针对目前太赫兹光谱技术在食用油品质检测方面存在定性多、 定量难的问题, 提出一种基于衰减全反射(ATR)式太赫兹时域光谱(THz-TDS)技术的食用油过氧化值定量分析方法。 首先采集不同种类、 不同氧化程度食用油样本的太赫兹时域光谱图, 筛选有效信号波段并提取得到光学常数, 经预处理算法校正后的光学常数, 结合多种化学计量学方法建立定量分析模型, 实现快速、 无损预测食用油的过氧化值。 70个实验样本包括大豆油、 菜籽油和玉米, 过氧化值覆盖范围0.41~10.23 mmol·kg^-1, 且样本的过氧化值均匀分布。 采用TeraView公司生产配有ATR检测模块的TeraPulse 4000太赫兹脉冲光谱系统采集样本THz-TDS信号, 根据THz-TDS谱图信号特征筛选有效波段10~86.78 cm^-1用于建模分析。 通过快速傅里叶变换得到频域信号并从中提取光学常数: 折射率和吸收系数, 采用Savitzky-Golay7点卷积平滑分别对折射率和吸收系数进行预处理, 去除干扰信号。 运用SPXY算法按照3:1比例划分校正集和预测集样本, 结合主成分回归法、 偏最小二乘法两种化学计量学分析方法, 分别建立基于折射率和基于吸收系数的过氧化值分析模型。 对模型评价指标均方根误差和相关系数进行分析, 基于折射率的过氧化值分析模型采用偏最小二乘算法建模预测精度最理想, 选取最优主成分数为6时, 其校正集均方根误差RMSEC为0.168%、 决定系数 R²为0.981, 预测集均方根误差RMSEP为0.231%、 决定系数 r²为0.977; 基于吸收系数的过氧化值分析模型则采用主成分回归算法建模预测模型稳健度最好, 选取最优主成分数为10时, 其校正均方根误差RMSEC为0.192%、 校正集决定系数 R²为0.979, 预测均方根误差RMSEP为0.262%、 预测集决定系数 r²为0.97。 该研究结果的得出, 验证了THz-TDS技术用于食用油过氧化值定量分析的可行性, 为食用油的品质评价找到一种高精度、 性能稳定、 快速无损的检测方法。

At present, Terahertz (THz) spectroscopy techniques are mainly used for qualitative analysis, but the application of THz technology can hardly be found in quantitative analysis in the detection of edible oil’s quality. This paper presents an approach to analyze edible oil quality based on Attenuated Total Reflection (ATR) and Terahertz Time Domain Spectroscopy(THz-TDS). Firstly, the THz-TDS of edible oil samples with different types and degrees of oxidation were collected, the effective signal band was filtered and the optical constants were extracted, the preprocessing algorithm corrected the optical constants, a variety of chemometrics methods were used to establish quantitative analysis models, in order to quickly and accurately predicted the peroxide value of edible oils. 70 experimental samples were used, including soybean oils, rapeseed oils and corn oils, the peroxide value ranged from 0.41 to 10.23 mmol·kg^-1, and the peroxide value distribution of the samples was evenly distributed. A TeraPulse 4000 terahertz pulse spectroscopy system equipped with an ATR detection module belonging to TeraView was used to collect samples’ THz-TDS signals. According to THz-TDS characteristics, the effective band 10 to 86.78 cm^-1 was selected for modeling analysis. The frequency domain signals were obtained by fast Fourier transform, and the optical constants were extracted: refractive index and absorption coefficient. Refractive index and absorption coefficient were preprocessed separately through Savitzky-Golay 7-points convolution smoothing, which had achieved the purpose of removing interference signals. The SPXY algorithm was used to divide the calibration set, and prediction set samples in a 3:1 ratio. The peroxide value analysis models based on refractive index and absorption coefficient were established by the principal component regression algorithm and partial least square algorithm. The root mean square error and correlation coefficient of the model evaluation indexes were analyzed, the peroxide value analysis model based on the refractive index was modeled by partial least squares algorithm had ideal prediction accuracy. When the optimal principal component number was selected to be 6, RMSEC is 0.168%, R² is 0.981, RMSEP is 0.231%, r² is 0.977. The principal component regression algorithm modeled the peroxide value analysis model based on the absorption coefficlent, and the prediction model had the best robustness. When the optimal principal component number was selected to be 10, RMSEC is 0.192%, R² is 0.979, RMSEP is 0.262%, r² is 0.97. This study verifies it is feasible to detect the peroxide value of edible oil by THz technology, and the more important innovation is a high-precision, stable performance, fast and non-destructive detection method for the evaluation of edible oil quality has been found. Furthermore, this research has important guiding significance for improving the safety of edible oil quality and building edible risk assessment systems.