太赫兹(Terahertz, THz)波通常是指位于微波和近红外之间的电磁波。 由于很多化学和生物分子的振动和转动模式正好都位于THz波段, 因此可以利用物质的这些“指纹谱”特性开展定性和定量分析研究。 目前用于THz光谱的定量分析主要有主成分回归(PCR)以及偏最小二乘回归(PLSR)等方法, 这些算法在建模时往往需要大量的样本进行监督学习, 模型精度对训练样本依赖性较高, 同时模型的外推性不易保证, 在样本量不足或者外推性要求较高的场合, 这些算法的使用会受到一定限制。 针对这些问题, 该研究提出一种利用光谱的参数化模型——间接硬建模方法(IHM), 进行混合物太赫兹光谱解析和量化分析技术方案的研究。 首先使用S-G平滑滤波方法滤除光谱中的噪声影响; 同时考虑到太赫兹光谱特性, 消除了人工基线对光谱解析产生的影响; 随后, 开展了IHM建模与分析研究, 重点讨论了在两个训练样本数目情况下模型预测准确度的问题; 为了验证该算法的可行性, 制备了利福平、 乳糖一水合物、 微晶纤维素以及硬脂酸镁的四元混合物进行实验与建模分析; 使用回归相关系数R和均方根误差RMSE对定量模型进行评价。 将IHM方法和PLSR方法进行了比较, 理论分析和实验结果表明, 相对于传统方法, IHM方法建模所需的训练样本数量可减少至2个, 与此同时量化分析准确度获得了提高, 同时外推性也有所提升。

Terahertz (THz) radiation is the electromagnetic spectrum lying between the microwave and infrared portions, and covers wavelength between 3mm and 30um. Terahertz spectroscopy has been used in qualitative and quantitative analysis since molecular vibrations and rotations of a great large number of materials lie in this region. Many techniques, such as principal component regression (PCR) and partial least square regression (PLSR), have been applied in quantitative analysis of terahertz time-domain spectroscopy. However, those “soft-modeling” methods, though effective in many cases, all require a large number of samples and have poor extrapolation capability, and the utilization is also restricted in some cases (samples lack or requirement of extrapolation). A parametric method of spectrum, Indirect Hard Modeling (IHM), was proposed to resolve terahertz spectral matrix and quantify the mixtures. Savitzky-Golay method was used to eliminate the influence of noise. Moreover, the features of terahertz spectra were considered and the artificial baseline was also resolved as a signal spectrum to improve spectrum extraction accuracy. Subsequently, the analysis of IHM was carried out, and the ability of prediction and accuracy of the model in the case of low training samples were discussed. In order to validate the capability of the proposed method, quaternary systems of Rifampicin, Lactose monohydrate, microcrystalline cellulose and magnesium stearate were tested. The regression correlation coefficient (R) and root mean square error (RMSE) were employed to estimate the performance of the model. Finally, the comparison of IHM and PLSR were carried out, and theoretical analysis and experimental results indicated that the number of samples required by IHM modeling could be reduce to two during calibration, at the same time, the accuracy of quantitative and extrapolation capabilities were also improved.