Prediction Of Starch Moisture Content Based On Near Infrared Spectroscopy

Starch quality is closely related to its water content.Although the traditional detection technology can achieve the purpose of prediction,but it is often accompanied by time-consuming,low efficiency,poor accuracy,damage to the sample to be tested and many other shortcomings.Therefore,it is very important to study a rapid,convenient,accurate and nondestructive detection method.This paper uses near infrared spectroscopy technology for nondestructive detection of starch samples,combined with chemometrics algorithms,making preparatory work to starch samples collected,including spectral preprocessing,extracting principal factor number,identifying and eliminating singular samples,and then establish the quantitative calibration model,finally realize the prediction of starch content the.In this paper,using partial least squares regression(Partial Least Squares Regression,PLSR)to establish the prediction model of starch content,because data preparation of modeling prophase has great influence on the stability and prediction ability of the model,this paper focus mainly on the first phase preparations of the model,including the selection of reasonable preprocessing method,extraction the best principal component and the identification and elimination of the singular sample concentration correction.By comparing the evaluation index of the model of quality under different pretreatment methods,choosing a standard normal variate(SNV)preprocessing method which is suitable for the solid sample;and the choice of the best principal component in the modeling is chosen by means of the cross validation mean variance method.The identification of the singular sample design a Monte Carlo method of cross validation(Monte Carlo cross validation,MCCV),the method based on probability statistics theory of Monte Carlo,building a large number of PLSR models,getting the mean and variance of prediction residual difference of all the calibration samples,making the mean-variance distribution curve,a sample that is located in a region of high mean or high standard deviations is tentatively assumed to be a suspect sample.Then using three sets of contrast experiments and t test to verify the singular samples.The comparative experiments are as follows:retention of suspicious samples modeling,elimination of suspicious sample modeling and random elimination of samples with the same number of suspicious samples modeling.Firstly,the evaluation indexes of each model were recorded,and then the t test method was used to analyze the differences among the indexes.The results of t test were used to judge whether the difference was significant or accidental.If there is a significant difference,the identified sample is a singular sample;otherwise,the suspicious sample is not a singular sample.In this paper,Monte Carlo cross validation method was used to identify and validate the singular samples of 100 starch samples involved in modeling.The singular samples were successfully screened,and the reliability of the method was verified.A prediction model of water content was established by using the pretreated starch samples,the test set starch spectrum is input to the model.The feasibility of the prediction model is verified by analyzing the predictive value and actual value of water content.Therefore,near infrared spectroscopy is applied to the design of starch moisture content prediction system.The software design uses Matlab and SPSS.Matlab calls the data in the database for analysis and modeling,and implements the display of the simulation results.SPSS implements data statistics and analysis.