D-S Evidence Theory Fusion Classification Based On Apple Internal Quality

Near infrared nondestructive testing technology has been widely used in nondestructive testing and classification of fruit internal quality due to its advantages of short testing time,simple operation and low testing cost.Near infrared is an indirect technology and needs to establish a prediction model.At present,the research mainly focuses on the construction and improvement of spectral data preprocessing model.However,the current classification model still has some problems,such as poor adaptability of single prediction model and inaccurate classification caused by hard segmentation: different application objects are suitable for different modeling;The data update model needs to be retrained;The reliability of data affects the accuracy of prediction,which seriously affects the classification accuracy in hard segmentation.Therefore,in order to solve the influence of uncertain incomplete information on a single classification model,this thesis studies the problem of multi-model fusion based on DS evidence theory,and combines the advantages and effectiveness of each model to improve the stability and prediction accuracy of the classification model.Near infrared spectroscopy and Soluble Solid Content of 439 Red Fuji apples were used as sample data.In order to make the fusion of multiple models better and the characteristic of single model,this thesis uses Probabilistic Neural Networks and Partial Least Squares to predict the SSC of apple respectively,in which Multivariate Scattering Correction is selected to correct the baseline of the original spectral data,and Genetic Algorithm is used to screen the characteristic wavelengths to establish a stable single prediction model with high accuracy.Mass function generation is one of the key steps in D-S evidence theory,but most of its generation is not fixed,which is based on the evidence obtained by experts.The discounted factor trapezoidal linear function of triangle is designed according to SSC and the distance of classification boundary,and the result of triangle function combined with PNN and PLS model is found by Dempster combination rule.Furthermore,in order to optimize the distribution of classification samples,this thesis designs a quadratic polynomial with one variable and a nonlinear function of Gaussian distribution type,and compares it with Dempster combination rule to find the classification result after Gaussian function fuses the two models has the best classification result.Combination rules are also an important part of D-S evidence theory.In this thesis,the conflict problem in practical application is studied to solve the conflict problem by studying combination rules.In this thesis,the classification results of several functions generated under the action of Yager combination rule,Sun Quan combination rule and Li Bicheng combination rule show that the combination of Gaussian function generation and Li Bicheng combination rule has the best effect on apple classification.To sum up,the purpose of this thesis is to study the design of functions in D-S evidence theory and the influence of functions and combination rules on apple classification.Through examples and simulation experiments,the combination suitable for apple classification is studied This method can effectively solve the instability problem of Apple single classification prediction model based on near infrared technology,better express the uncertainty in hard segmentation and improve the prediction accuracy of the fusion model.