| China is a large country in the production of agricultural products,but there are still many problems such as overcapacity and uneven quality,especially in the fruit and grain industries.With the development of society,the demand for fruits,cereals and their products has transformed from quantity to quality.Therefore,in order to realize the diversity and high-quality development of China's fruit and cereal products,it is necessary to focus on and improve the post-harvest commercialization processing technology of fruits and cereals,especially the technology that can realize rapid,non-destructive and online detection.Studying online detection methods for the internal quality of fruits,cereals and their products and developing rapid,non-destructive,high precision,online detection system will be extremely important for improving commercialization,reducing loss and enhancing the international competitiveness.Based on the near infrared spectroscopy technology,taking strawberries and flour as experimental subjects,this paper studied an online detection method for strawberry quality and shelf life by vis/near infrared spectroscopy and developed an online detection system for flour quality detection by C++.This paper also studied a nonlinear modeling method based on machine learning and deep learning,and constructed a strongly generalized model for rapid online detection by near infrared spectroscopy.The main research contents and results are as follows:(1)The variation regulation between the quality change and the near-infrared spectrum during strawberry storage was analyzed,and the discriminant model of strawberry shelf life at different speeds(0.05,0.10 and 0.15m/s)was established.The results showed that the LDA model established by the original spectrum was affected by the speed and has poor prediction ability.Based on the characteristic wavelength which extracted by CARS algorithm,the PLS-DA model obtained an optimal classification prediction performance,and the model was less affected by speeds.The correct rate of discriminations at three speeds(if not specified,it refers to three speeds from low to high)were greater than 90%,which were 95.1%,97.4%,and 93.3%,respectively.(2)Based on the PLS algorithm,a model for the SSC prediction in strawberries was established using the strawberry spectral data obtained online.The best model obtained in the whole band scope was preprocessed via MSC,SG smoothing and 2~ndd derivative,and the RPD values of models at three speeds were increased from 1.39,1.27 and 1.07 to 1.60,1.53 and 1.48,respectively.These pretreatment methods made the model have a certain predictive ability and also reduced the impact of speed on the accuracy of the model.Extracting the characteristic wavelength by CARS algorithm not only simplified the model,but also improved the computational efficiency and enhanced the robustness of the model.The R_p~2 of models with characteristic variables at three speeds were 0.702,0.733 and 0.707,respectively;RMSEP values were 0.892,0.699 and 0.761°Brix,respectively;RPD values were 1.80,1.96 and 1.87,respectively.(3)An online detection system for flour quality was developed with an online platform using the NIR 1.7/S micro-spectrometer from INSION of Germany and the diffuse reflection integrating sphere with integrated light source.The C++was used for secondary development of the spectrometer,and a control software NIRspec for online detection of flour was independently developed.The experiment of flour moisture content online detection showed that the absolute error of predicted moisture content of all samples was below 3%,and the prediction error of 66%samples was below 2%.The coefficient of determination(R~2)of the actual value and the predicted value was 0.883,and the root mean square error(RMSE)was 0.206%.(4)In response to the limitations of the traditional linear PLS model,a model for online prediction of flour moisture content was constructed by using BP-ANN and SVR algorithm based on machine learning.Based on this,the paper further explored the CNN model based on theories of deep learning.Through experiments,the structure and optimal parameters of the three models were determined respectively,and the results were compared with the results of traditional linear PLS models.It was found that the generalization ability of nonlinear models was stronger,and the prediction accuracy and robustness of SVR models were the best with R_p~2 of 0.887 and RMSEP of 0.277%.At the same time,the results showed that the deep learning-based CNN model also had good predictive ability,with R_p~2 of 0.776 and RMSEP of 0.339%,indicating that the deep learning had certain potential in near infrared spectroscopy. |
PDF Full Text Request