基于激光诱导击穿光谱法和反向传播神经网络建立废钢分类和定标模型及其应用  

Establishment and Application of Classification Model and Calibration Model of Steel Scrap Based on Laser Induced Breakdown Spectroscopy and Back Propagation Neural Network

作　　者：刘艳丽 刘宏强 黄晓红 孙永长 卫广运 LIU Yanli;LIU Hongqiang;HUANG Xiaohong;SUN Yongchang;WEI Guangyun(He Steel Group Central Iron and Steel Research Institute,Shijiazhuang050000,China;School of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]河钢集团钢研总院,石家庄050000 [2]华北理工大学人工智能学院,唐山063210

出　　处：《理化检验（化学分册）》2022年第12期1389-1394,共6页Physical Testing and Chemical Analysis(Part B:Chemical Analysis)

基　　金：河北省自然科学基金(E2016318007)。

摘　　要：基于自建激光诱导击穿光谱(LIBS)设备、软件以及反向传播(BP)神经网络,利用18个钢标准样品的LIBS光谱数据分别和钢标准样品类别以及Si、Mn、Cr、Ni、Cu元素的认定值建立分类模型和定标模型,并用于实际样品的检测。以各元素优选谱线(Si、Mn、Cr、Ni、Cu元素谱线分别为251.60,293.86,286.41,227.01,213.60 nm)与对应铁元素谱线(对应Fe元素谱线分别为263.54,292.66,271.44,263.54,206.98 nm)下相对强度作为输入变量,建立3层BP神经网络模型。分类模型的最大迭代次数为500,学习率为0.01,360组数据带中训练集和测试集的数量比为3∶1;Si、Mn、Cr、Ni、Cu定标模型的最佳迭代次数分别为200,200,200,160,280次,360组数据中训练集和测试集的数量比为4∶1,模型性能通过线性相关性拟合度(R^(2))、均方根误差(RMSE)、平均百分比误差(MPE)和残差平方和(PRESS)等4个指标评价。结果显示:分类模型对测试集分类的预测准确率达到100%;测试集中5种元素的定标模型R2分别为0.941,0.983,0.983,0.988,0.987,RMSE分别为0.0612,0.0607,0.0425,0.0496,0.0169。定标模型对实际样品的预测值和参考GB/T 4336-2016所得测定值基本一致。所建方法可用于钢铁行业中废钢样品的分类及其中成分的快速检测。Based on self-built laser induced breakdown spectroscopy(LIBS)equipment,software and back propagation(BP)neural network,the classification model and calibration model were established using the LIBS spectral data together with category and certified values of Si,Mn,Cr,Ni,Cu of 18 steel standard samples,which were used for the detection of actual samples.A three-layer BP neural network model was established with the relative intensity under the optimized spectral line of each element(the spectral lines of Si,Mn,Cr,Ni,Cu were 251.60,293.86,286.41,227.01,213.60 nm)and the corresponding spectral line of iron element(the corresponding spectral lines of Fe were 263.54,292.66,271.44,263.54,206.98 nm)as input variables.For classification model,the maximum number of iterations was 500,the learning rate was 0.01,and the number ratio of the training set to the test set in the 360 groups of data was 3∶1.For calibration model,the optimal number of iterations for Si,Mn,Cr,Ni,and Cu were 200,200,200,160,280,the number ratio of the training set to the test set in 360 groups of data was 4∶1,and the performance of the calibration model was evaluated by the 4 indices,including linear correlation fit(R^(2)),root mean square error(RMSE),mean percentage error(MPE)and sum of squares of residuals(PRESS).As found by the results,the predicted accuracy of the test set classification with the classification model was 100%.R~2 of the calibration models of the 5 elements in the test set were 0.941,0.983,0.983,0.988,0.987,and RMSE were 0.0612,0.0607,0.0425,0.0496,0.0169,respectively.The predicted values of actual samples with calibration models were consistent with those given by GB/T 4336-2016.The established method could be used for the classification of steel scrap samples and the rapid detection of their components in the steel industry.

关 键 词：激光诱导击穿光谱法 反向传播神经网络 废钢 分类 定量分析 

分 类 号：O657.319[理学—分析化学]