Research On Optimization Of Blueberry Pomace Detection Model Based On Near Infrared Spectroscopy

Blueberry pomace is rich in phenolic substances,of which anthocyanin is the main object of value-added research of blueberry pomace.Compared with traditional detection methods,it is expensive,time-consuming and laborious,toxic and easily pollutes the environment the low-cost technical features can be used as a new detection method for anthocyanins in blueberry pomace,but there is little research on anthocyanins in blueberry pomace by near infrared spectroscopy.This paper uses near infrared spectroscopy technology to optimize the detection model of the three blueberry pomace of Northland,Bluebeauty No.1 and Britewell.The main research contents are as follows:(1)In view of the current situation of less research on the classification model of blueberry pomace varieties,the near infrared spectroscopy technology was used to optimize the classification model of different blueberry pomace varieties.First,principal component analysis-Mahalanobis distance(PCA-MD)was used to remove the abnormal spectrum,and then the random selection(RS)and kennard-stone(K-S)were used as sample sets partition method,and finally established and trained convolutional neural networks(CNN)and support vector machine(SVM)models respectively.The results showed that the established CNN classification model has an accuracy rate of 100%for unknown samples,which the classification speed was 19.9ms/pc,and the classification effect was better.(2)The sample set division,pretreatment,wavelength screening method selection and combination comparison of the obtained blueberry pomace near infrared spectrum data were performed in turn,and the effects of different method combinations on the performance of the PLS model were compared to optimize the PLS model.The results showed that the sample set divided by the K-S method,and then using orthogonal signal correction+multiplication scattering correction(MSC)+savitzky-golay(SG)convolution smoothing+savitzky-golay convolution second derivative(SGCSD)preprocessing,the partial least squared(PLS)model established by competitive adaptive reweighted sampling(CARS)screening 25 characteristic wavelengths had the best performance,and its R_c~2>0.85,ratio of standard deviation of the validation set to standard error of prediction(RPD)>2.5 verified the accuracy and effectiveness of using near infrared spectroscopy to detect anthocyanins in blueberry pomace.(3)Compared with the performance of the berry anthocyanin detection model in the existing research,the performance of the PLS model needs to be further improved.In this regard,a CNN quantitative model of anthocyanins of blueberry pomace was established,and CNN-GA was proposed and used for parameter initialization method,data enhancement factor and hidden layer composition,and automatic parameter adjustment of hyperparameters to realize the CNN quantitative model.optimization.The results show that the CNN optimal quantitative model has R_c~2 of 0.9808,R_p~2 of0.9576,and RPD of 4.8194,which are superior to the blueberry pomace anthocyanin PLS optimal model and the existing berry anthocyanin detection model,and its detection speed is 5.8 ms/pc.