model is 80%,and the accuracy of the test set is 80%;the accuracy of the training set of the MFO-SVM model is 100%,and the accuracy of the test set is 95.625%.Finally,four models were used to identify water inrush fluorescence spectra of the other three uneven groups.The research results show that the MFO-SC algorithm is effective in identifying the type of water inrush,and can accurately detect the type of water inrush,which has great significance on the safety of coal mine production.