Study On Near-infrared Spectroscopy Detection Technology And Calibration Transfer Of Fruit Quality

Near-Infrared spectroscopy(NIRS)technology is a widely studied and applied nondestructive testing technology,which can be used to quickly analyze the internal quality of fruits,to improve the quality,scale,and international competitiveness of fruit export in China.However,the internal hardware differences of the spectrometer,the ambient temperature differences,and the different years or places of fruit collection will affect the prediction results of the model,leading to the decline of the prediction performance of the model based on the spectrum of the main instrument when it is directly applied to other secondary instruments.Therefore,to ensure the accuracy of the model and save the modeling cost,the model needs to be standardized to enhance its robustness of the model,which is an urgent problem to be solved in the practical application of NIRS technology.Model transfer(model transfer)is the main method used to deal with this problem.In this paper,a new band selection algorithm based on model transfer is proposed for apple and mango,and the robustness of the new algorithm to improve the transfer model is studied.The main findings include:Firstly,a Relative Error analysis(REA)band screening method based on model transfer is proposed based on the current band screening problem in the field of model transfer.REA can further optimize the transfer effect based on the traditional model transfer algorithm by screening the bands with a consistent and stable spectral response between master and slave instruments and applying them to model transfer.REA method,based on the Mean Relative Error(MRE)of each band,can effectively distinguish the bands with a consistent signal response between master and slave instruments,and eliminate the bands with a large difference in response,to achieve the purpose of the model modification.In this paper,apple and mango experimental data are taken as the object of near-infrared spectrum analysis,and calibration is carried out with and without a standard sample model transfer algorithm to verify the universality of the REA algorithm.Second,based on the model transfer of Soluble solids content(SSC)prediction model of apple and mango,an optimization strategy was adopted combining the band selection algorithm with the standard sample model transfer algorithm.In this paper,Piecewise Direct standardization(PDS),Slope/bias correction(Slope/bias correction,SBC),and Spectral Space Transformation(SST)are used to calibrate the model,and REA is combined with Competitive Adaptive Reweighted Sampling(CARS),continuous projections algorithm,SPA)and Uninformative variables elimination(UVE).The results showed that the prediction performance of the apple and mango SSC detection model was greatly improved after SST was combined with REA.The Root means square Error of Prediction(RMSEP)decreases from 1.109 and 2.433 to 0.625 and 1.220,respectively.The Rp increases from-0.563 and0.470 to 0.735 and 0.897,respectively.Therefore,the strategy combining REA with the model transfer algorithm with standard samples can effectively realize the transfer of the Mango SSC prediction model between NIRS devices.Thirdly,based on the model transfer of apple origin NDT model between instruments,an optimization strategy combining band selection algorithm and model transfer algorithm without standard samples was adopted.This paper introduces three Transfer Components analysis(TCA),Balanced Distribution Adaptation(BDA)algorithm,and Manifold Embedded distribution alignment(MEDA)are used to transfer models without standard samples.Four band screening algorithms(REA,CARS,SPA,and UVE)were used to optimize the transfer model.The results showed that MEDA combined with REA had the best correction effect on the transfer model,and the accuracy of the model increased from 84.45% to 94.54%,which was very close to the test result of the model on the original platform(accuracy 96.22%).Therefore,REA can improve the correction performance of model transfer without standard samples to a certain extent,and this method is expected to be popularized and applied to the model transfer of near-infrared portable instruments,promoting the rapid development of NIRS analysis technology in the field of nondestructive testing.