模型转移是解决近红外光谱仪器间存在差异导致校正模型难以在多台仪器间通用问题的重要方法。 利用主成分-马氏距离方法判断样品在不同仪器间的光谱差异性, 然后通过吉洪诺夫正则化约束和校正模型参数, 提出新的模型转移算法, 实现模型在不同近红外光谱仪器上的共享和使用。 首先使用一组标准样品光谱, 建立主机和子机近红外光谱模型预测误差最小化函数。 通过约束主机和子机的模型参数的差异, 求出子机的模型参数, 从而达到模型转移的目的。 该方法应用于药物活性成分和烟叶中总植物碱与总糖的含量分析, 结果表明使用15个标准样品时, 子机光谱样本的预测均方根误差(RMSEP)分别从8.3 mg、 0.49%和1.91%降到3.9 mg、 0.09%和0.83%。 转移后模型预测相对分析误差(RPD)均大于3.0, 子机光谱样本的预测效果得到明显提高。 该方法理论明确、 直观, 在实际应用中样品预测准确性较好, 为具有标准样品的模型转移方法提供一种新思路。

Model transfer plays an important role in solving the problem of the difference between near infrared (NIR) spectroscopic instruments and the prediction difficulty of models. The spectral differences between the same samples taken on different NIR instruments were identified using the principal component-Mahalanobis distance method. Based on the constrain conditions of Tikhonov regularization (TR) and model parameter correction, a new algorithm (called new Tikhonov regularization-based calibration transfer, NTRCT) was proposed for calibration transfer between NIR instruments, so as to facilitate the share and use of the calibration models. The spectra of a set of standard samples were first utilized to establish a specific function that could minimize the prediction errors obtained from the master and slave instrumental models. By constraining the difference of the model parameters, the parameters of the slave instrument model were then determined, to achieve the purpose of model transfer from the master instrument to the slave one. This method was applied to analyze the content of the active pharmaceutical ingredient (API) of tablets and quantify the contents of total alkaloids and total sugars in tobacco leaves respectively, by means of their NIR spectra acquired on different instruments. The results showed that the root means square error of prediction (RMSEP) of samples taken on the salve instrument was reduced from 8.3 mg, 0.49% and 1.91% to 3.9 mg, 0.09% and 0.83% respectively; when 15 standard samples were employed for modelling. As the calibration transferred all the resulting RPD values were larger than 3.0, and the sample predictions from the salve instrumental spectra were thus significantly improved. The method was explicit and intuitive in theory, and had good accuracy in sample prediction in practical applications. It provided a new idea for calibration transfer method with standard samples.