A Study On Method Of IR Spectroscopy Integration And Model Optimization For Measuring Linoleic Acid And Linolenic Acid In Vegetable Oil

Infrared (IR) and near-infrared (NIR) spectroscopy are used in quantitative research, have advantages of quick analysis, free pretreatment and non-destructive, etc., and have been widely applied in food, agricultural products and pharmaceutical and other industries. Spectral analysis is an indirect measurement technique, and relies on specific mathematical model. Therefore, increasing the accuracy and robustness of prediction model is one of the main content of researches in spectroscopy and chemometrics.In this study, a large number of samples including market and homemade oil were collected. The samples were randomly divided two groups, calibration set and prediction set, the number were 201 and 101 respectively. They were esterified, and the content of linoleic acid and linolenic acid were analyzed by gas chromatography.The raw NIR and IR spectra were preprocessed by first-order differential and autoscale, and then were build model with method of PLS. The results showed that NIR spectra model has high prediction accuracy for linoleic acid, and the IR spectra model was good in detecting of linolenic acid. The spectrum was split into 10 intervals, iPLS and siPLS methods were used to build model. For linoleic acid and linolenic acid, siPLS method was better, and the prediction accuracy was higher than other methods.The elimination of singular sample has important role in improving of model robustness and accuracy. RMSEPm combined with RMSECc in Monte Carlo cross-validation was adopted to eliminate the singular sample, which was more reasonable than the single one adopted. It was better than method of leverage value and others. Monte Carlo strategy was also used to optimize intervals in siPLS. Results showed that it selected the best combination of intervals with smaller calculation workload and achieved a better model.Methods of uninformative variable elimination were used to selected variables in building model. Firstly, methods of the introduction of variables including UVE-PLS, UVE-GA-PLS, nUVEvote-PLS and MP-UVE-PLS were used. The results show that the MP-UVE-PLS was the best, has a high capacity in selecting variables and improved the accuracy and robustness of model. The secondly, Monte Carlo methods including the LOO-UVE-PLS, MC-UVE-PLS and MCvar-UVE-PLS was used to eliminate variables, and results show that MCvar-UVE-PLS was the best method.Spectrum integration technology was studied. NIR and IR spectra were integrated with different ways, and the prediction accuracy was increased than a single spectrum, the best way was the first derivative of NIR connected with IR. When uninformative variables integrated spectra were eliminated by UVE-PLS, the accuracy of prediction was increased. Especially for the MP-UVE-PLS, through which many uninformation variables were eliminated, and the best prediction accuracy was achieved. Spectral variables were significantly reduced by wavelet compressionby wavelet with scale 2,4,8 respectively, the model accuracy was high than uncompressed spectra, which suggested that wavelet compression is a good method in increasing model prediction. A large number of uninformative variables of approximate coefficients from wavelet transform were eliminated, and the accuracy of the model was as good as the PLS model with full approximate coefficients. Scale 2 is better, scale 8 is the worst. Therefore, the approximate coefficients from wavelet compression can obtain better PLS model than the original spectrum. However, in the case of uniformative variable elimination, the original spectrum should be adopted in order to obtain better results.Interferogram was denoised by wavelet. For directed air 100% transmission line, when interferogram was denoised by wavelet, SNR (p-p) was increased 41.19%. In polystyrene test, the spectral peak was intact, the spectral shape in the low frequency region was consistent with the original spectra, and there are subtle differences in high-frequency region. For signal of weak SNR from the ATR attachment, wavelet denoising achieved the purpose of denoise. However, SNR (P-P) was improved less, only 4.97%. For signal of weak SNR of oil, spectral shape was unchanged when bior2.4 wavelet was used. Therefore, the strategy of that the improvement of SNR through interferogram denoising by wavelet was achieved.