Research On Detecting Methods For Soybean Oil Quality Based On Near-infrared Spectrum Analysis

Edible oil is an important part of the human diet and one of the three major nutrients of the human. Its quality situation will affect the development of food industry, health security of consumers and social harmony and stability. There are many indicators evaluating edible oil quality, which should be detected constantly in processing. At present, our country adopts mainly the traditional chemical methods to determine edible oil quality. There are many ubiquitous problems, such as trivial determination program, slow determination speed, more chemical reagent and the complex sample pretreatment and so on. These methods are only adapted to laboratory test, can not satisfy the requirement of the modern society for detecting edible oil quality which should be simple, rapid, accurate and on-site.In recent years, near infrared spectrum analysis technology is rising rapidly in process analysis and industrial control areas, its advantage is rapid, low consumption, no pollution, without sample pretreatment, and can meet multiple analysis requirement synchronously, more suitable for quality inspection and quality control in processing. Using near infrared spectrum analysis technology to detect main quality parameters of edible oils rapidly can overcome the drawbacks of conventional methods, and possesses of important realistic significance to improve product quality of edible oil, realize dynamic monitoring oil processing.Taking soybean oil for example, which is the first edible oil in our country, on the basis of analyzing real requirement for parameters detection in oil processing and storage, this paper presented research project of detecting three major parameters evaluating oil quality named acid value, peroxide value and color by near infrared spectrum analysis technology, collected representative soybean oil samples, determined accurately their chemical values by standard methods, made data acquisition of near infrared transmission spectrum synchronously, researched in detail every kind of data processing and modeling methods adopted in soybean oil near infrared spectrum analysis.Firstly, the regression calibration model between chemical value and spectral data of oil acid value is established using classical partial least-square linear modeling method. The best modeling band is selected. In order to improve the model’s prediction accuracy, denoising method of oil near infrared spectrum based on wavelet transform is researched. The Daubechies series wavelets are selected to preprocess near infrared spectral signal of oil acid value. Several aspects including selection of wavelet base, determination of decomposing scale and threshold way are compared and analyzed in detail. The optimal wavelet parameters suitable for denoising oil near infrared spectrum are chosen. Comparing to other conventional spectral data pretreatment methods, the superiority of wavelet denoising is showed. The prediction accuracy of model has been markedly improved after wavelet denoising, decision coefficient R2 and RMSEP of prediction set achieved 0.9936 and 0.0610, respectively, relative standard deviation of prediction is 3.629%. The model can meet practical detecting requirement.Secondly, aiming at nonlinear response problem in near infrared spectrum analysis, application of BP artificial neural network nonlinear modeling method in oil peroxide value near infrared spectrum analysis is investigated. The BP neural network with three layer structure is designed. The main parameters affecting network performance including the number of hidden neurons, momentum factor, learning rate, learning times and so on, are optimized. The BP neural network correction model of oil peroxide value is established. The prediction result is better than that of PLS linear correlation model. In order to improve the model’s prediction accuracy further, wavelet denoising method is applied again to preprocess oil peroxide near infrared spectrum. The BP neural network correction model utilizing reconstructed spectrum after wavelet denoising is established. The decision coefficient R2, RMSEP achieved 0.9938 and 0.2379, respectively, the prediction relative standard deviation is 3.512%. The new modeling method for near infrared spectrum analysis of oil peroxide is offered.Next, the near infrared spectrum analysis method for oil color detection based on SVM is discussed. According to the particularity of soybean oil color, the three oils with different Lovibond yellow values are classified by C-SVM firstly. The comparison analysis is made from the aspects of spectrum preprocessing, kernel function selection, kernel parameter optimization, etc. The SVM classifier suitable for near infrared spectral recognition of soybean oil color is designed, recognition correct rate of different level soybean oils achieved 100%. On this basis, aiming at different oils with different Lovibond yellow values, the regressions between near infrared spectral data and Lovibond red values are made by usingε? SVM. The influences of different kernel functions and different kernel parameters on the prediction accuracy are analyzed, and the best SVM correction models aiming at different level soybean oils are established, the prediction errors are within 0.2 Lovibond unit. Research result demonstrated that using near infrared spectrum analysis to detect soybean oil color is feasible.At last, feature wavelength variable optimization of near infrared spectrum based on Kalman filter is lucubrated. The principle of Kalman filter for selecting optimal wavelength variable is analyzed. The wavelength selection algorithm is designed and programmed. This algorithm is applied to select optimal wavelength variables of soybean oil acid value and peroxide value in modeling bands. The correction models are established using a few selected feature variables which affected modeling effect heavily, and compared with the models using full spectrum, the prediction results are quite. The complexity of models is decreased. The amount of calculation is reduced, at the same time the robustness of models is improved. The research offered important reference for developing special oil near infrared spectrum analysis instruments on next step.