Study On Optimization Method For Rapid Detection Of Soybean Protein Based On Portable Near Infrared Spectrometer

Soybean is China’s main economic crop.The characteristics of wide distribution and small-scale planting make it difficult to purchase on-site.At present,the main basis for soybean purchase pricing in the market is the crude protein content of soybean.For small soybean purchasers and soybean farmers,traditional methods of determining protein content of soybeans based on experiences lack credibility,which make them difficult to get convinced in the circulation of soybean market.However,the national standard method of soybean protein content detection has such problems as complex operation,long detection time,and interference from human factors etc.As a rapid detection method,near-infrared spectroscopy has the characteristics of fast sampling speed and simple operation,and it is one of the important methods for non-destructive testing of agricultural products quality.The portable near-infrared spectrometer based on micro-electromechanical system(MEMS)has the characteristics of small volume,shock resistance and high accuracy.It has been widely used in field rapid analysis and detection in recent years.Under guidance of the instructor of our school and the corporate tutor of Wuxi Xunjie Guangyuan Technology Co.,Ltd.,this paper studies the optimization of the hardware structure and modeling algorithm of the IAS-2000 portable near infrared spectrometer of the company,and improves its influence on the stability of soybean crude protein detection due to the reasons of soybean granularity and surface inhomogeneity,so as to improve the instrument’s influence on soybean crude protein detection.The accuracy and stability of protein determination ensures the interests of soybean farmers and purchasers,standardizes the circulation market behavior,and provides reliable and accurate detection methods to meet the needs of rapid detection.The main research contents are as follows:(1)IAS-2000 instrument performance index test.The performance of IAS-2000 was tested from the instrument’s absorbance linear range,wavelength accuracy,wavelength repeatability,absorbance repeatability and spectrometer baseline drift.The test results show that the linear range of absorbance at wavelengths of 1 064 nm and 1 550 nm is better than 0.1 AU ~ 1.7 AU;the wavelength accuracy of the instrument is 0.64 nm;the wavelength repeatability is better than 0.04 nm;the absorbance repeatability ranges from 900 nm to 1650 nm;the repeatability is 0.000 5 AU,and the wavelength drift is large in 1 650 nm ~ 1 700 nm band.The standard deviation of the 100% T line is 0.000 3 during the baseline drift test.It can be seen that the instrument has a high detection accuracy and stability at 900 nm ~ 1 650 nm.(2)In order to verify the feasibility of using IAS-2000 to detect the crude protein content of whole soybean,60 soybean samples from Northeast China were selected and prediction accuracy of the model was compared by establishing a calibration model of soybean powder and whole soybean.The results showed that Rp,RMSECV and RMSEP of the crude protein model of soybean powder samples were 0.910,0.795 and 0.842,respectively.The Rp,RMSECV and RMSEP of the crude protein model of whole soybean samples were 0.706,1.114 and 1.318,respectively.The experiment proves that there is a good mathematical relationship between the near-infrared spectrum and the crude protein content of soybean powder or whole soybean in the range of 900 nm ~ 1 650 nm.However compared with soybean powder,the mathematical relationship of the whole soybean spectrum is poor.In order to obtain the ideal detection results,it is necessary to optimize the hardware structure and modeling method of IAS-2000.(3)In order to increase the capability of IAS-2000 to receive effective information from the near-infrared spectrum of whole soybeans and improve the accuracy of crude protein detection,from the aspects of light source intensity and light spot collection,the effects of hardware configuration of light source intensity(single source(10 W)and dual sources(20 W))and the incident slit(25 μm and 50 μm)on prediction accuracy of calibration model were tested.The results showed that the whole soybean crude protein model with single light source +50 μm slit configuration had the best prediction accuracy,among which Rp,RMSECV and RMSEP were 0.862,0.693 and 0.776,respectively.It is concluded that the selection of 10 W single source + 50 μm slit configuration in the whole soybean crude protein detection can increase effective information of IAS-2000 on the near-infrared spectrum of whole soybean and improve the detection accuracy of crude protein.(4)In order to establish a calibration model of whole soybean crude protein with high prediction accuracy by IAS-2000,587 soybean samples from Northeast China were taken as the research object,and the prediction accuracy of the calibration model was used as a criterion to compare the effects of various pretreatments,sample set partition,wavelength selection and modeling methods.The results showed that,when MCSV outlier +SPXY method was used to divide the sample set,the crude protein PLS calibration model established by using SNV pretreatment and 49 wavelength points from CARS screening had the best quality.Its main factor number was 9,RMSECV and RMSEP were 0.389 and 0.418 respectively,the prediction effect was better than other models.The test result proved that the detection precision of whole soybean crude protein based on IAS-2000 can meet the needs of rapid detection of crude protein in soybean production and circulation.(5)Regarding model sharing problem of IAS-2000,the model transfer algorithm such as slope/intercept algorithm(S/B)and direct correction algorithm(DS)is used to transfer the host model to the same type and different models from IAS-2000.On-board applicability research was conducted through using 7 standard sets,20 predictive samples,the prediction accuracy after transfer as the standard to find the optimal transfer matrix size and transfer algorithm.The results show that after the DS algorithm is transferred,for the same type of instrument,when the number of standard samples is 20,the predicted results of RMSEP and Rp are 0.569 and 0.945.For different types of instruments,when the number of standard samples is 25,the predicted results of RMSEP and Rp are 0.649 and 0.922.These experiments show that the transfer model between the master and the slave can be realized by a certain model transfer algorithm.The optimal model transfer algorithm between different instruments is different.The model transfer can eliminate or partially eliminate the difference between near-infrared spectroscopy instruments,meeting the needs of large-scale popularization and application of near-infrared instruments.