Ye Shu-bin, Xu Liang, Li Ya-kai, Liu Jian-guo, Liu Wen-qing
Guang Pu Xue Yu Guang Pu Fen Xi. 2017 Mar;37(3):749-54.
With the developing of catering trade, cooking oil fume has became one of the three major air pollution sources in some cities. In recent years, a lot of research on the cooking oil fume have been done for its high threaten to human health. The cooking oil fume contains a large amount of unsaturated hydrocarbons produced by pyrolysis of edible oil, which are harmful to human health. The characteristics of the composition and content of edible oil fumes produced by pyrolysis of different edible oil are different. For classification and identification of edible oil, two kinds of classification and identification mathematical model are constructed. The spectrum data of different edible oil fume are collected by Fourier transform infrared spectrometer which is independent research and development. At the same time, different classification algorithms of the principal component analysis (PCA) combining probabilistic neural network (PNN) and the error back propagation artificial neural network (BPANN) are constructed respectively. Two kinds of classification algorithms are used to analyze the Fourier transform infrared spectrum data of different cooking fume gas. The mathematical models are trained by the sample data, and the trained mathematical model are used to analyze the unknown spectral data to determine the type of edible oil. The experimental results show that the two algorithms can classify and identify different types of oil fume. In the whole band recognition, the recognition rate is 90.25% and 97% respectively. By analyzed spectral data of flue gas absorption band, spectrums of atmospheric window and the strong absorption feature bands of volatile organic compounds (VOCs) (from 1 300 to 700 cm-1 and from 3 000 to 2 600 cm-1) were extracted. The absorbance data are divided into two parts with separated absorption band, and the two algorithms in 3 000～2 600 cm-1 band have better recognition rate. PCA-PNN algorithm recognition rate is 90.25% and PCA-BPANN algorithm recognition rate is 92.25%. Obviously, two kinds of artificial neural network algorithm combining principle component analysis respectively can effectively identify the types of edible oil fume.
随着餐饮行业的发展，食用油油烟已成为一些城市的三大主要空气污染来源之一。近年来，针对食用油油烟对人体健康的高威胁性，已开展了大量相关研究。食用油油烟中含有大量由食用油热解产生的不饱和烃，这些物质对人体健康有害。不同食用油热解产生的油烟在成分和含量特征上存在差异。为了对食用油进行分类和识别，构建了两种分类识别数学模型。利用自主研发的傅里叶变换红外光谱仪采集不同食用油油烟的光谱数据。同时，分别构建了主成分分析（PCA）结合概率神经网络（PNN）和误差反向传播人工神经网络（BPANN）的不同分类算法。运用这两种分类算法对不同烹饪油烟气的傅里叶变换红外光谱数据进行分析。通过样本数据对数学模型进行训练，再利用训练好的数学模型分析未知光谱数据以确定食用油的类型。实验结果表明，这两种算法能够对不同类型的油烟进行分类和识别。在全波段识别中，识别率分别为90.25%和97%。通过分析烟气吸收带的光谱数据，提取了大气窗口光谱以及挥发性有机化合物（VOCs）的强吸收特征带（1300至700 cm-1和3000至2600 cm-1）。将吸光度数据按分离的吸收带分为两部分，两种算法在3000～2600 cm-1波段具有更好的识别率。PCA - PNN算法识别率为90.25%，PCA - BPANN算法识别率为92.25%。显然，两种分别结合主成分分析的人工神经网络算法能够有效识别食用油油烟的类型。
