| With the improvement of living standards,consumers have also put forward higher requirements when purchasing fruits,from the initial effort to external quality(size,color,shape),to pay more attention to internal nutrition(such as sugar,acidity,vitamin content,etc.).Near-infrared spectroscopy technology has the advantages of non-destructive,rapid and accurate,and is widely used in non-destructive testing of internal quality of fruits.Due to the large differences in the internal physical and chemical properties of different types of fruit,the resulting spectral response is also different.Therefore,it is usually necessary to model the fruits of each variety individually,which makes the maintenance and optimization of the model time-consuming and labor-intensive.In this paper,based on the machine learning algorithm,the spectral data of various kinds of fruits were excavated,and the modeling method of general model was studied to improve the fruit detection accuracy.In this paper,four different kinds of fruit sugar(apple and pear)as the research object,using near infrared diffuse reflectance spectroscopy,combined with machine learning algorithms for data preprocessing,modeling,wavelength optimization and other analysis.A variety of spectral pretreatment methods were compared,and by anon-linear model random forest to mode and verify the feasibility of the universal model for the detection of sugar in four types of fruits.The model is optimized and improved the prediction accuracy.by wavelength and band optimization.Spectral measurements were performed using a miniaturized spectroscopic module based on Digital Light Procession(DLP)spectroscopy to verify the feasibility of the general model and wavelength optimization method on the spectroscopic module.The main research contents and conclusions are as follows:First,the pretreatment method of spectral data was studied.The near-infrared spectral data was preprocessed by smooth denoising,baseline correction,derivation,multiple scattering correction,and compared with the model established by the original spectrum to optimize the pretreatment method.The analysis showed that the effect of pretreatment on a single variety of fruit was not significantly improved,and the multiple scattering effect was slightly better.Second,the study of a variety of generic models of fruit.The Random Forests was used for many kinds of fruit spectrum,by comparing with the effect of partial least squares,multiple linear regression model,it is predicted that R_p~2 increased from 0.731 to 0.888 and RMSEP decreased from 1.148 to 0.334,which greatly improved the prediction effect of the model.It was proved that the feasibility of using this method to study the general model of various types of fruit sugar.Then,the study of characteristic wavelength optimization methods.Binary particle swarm optimization(BPSO)and genetic algorithm(GA)are combined with partial least squares to perform spectral wavelength optimization.BPSO-PLS not only reduced the wave number from 1557 to 817,but also reduced the computational load.Moreover,the R_p~2 of the model increased by 0.831 from0.731,and the RMSCP decreased from 1.149 to 0.742,which improved the prediction accuracy.GA-PLS selects 7 bands out of 1557 wave numbers,with a total of 210 wave numbers,which greatly reduces the amount of calculation,and the difference between the model effect and the full wave number was not significant.It has been proved that the preferred wavelength can not only reduce noise,reduce the amount of calculation,but also improve the accuracy of model prediction.Finally,a portable spectroscopic module based on Digital Light Procession(DLP)spectroscopy is used for the portable and rapid determination of fruit sugar.The module was used to measure the diffuse reflectance spectra of fruits.The Random Forest and binary particle swarm combined with partial least squares were used for model analysis.The feasibility of a general model modeling method and wave number optimization method for various kinds of fruit sugars was verified. |
PDF Full Text Request