基于LVQ与SVM算法的近红外光谱煤产地鉴别

摘要

Traditional coal origin identification method generally take the calorific value,volatiles,caking index,hardgrove index and crucible swelling number as the classification index,process complicated,use manpower and material resources and can’t get coal sample origin directly,take advantages of the near-infrared spectrum technology fast nondestructive testing,due to be col-lected in the original spectrum that contains some or false spectral data,using Leave-one-out cross validation based on SVM to e-liminate abnormal sample of spectral data set,get the correct spectral information of coal sample spectra data sets,and construct the qualitative analysis model based on SVM algorithm and LVQ algorithm,complete based on near-infrared spectral analysis technology of coal origin identification,don’t need to make summary and coal samples of various indicators forecast.In view of the random parameter optimization problems in SVM model,the PSO-SVM model of loss parameters (C)and the radius of ker- nel function (g)are improved,get the optimal parameters,finally,calculation accuracy of the method above contrast model is introduced to evaluate and analysis.Experiments collect the near infrared spectrum of Canada,Russia,Australia,Indonesia and China’s five regions,all the data sets,a total of 305 samples,of which 10 simples is abnormal samples and the first 31 groups of the coal spectra were selected as training samples,6 sets of data after as test samples.Results show that the classification accu-racy of classification model can achieve 75% above,including the analysis of the SVM model based on PSO algorithm to improve the accuracy can reach 96.67%,only a sample appear problem,it will be realized quickly and efficiently based on near-infrared spectral analysis technology of coal origin identification.%传统煤产地鉴别方法一般以发热量、挥发分、粘结指数、哈氏可磨指数和坩埚膨胀序数作为分类指标，过程复杂耗时较多、耗费巨大的人力、物力并且无法直接快速的得到煤样产地等问题，借助近红外光谱技术快速无损检测的优势，利用基于 SVM的留一算法对光谱数据集进行异常样本剔除，得到包含正确光谱信息的煤样光谱数据集，构造基于 SVM算法与 LVQ算法的定性分析模型，完成基于近红外光谱分析技术的煤产地的快速鉴别，无需对煤样的各种指标进行汇总并且人为预测。针对 SVM分析模型中存在随机参数优化问题，引入PSO算法对SVM模型中的损失参数C和核函数半径g 进行改进，得到最优参数，最后引入计算准确率的方法对比以上模型并进行评价分析。实验一共收集了加拿大、俄罗斯、澳大利亚、印度尼西亚、中国内蒙等5个地区的煤样光谱数据集，数据集共计305组煤样样本，其中异常样本共计10组，分别选择各国煤炭光谱的前31组作为训练样本，后6组数据作为测试样本，结果表明各分类模型的分类准确率均能达到75%以上，其中基于 PSO 算法改进的 SVM分析模型的准确率可达到96.67%，仅一个样本出现问题，可快速高效地实现基于近红外光谱分析技术的煤产地的鉴别。