紫米是生活中常见的食材, 具有丰富的营养价值。 由于紫米价格较高导致染色紫米大量流入市场。 本文使用太赫兹时域光谱技术结合化学计量学方法探索紫米掺假的快速检测方法。 采用太赫兹时域光谱技术(THz-TDS)采集0～7 THz范围内紫米掺假的光谱数据, 并选择0.5～2.5 THz波段的吸收系数谱和折射率谱进行分析并采用化学计量学方法对光谱数据进行建模分析。 分别采用Savitzky-Golay卷积平滑（SG Smoothing, SG平滑）、 基线校正（Baseline）、 归一化（Normalization）、 多元散射校正（MSC）等方法进行光谱预处理, 结合偏最小二乘判别分析(PLS-DA)对紫米、 紫米掺染色大米和紫米掺染色黑米进行定性分析。 定性分析结果显示, 通过主成分分析（PCA）的三种样品平面分布存在明显差异; 经过基线校正的光谱数据建立的PLS-DA模型效果最佳, 误判率为0。 接着使用偏最小二乘法(PLS)结合SG平滑、 Baseline、 Normalization、 MSC等预处理方法分别对紫米中掺染色大米和紫米中掺染色黑米的光谱数据建立PLS定量模型。 结果显示, 采用基线校正预处理方法的PLS建模效果最佳, 紫米掺染色大米的预测集相关系数为0.936, 预测集均方根误差（RMSEP）为0.095。 紫米掺染色黑米的预测集相关系数为0.914, 预测集均方根误差为0.096。 为对比分析线性（PLS）与非线性（LS-SVM）两种定量模型方法的预测精度, 采用相同预处理方法后的紫米掺假含量光谱数据建立最小二乘支持向量机（LS-SVM）预测模型, 选用径向基函数（RBF）作为核函数。 结果表明采用基线校正处理后LS-SVM模型效果最佳, 紫米中掺染色大米的预测集均方根误差（RMSEP）为0.092, 预测集相关系数（Rp）为0.979; 紫米中掺染色黑米的预测集均方根误差（RMSEP）为0.093, 预测集相关系数（Rp）为0.948。 对比发现对紫米掺假的含量建立LS-SVM预测模型较PLS模型的稳定性更好、 精确度更高。 研究表明, 太赫兹时域光谱结合化学计量学方法可为紫米掺假的定性定量分析提供快速精确的分析方法。

Purple rice is a common ingredient in life and has rich nutritional value. Due to the high price of purple rice, the dyed purple rice has entered the market in large quantities. In this paper, terahertz time-domain spectroscopy combined with chemometric methods is used to explore the rapid detection method of purple rice adulteration. The spectral data of purple rice adulteration in the range of 0～7 THz was collected by Terahertz Time domain Spectroscopy (THz-TDS), and the absorption coefficient spectrum and refractive index spectrum of 0.5～2.5 THz band were selected for analysis and adopted. The chemometric method models and analyzes the spectral data. Savitzky-Golay convolution smoothing (SG smoothing), baseline correction (Baseline), normalization (Normalization), multiple scattering correction (MSC) and other methods are used for spectral preprocessing. Qualitative analysis of purple rice, purple rice mixed with rice and purple rice mixed with black rice was carried out by partial least squares decision analysis (PLS-DA). Qualitative analysis showed that there were significant differences in the plane distribution of the three samples by Principal Component Analysis (PCA); the PLS-DA model established by baseline corrected spectral data had the best effect, and the false positive rate was 0. Then using partial least squares (PLS) combined with SG smoothing, Baseline, Normalization, MSC and other pretreatment methods to establish a PLS quantitative model for the spectral data of the black rice mixed with dyed rice and purple rice. The results showed that the PLS model with baseline correction pretreatment method had the best effect. The correlation coefficient of the prediction set of purple rice-doped rice was 0.936, and the root means square error of prediction (RMSEP) was 0.095. The correlation coefficient of the prediction set of purple rice blended black rice was 0.914, and the root mean square error of the prediction set was 0.096. In order to compare and analyze the prediction accuracy of linear (PLS) and nonlinear (LS-SVM) quantitative model methods, the least squares support vector machine (least squares support vector) is established by using the same pretreatment method. Machine, LS-SVM) predictive model, using radial basis function (RBF) as the kernel function. The results showed that the LS-SVM model had the best effect after baseline correction. RMSEP of the predicted rice with purple rice was 0.092, and the correlation coefficient (Rp) of the prediction set was 0.979. RMSEP of the meter is 0.093, and the prediction set correlation coefficient (Rp) is 0.948. The comparison found that the LS-SVM prediction model for the content of purple rice adulteration is better and more accurate than the PLS model. Studies have shown that terahertz time-domain spectroscopy combined with chemometric methods can provide a fast and accurate analytical method for qualitative and quantitative analysis of purple rice adulteration.