基于SiPLS特征提取和信息融合的汽油中乙醇含量的多光谱检测

发布时间：2018-06-23 19:45

本文选题：紫外可见光谱 + 近红外光谱　；参考：《光谱学与光谱分析》2017年02期

【摘要】：采用紫外可见(ultraviolet/visible,UV-Vis)光谱技术和近红外(near-infrared,NIR)光谱技术及信息融合技术对乙醇汽油中乙醇含量进行了检测。首先采用组合区间偏最小二乘(synergy interval PLS,SiPLS)算法作为特征提取方法,分别建立了基于UV-Vis和NIR光谱的偏最小二乘(PLS)回归模型;再根据油品的实际情况,运用信息融合理论将UV-Vis和NIR光谱信息进行融合,建立了数据级融合(low level data fusion,LLDF)和特征级融合(mid-level data fusion,MLDF)模型,并与单谱源模型效果进行了比较,确定了最优模型为数据级融合后再进行矢量归一化的模型(LLDF-VN1);最后分别用高乙醇含量样品和市售汽油样品的光谱数据对该最优模型进行了通用性检验。结果表明:UV-Vis和NIR光谱数据单独建模均能很好的检测并提供较好的预测结果;而UV-Vis和NIR光谱数据直接融合在基于校正集的回归模型中效果最好,其校正集相关系数rc=0.999 9,校正集交叉验证均方差RMSECV=0.125 8,校正集整体评价偏差Bias_c=0.000 6;而采用数据级融合后再进行矢量归一化的模型(LLDF-VN1)的预测效果为最佳,其r_p=0.999 1,RMSEP=0.352 7,Bias_p=-0.073 8;自配溶液对最优模型(LLDF-VN1)的通用性验证中,r_p=0.999 7,RMSEP=0.329 1,Bias_p=0.102 2;市售汽油对最优模型(LLDF-VN1)的通用性验证中,r_p=0.990 1,RMSEP=0.892 7,Bias_p=0.675 1。实验结果说明通过将UV-Vis和NIR光谱信息进行数据级融合可以快速、准确的检测出乙醇汽油中乙醇的含量,并能实现乙醇浓度的宽范围检测,为进一步实现混合油品中物质的快速检测奠定了基础。
[Abstract]:The content of ethanol in ethanol gasoline was determined by ultraviolet / visible (UV-Vis), near-infrared (NIR) and information fusion techniques. The partial least squares (PLS) regression models based on UV-Vis and NIR spectra are established by using the combinatorial interval partial least squares (synergy interval) (synergy interval SiPLS algorithm as the feature extraction method, and then according to the actual conditions of the oil, the partial least squares (PLS) regression models based on UV-Vis and NIR spectra are established respectively. UV-Vis and NIR spectral information were fused by using information fusion theory, and the data level fusion (low level data fusion model and the characteristic level fusion model were established, and the results were compared with those of the single spectral source model. The optimal model was established as a vector normalization model (LLDF-VN1) after data level fusion, and the general use of the optimal model was tested with the spectral data of high ethanol content sample and gasoline sample respectively. The results show that the single modeling of the spectral data of: UV-Vis and NIR can well detect and provide better prediction results, while the direct fusion of UV-Vis and NIR spectrum data is the best in the regression model based on correction set. The correlation coefficient of correction set is rc=0.999 9, the mean square deviation of correction set cross validation is 0.1258, and the deviation of whole evaluation of correction set is Biasspar 0.000 6. The model of vector normalization after data level fusion (LLDF-VN1) is the best one. Its r_p=0.999 1 / RMS EPN 0.352 7 / Biasp _ (+) -0.073 8; RMP ~ (0.999 7) RMS EP0.329 ~ (1) Biasp ~ (0.102); rp0. 990 ~ (1) RMSEPN ~ (0.892 7) Biasp ~ (0.675); rp0. 990 ~ (1) RMSEPN ~ (0.892 7) Biasp _ (0.675)] in the verification of the generality of the optimal model (LLDF-VN1) of gasoline sold in the market. The experimental results show that the data level fusion of UV-Vis and NIR spectrum information can quickly and accurately detect the ethanol content in ethanol gasoline, and can achieve a wide range of detection of ethanol concentration. It lays a foundation for the rapid detection of substances in mixed oil products.
【作者单位】：燕山大学信息科学与工程学院河北省特种光纤与光纤传感重点实验室;内蒙古民族大学物理与电子信息学院;
【基金】：国家自然科学基金项目(61475133) 河北省科技计划项目(15273304D,14273301D,16273301D,16213902D)资助
【分类号】：O657.33;TE626.21

【相似文献】