Study On Static And Dynamic Nondestructive Methods For Moldy Detection In Maize

After being harvested,maize is highly susceptible to mold contamination,and then mold metabolism produces mycotoxins which harmful to humans and animals.Aiming at the problems of poor timeliness and low sensitivity of traditional detection methods for grain mildew,this paper mainly studies the change rules of spectrum,image information and volatile odor of maize contaminated by harmful mold.Firstly,the static detection methods based on Fourier transform near-infrared spectroscopy and electronic nose technology are studied.Then,this paper makes a further study of the dynamic detection methods based on visible/near infrared spectroscopy and image processing technology.The main conclusions of this study are as follows:1.Having established a discriminant model of maize mildew based on spectrum and electronic nose under static.The results of Fourier transform near-infrared spectroscopy showed that the main absorption peaks of maize samples were located at4538cm^-1,4858cm^-1,5699cm^-1,5857cm^-1and 6966 cm^-1,it closely related to the frequency combination and doubling of C-H,N-H and O-H groups in proteins,linear lipids and aromatic hydrocarbons.The results of electronic nose showed that,as the mildew degree deepens,the response signals of T30/1,T70/2,PA/2 and P10/1 which represent hydrocarbons,aromatic compounds and nitrogen compounds gradually weakened,while the response signals of LY2/G,LY2/AA and LY2/GH which represent alcohol and ketone gradually increased.By comparing the qualitative and quantitative models of maize mildew based on spectroscopy and electronic nose,it can be known that the discriminant accuracy of near-infrared spectroscopy is better than electronic nose's.The prediction correct rate of the optimal Linear Discriminant Analysis?LDA?model for all maize samples is 90%,which is 17.3%higher than that of the electronic nose model;the root mean squared error of prediction?RMSEP?of the optimal Partial least squares regression?PLSR?model is 0.620 log CFU/g,which is 58%lower than that of the electronic nose model;and the residual predictive deviation?RPD?is 3.45,which is 62.7%higher than that of the electronic nose model.The results showed that the technologies of Fourier transform near-infrared spectroscopy and electronic nose were feasible for maize mildew detection.2.Having established a dynamic discriminant model of maize mildew based on spectrum and image.The results of visible/near infrared spectroscopy showed that the main absorption peaks of mildewed maize samples were at 962,1143 and 1411 nm,which corresponded to the second fundamental vibration of C=O bond,N-H bond in protein,C-H bond in fatty acid and the first fundamental frequency vibration of C-H bond in starch,respectively.The results of image processing showed that the mold metabolism could cause the external quality changes of maize,such as dim surface color and low gloss.Having established LDA and PLSR models of all maize samples based on visible/near infrared spectroscopy.At high speed,the prediction correct rate of the optimal LDA model was 86.7%,RMSEP and RPD of PLSR models were 0.823 log CFU/g and 2.46,respectively.LDA and PLSR models were established for all maize samples based on images,at high speed,the predictionc orrect rate of LDA models was82.2%,RMSEP and RPD of PLSR models were 0.717 log CFU/g and 2.93,respectively.Comparing the results of maize mildew models at different speeds,the difference between robustness of models at low and high speeds is not significant.3.Having established a dynamic discriminant model of maize mildew based on the combination of spectral and image information.The results showed that the spectral and image information fusion optimized the correct rate and improved the robustness of the model.At high speed,the prediction correct rate of LDA model of combination of spectral and image was 91.1%,which was 5.1%and 10.8%higher than that of spectral and image models respectively;RMSEP of PLSR model was 0.665 log CFU/g,which was 7.3%and 10.8%lower than that of spectral and image models;RPD value was 3.06,which was 24.4%and 4.4%higher than that of spectral and image models.4.Comparing with the static and dynamic models,the accuracy of the static spectral model was higher than that of the dynamic spectral and image models.The interference of on-line detection to the model was reduced by the spectral and image information fusion.The qualitative analysis model of maize mildew was optimized,and the accuracy of the quantitative model was still lower than that of the static model.The results showed that the combination of visible/near infrared spectroscopy and image processing technology had practical application potential for on-line detection of maize mildew.