化学学报 ›› 2012, Vol. 70 ›› Issue (18): 1969-1973.DOI: 10.6023/A12050265 上一篇    下一篇

研究论文

训练字典及其稀疏表示在近红外光谱法检测柴油中的应用

周扬a,b, 戴曙光a, 吕进b, 刘铁兵b, 施秧b   

  1. a 上海理工大学 光电信息与计算机工程学院 上海 200093;
    b 浙江科技学院 近红外应用技术研究室 杭州 310023
  • 投稿日期:2012-05-31 发布日期:2012-08-06
  • 通讯作者: 周扬
  • 基金资助:

    项目受国家自然科学基金(No. 51075280)、浙江省重大科技专项和优先主题计划(No. 2010C11060)和浙江省自然科学基金(No. Y4110235)资助.

Application of Training Dictionary and Sparse Representation in Diesel Property Detection Using Near-infrared Spectroscopy

Zhou Yanga,b, Dai Shuguanga, Lv Jinb, Liu Tiebingb, Shi Yangb   

  1. a College of Optics & Electronics Engineering, University of Shanghai for Science and Technology, Shanghai 200093;
    b Laboratory of NIR-Applications, Zhejiang University of Science and Technology, Hangzhou 310023
  • Received:2012-05-31 Published:2012-08-06
  • Supported by:

    Project supported by the National Natural Science Foundation of China (No. 51075280), the Major Science and Technology Projects of Zhejiang Province (No. 2010C11060), the Natural Science Foundation of Zhejiang Province (No. Y4110235).

为提高柴油组分近红外法检测的精度, 提出了一种基于训练字典稀疏表示下的建模方法并用于柴油十六烷值、沸点和芳烃总量的检测. 该法先用柴油光谱结合K 均值奇异值分解(K-SVD)算法完成对冗余字典的训练, 再用正交匹配追踪算法(OMP)寻找柴油光谱在该训练字典下的稀疏表示系数, 用该系数建立了柴油十六烷值、沸点和芳烃总量偏最小二乘预测模型. 实验比对了训练字典、傅里叶字典和小波字典稀疏表示下的柴油组分预测模型性能, 其中训练字典的表示系数建模性能最优且比其他两种字典的预测性能有较大幅度改进, 验证了该法在近红外光谱检测建模领域推广能改善预测的准确度和稳健性.

关键词: 近红外, 柴油组分, 稀疏表示, 冗余字典, 字典训练

It is an important method for predicting the diesel quality parameter by combination of near-infrared spectroscopy and chemometrics. In order to improve the accuracy of prediction model for diesel property, this paper puts forth a modeling method based on training dictionary and its sparse representation, and this method can test the cetane number, boiling point and total aromatics fast. The first stage of the method searches the particular redundant dictionary for diesel oil calibration samples, and the redundant dictionary can be trained by using K-means singular value decomposition algorithm (K-SVD). After finding of this redundant dictionary, it uses the orthogonal matching pursuit algorithm (OMP) to search the sparse representation coefficients of diesel spectroscopy under the training dictionary. This coefficients of the training dictionary, which is similar with the sparse representation coefficients of wavelet dictionary and fourier dictionary, is scale expansion of spectral characteristics that can effectively use to establish the pre-test partial least squares prediction (PLS) model of cetane number, boiling point and total aromatics for diesel. The spectra of diesel samples with different quality parameters from same factory are collected and the quality parameters of diesel samples are measured by ASTM certification method. The sparse representation coefficients of diesel spectroscopy under fourier dictionary and wavelet dictionary are figured in the experiments, inter-compared the performance of prediction model for diesel property under sparse representation coefficients of training dictionary, fourier dictionary and wavelet dictionary. Furthermore, the PLS model correlation coefficient (R) and standard deviation of prediction set (SEP) under sparse representation coefficients in training dictionary achieve 0.99 and 0.05 separately, it has optimal performance and also has considerable improvement to compare with the prediction performance of fourier dictionary and wavelet dictionary, which validated the efficiency of the method in predicting the diesel quality parameter using near-infrared spectroscopy and the proposed method can substantially improve the accuracy and stability of near-infrared spectroscopy prediction method.

Key words: near-infrared spectroscopy, diesel property, sparse representation, redundant dictionary, dictionary training