Optimization Of Near Infrared Model Of Wheat Flour Protein By Dragonfly Algorithm

The protein content in wheat flour production affects the quality of the final product,so it is necessary to quickly detect the protein content of wheat flour in the production process,so as to adjust the production process parameters and ensure the high quality of wheat flour production.The traditional chemical determination method is time-consuming and unable to achieve timely feedback.The near infrared spectroscopy technique has the characteristics of high efficiency,fast speed,simple and non-destructive,and can detect the protein content of wheat flour online and in real time.Due to the diversity of samples,the complexity of measurement information and the overlap of absorption bands of each molecular group in the NIR region,NIR spectral analysis requires the extraction of characteristic information from complex,overlapping and changing spectra by computer and chemometrics methods for qualitative and quantitative prediction.In the chemometrics application research of NIR spectroscopy,it is an important means to improve the accuracy and reliability of model analysis to introduce various swarm intelligence algorithms to optimize the modeling parameters of NIR spectroscopy.In this paper,the feasibility of the application of Binary Dragonfly Algorithm(BDA),a new swarm intelligence optimization technology,was investigated for the selection of characteristic wavelength,calibration set and model transfer standard sample set optimization in the near infrared analysis modeling of wheat flour protein.The aim of this study was to establish an optimal model with high accuracy and strong robustness and to improve the universality of NIR analysis model of wheat flour protein between different instruments.The research contents are as follows:(1)Attenuation Elimination Binary Dragonfly Algorithm(AE-BDA)was proposed by using exponential decay function and linear decay function to improve the BDA Algorithm,which was used to select the characteristic wavelength of wheat flour protein in NIR spectrum.Taking NIR spectra of 160 samples of different wheat flour as the research object,the Single-BDA and AE-BDA respectively were employed to select characteristic wavelength of wheat flour proteins.The quantitative model of wheat flour protein was established by partial least square regression method to evaluate the wavelength selection effect.Compared with Single-BDA,the selected wavelengths by AE-BDA were of fewer number but more stable as well as the established model had better prediction performance with the prediction determination coefficient(R_p~2)of 0.9727 and the root mean square errors of prediction(RMSEP)of 0.2811.The average number of characteristic wavelengths selected from 8 experiments was 15.8,accounting for 12.6%of the original wavelengths,in which 3 wavelengths were targeted each time.According to the interpretive near-infrared spectroscopy,the selected wavelengths were contained in the main absorption bands of wheat flour protein and background components.The results indicate that AE-BDA can select less characteristic wavelengths from near infrared spectrum of wheat flour with high computational efficiency to establish model for protein analysis with higher accuracy and stability.The proposed method can provide a much simple and more effective wavelength optimization strategy for NIR analysis modelling.(2)The K/S-BDA algorithm is combined with the BDA algorithm and the K/S algorithm,and the prediction performance of the model established by the selected correction set is discussed.In the iterative optimization process,the fitness function is constructed from the sum of the standard error of cross validation of the calibration set and root mean square error of prediction of the validation set.The spectra of 160 wheat flour samples were collected by NeoSpectra Micro Fourier near-infrared spectrometer.The quantitative prediction model of wheat flour was established by 7 different pretreatment methods and partial least squares algorithm,and the prediction set was predicted.The calibration set is optimized by K/S algorithm,SPXY algorithm,K/S-BDA algorithm and K/S-multi-BDA algorithm respectively.Compared with other methods,the prediction effect(R_p~2 of 0.9564,RMSEP of 0.2781)of the calibration set optimized by K/S-BDA method is the best,and the sample number of the optimized calibration set is about 30%of the original calibration set.The results showed that after the calibration set was optimized by K/S-BDA,the NIR analysis model of wheat flour protein established by PLSR was more stable,has a higher prediction accuracy which provided a more convenient and effective optimization method of calibration set for NIR application.(3)The BDA-DS integration algorithm which combines BDA and Direct Standardization(DS)was investigated to evaluate the performace on model transfer.In this experimental research,Lengguang S450 grating scanning spectrometer was defined as the master instrument while NeoSpectra Micro Fourier transform(FT)NIR spectrometer was the target instrument.Both of them were adopted to collect the spectra of 160 samples respectively,and then established the correlation model between NIR spectra and wheat flour protein by Partial Least Squares Regression(PLSR)method.After the model transfer of BDA-DS,the master model has a significant predictive performance on the target prediction set.The prediction determination coefficient of 0.9812,the root mean square errors of prediction of 0.1838,and the average Mahalanobis distance between the master and target spectra decreases from 22.34 to 1.40,all of which are close to the prediction accuracy of master prediction set.The conventional K/S-DS method was also employed for comparative investigation,and the results show that the transfer set selected by the BDA-DS method not only can better characterize the differences between the master and target instruments with fewer transfer sets but also improve the prediction accuracy of the transferred spectra measured on target instrument.The proposed method can provide a more effective selection strategy to optimize the standardization samples for NIR calibration model transfer.