多物种血液鉴别对于进出口检验检疫、 刑事侦检以及野生动物保护等领域尤为重要。 传统的血液鉴别方法, 在鉴别时常常会对血液样本造成破坏, 而Raman光谱作为一种振动光谱可获得物质分子振动、 转动信息, 进而分析物质组成, 为无损血液鉴别技术提供了可能。 目前, 已经有基于Raman光谱进行血液鉴别的报道, 但存在如下两个问题: 单一物种样本数量较少, 易导致模型欠拟合; 均采用线性分类模型, 忽略了光谱中非线性因素的影响, 降低了模型的分类性能。 因此, 将支持向量机沿用至Raman光谱血液鉴别中, 克服了线性模型只能为光谱中线性关系建模的缺点, 有效地吸收了Raman光谱中的非线性关系, 实现了对人、 犬及兔血液的三分类。 实验通过激发波长为785 nm的海洋Raman光谱仪测得共326例样本数据(人110例、 犬116例、 兔100例), 利用Savitzky-Golay平滑滤波、 加权最小二乘多项式拟合基线以及矢量归一化等方法对Raman光谱数据进行预处理, 并选择2/3的样本数据作为校正集用于模型训练, 余下1/3作为测试集用于盲测。 与线性分类模型对比实验结果显示, 该模型的校正集分类正确率达100%, 盲测集分类正确率达9352%, 均优于线性分类模型。 实验结果表明, 基于支持向量机的分类模型可以用于Raman血液光谱鉴别, 具有重要的研究价值和广泛的应用前景。

The identification of multiple species blood is particularly important for entry-exit inspection and quarantine, forensic investigation and wildlife protection. The traditional methods often destroy blood samples and make further analysis of samples impossible. Raman Spectroscopy is a vibrational spectrum, which can obtain the information of molecular vibration and rotation so as to analyze the chemical composition of the material. It provides the possibility of non-destructive blood identification. Currently, there are several methods of blood identification based on Raman spectroscopy, but these methods use the linear classification model, ignoring nonlinear relationship between the spectrum and sample, and lead to the bad performance of the model. Moreover, the small sample number of each species usually results in the under-fitting the model. Therefore, this paper set up a classification model for the nonlinear relationship using the support vector machine to identify Raman spectra of blood, overcame the shortcoming of the linear classification model which emphasizes the linear characteristic of the spectrum in the training, and absorbed the nonlinear relationship in the Raman spectrum effectively, realizing the three classification of human, dog and rabbit blood. There are a total of 326 samples which were measured by Ocean Raman spectrometer with excitation wavelength of 785 nm, including 110 humans, 116 dogs and 100 rabbits. Savitzky-Golay smoothing filter, weighted least squares baseline correction, and vector normalization were used to preprocess them. The 2/3 of these samples were used as calibration set for training and the remaining samples were used as test set for blind testing. Experimental results showed that the classification accuracy of proposed model for the calibration set and the blind test were 100% and 9352%, and outperformed the existing linear classification models. This indicates that proposed classification model has good application prospects and research value.