Research Of Near Infrared Spectroscopy Analysis Prediction Model Based On Intelligent Method

Detection of internal content of fruits, vegetables and other agricultural products is animportant research topic. With the development of molecular spectroscopy, the possibility ofnon-destructive testing comes true. Replacing traditional chemical processes, near infraredspectroscopy analysis technology is becoming a fast, effective detection instrumentality. Withapple and lotus root as research objects in this paper, and making use of chemometric methodsand theory, the research is carried out to establish the near infrared spectroscopy quantitativeanalysis model.This paper has collected a total amount of101apple samples and80lotus root samples,the total sugar content of apple, moisture and starch content of lotus root, is detected by usingstandard method respectively. According to the analysis of the sample set, the singular sampledata is distinguished and deleted,100apple samples and72lotus root samples are selectionand involved in modeling and forecasting. Then, spectral preprocessing methods are utilizedin the NIR data to smooth and denoise.The quantitative analysis model of the near-infrared spectroscopy is a kind of modelbased on data, so multiple regression method is need to be used to analyze the spectral dataand material content. This paper established basic correction model with partial least squaresregression method. In view of the linear characteristic of partial least squares regression.Nonlinear polynomial partial least squares regression and artificial neural network modelingmechanism is introduced and nonlinear quantitative analysis model is established. Bycomparison and analysis, PLS-BP method can access the best modeling results; the model’sperformance has been greatly improved. In Apple’s total sugar content model, the calibrationset correlation coefficient is0.992, the prediction set has a correlation coefficient of0.942. Inlotus root moisture, starch content model the correlation coefficient of the prediction set is0.923and0.938.To solve the problems of the huge data size and multicollinearity of the NIRspectroscopy, near-infrared wavelength region selection method is presented to find theimportant wavelength variable. Backward synergy interval partial least squares, ant colonyoptimization algorithm, and mean impact value method with noise contrast are involved inthis paper. By using the three methods above, wavelength interval and variable was reduced,the stability and performance of quantitative model were improved. In apple’s total sugarcontent model, the calibration set correlation coefficient is0.974, the prediction set has acorrelation coefficient of0.968.Finally, to solve the shortness of ANN, ensemble of neural network method isintroduced, Bagging-ANN model is established and the randomness of artificial neuralnetwork modeling is overcome. The model is stable and effective. The apple’s total sugarprediction set has a correlation coefficient of0.977...