应用数字傅里叶滤波方法提高近红外光谱多元校正模型稳健性的研究

在近红外光谱多元校正方法实际应用中，经常遇到这样的情况，近红外光谱校正模型仅适用于建模时的测量条件，而在测量条件稍有变化时就无法实现样品的准确预测。文章主要研究采用数字傅里叶滤波预处理方法提高近红外光谱多元校正模型稳健性。文章将数字傅里叶滤波预处理方法应用于葡萄糖水溶液的温度实验，实验1和实验2分别在恒温25 ℃和恒温30 ℃进行光谱测量;实验3在未控温的室内环境下进行光谱测量。采用实验1和实验2的样品作为训练集进行模型训练和优化，模型建立完毕之后，采用实验3的样品作为验证集进行模型预测能力评价。结果表明，如果训练集样品未经过预处理而直接建立偏最小二乘(PLS)多元校正模型，则验证集样品均方根预测误差(RMSEP)为664.47 mg·dL-1。而训练集和验证集样品经过傅里叶滤波预处理之后分别进行PLS建模和预测，验证集样品均方根预测误差(RMSEP)降低为58.43 mg·dL-1,样品预测值与参考值的相关性也得到提高。可见，采用数字傅里叶滤波预处理方法可以提高多元校正模型的稳健性。

Application of Digital Fourier Filtering Pretreatment Method to Improving Robustness of Multivariate Calibration Model in Near Infrared Spectroscopy

In the application of near infrared spectral analysis,it often occurs that the multivariate calibration model can only predict accurately the samples whose measuring conditions are the same as those of the training set. When the measuring conditions of the predicted samples change,the calibration model would predict the measured samples with greater prediction errors. In the present paper,digital Fourier filtering pretreatment method was used for the glucose aqueous solution temperature experiments. Experiment 1 was carried out at constant 25 ℃,while experiment 2 at constant 30 ℃. Experiment 3 was performed at ambient temperature. Samples of experiment 1 and experiment 2 were employed as the training set to optimize the calibration model. And then the samples of experiment 3 were used as the validation set to evaluate the prediction error of the model. The results showed that the root mean square of prediction error (RMSEP) of the validation set was 664.47 mg·dL-1 for the partial least squares (PLS) calibration model without digital Fourier filtering pretreatment,while the RMSEP of the validation set was reduced to 58.43 mg·dL-1 for the PLS calibration model with digital Fourier filtering pretreatment,and the correlation coefficient between prediction values and reference values increased. It is showed that the digital Fourier filtering pretreatment method could improve robustness of the multivariate calibration model.