将经典的卡尔曼滤波器与近红外光谱分析技术相结合, 提出了一种新的特征波长变量选择方法——卡尔曼滤波法。 分析了卡尔曼滤波器用于波长优选的原理, 设计了波长选择算法并将其应用到大豆油脂酸价的近红外光谱检测中。 首先利用偏最小二乘法(PLS)对油脂不同吸收波段建模, 初步筛选出4 472～5 000 cm-1油脂酸价特征波段共132个波长点, 然后进一步利用卡尔曼滤波器进行特征波长选择, 从中优选出22个特征波长变量建立PLS校正模型, 预测集决定系数R2、 预测误差均方根RMSEP分别为0.970 8和0.125 4, 与利用132个波长点建立的校正模型预测结果相当, 而波长变量数减少到原来的16.67%。 该波长变量选择算法是一种确定性的迭代过程, 无复杂的参数设置和变量选择的随机性, 物理意义明确。 优选出少数对模型影响较大的特征波长变量以代替全谱建模, 在简化模型的同时提高了模型的稳健性, 为开发专用油脂近红外光谱分析仪器提供了重要参考依据。

Combining classical Kalman filter with NIR analysis technology, a new method of characteristic wavelength variable selection, namely Kalman filtering method, is presented. The principle of Kalman filter for selecting optimal wavelength variable was analyzed. The wavelength selection algorithm was designed and applied to NIR detection of soybean oil acid value. First, the PLS (partial least squares) models were established by using different absorption bands of oil. The 4 472～5 000 cm-1 characteristic band of oil acid value, including 132 wavelengths, was selected preliminarily. Then the Kalman filter was used to select characteristic wavelengths further. The PLS calibration model was established using selected 22 characteristic wavelength variables, the determination coefficient R2 of prediction set and RMSEP (root mean squared error of prediction) are 0.970 8 and 0.125 4 respectively, equivalent to that of 132 wavelengths, however, the number of wavelength variables was reduced to 16.67%. This algorithm is deterministic iteration, without complex parameters setting and randomicity of variable selection, and its physical significance was well defined. The modeling using a few selected characteristic wavelength variables which affected modeling effect heavily, instead of total spectrum, can make the complexity of model decreased, meanwhile the robustness of model improved. The research offered important reference for developing special oil near infrared spectroscopy analysis instruments on next step.