| Grain is a strategic material related to the national economy and people’s livelihood,and a necessity for human survival.Food security is related to economic development and social stability,and is a major prerequisite for national security.Grain reserve is an important material basis for guaranteeing national food security,and it is an indispensable link from production to consumption.In the process of food storage,quantity loss and quality loss will occur,among which the quality loss will lead to the deterioration and spoilage of food.If human beings eat the deterioration and spoilage of food,it will have adverse effects on health.Therefore,it is of great theoretical value and practical significance to study how to reduce the loss of grain storage quality and improve the quality of grain storage in improving the level of national food security and ensuring people’s health.With the continuous development and innovation of machine learning methods,the rapid improvement of computer hardware and software,and the gradual application of cloud computing,the application prospects of machine learning in the analysis and prediction of grain condition data have become broader.The quality of stored grain is traditionally determined through laboratory testing methods such as physics and chemistry.This method requires complex steps of sampling and testing,which increases the operational decision cycle of the management of stored grain and the risk of serious deterioration of grain.The big data of grain condition has the characteristics of fast collection and large amount of data.Based on machine learning method,this paper proposed 2 prediction models of stored grain quality based on support vector regression and the corresponding optimization algorithms.On the basis of fully considering the characteristics of stored grain data,relevant storage factors were selected as the input characteristics of the models,and the prediction of fatty acid value and tasting assessment value of paddy during storage was carried out in depth,which gave full play to the advantages of machine learning method in the discrimination of stored grain state and quality prediction.The primary work and results are as follows:(1)Research on the analysis and modelling method of stored grain quality based on grain condition data.Through the collection of a large amount of stored grain data,the variation law of paddy quality during storage was studied.In the paper,the traditional data fitting method and machine learning prediction method for grain storage quality were discussed,the advantages and disadvantages of the traditional data fitting prediction method and machine learning prediction method in grain storage quality prediction were compared,and the applicability of machine learning prediction method in grain storage quality prediction was discussed.Based on this cognition,the general process of grain quality prediction based on machine learning was given.On the basis of grain condition data,the input characteristic parameters and prediction objectives in the modeling process were given.(2)Research on data preprocessing methods.The repeated,missing and abnormal situation of stored grain temperature historical data were analyzed and processed.The repeated detection data were repaired by means of average method,and the missing temperature data were repaired by linear interpolation method.A grain temperature prediction algorithm based on sliding window was proposed to judge abnormal temperature data.Longitude and latitude of grain depot,granary type,initial moisture content,detected moisture content of stored grain,the month when warehousing and sampling,average temperature of stored grain and granary,storage period,effective accumulated temperature of grain and granary,initial tasting assessment value and initial fatty acid value of stored grain were chosed,then,a comprehensive analysis of the interaction between these selected storage factors was carried out.Correlation analysis confirmed that there was a strong correlation among storage factors.Then,principal component analysis(PCA)was used to reduce and compress the original storage factors,and the first 6 principal components were extracted from 14 storage factors as the new independent variables of the model,which provided a basis for the parameter selection of the model for predicting grain storage quality.(3)Stored grain quality prediction model based on multi-kernel learning.The prediction accuracy of single kernel learning model depends largely on the choice of kernel function and its parameters,and the choice and construction of the kernel function has not yet a unified theoretical basis.It is often difficult to get the ideal fitting accuracy of the model built with single kernel function.Therefore,in this paper,multi-kernel support vector regression(MKSVR)was constructed on the basis of single-kernel support vector regression(SKSVR).And simple MKL algorithm was used to optimize the parameters of the MKSVR model,then,the MKSVR model was used to predict paddy storage quality.Based on the data set of paddy storage quality in Northeast China,the PCA-MKSVR model was established and compared with the PCA-SKSVR model with single radial basis kernel function,the PCA-MLR model with linear regression,and the MKSVR model,the SKSVR model and the MLR model without PCA.The experimental results showed that the multi-kernel learning model was superior to the single-kernel learning model in predicting accuracy and goodness of fit.Compared with similar models,namely,PCA MKSVR model and MKSVR model,PCA SKSVR model and SKSVR model,MLR model and PCA-MLR model,the prediction accuracy and goodness of fit of the model with PCA after dimension-reducing storage factor treatment were higher than that of the prediction model directly established with original storage factor.Therefore,the MKSVR model can be applied to predict the paddy storage character.(4)Stored grain quality prediction model based on multi-task and multi-kernel learning.Conventional single-task learning methods need to train different prediction models separately,and often ignore the potential relationship between multiple models,which limits the generalization performance of the models.Multitasking learning(MTL)is to put multiple tasks together and learn at the same time to fully explore the correlation between different tasks and realize the sharing of information among multiple models or tasks.In view of the two key indicators of paddy storage character,the paper proposed the multi-task learning method for predicting the quality of stored grain based on single-task learning,established the multi-task and multi-kernel learning model(MTMKL),optimized the parameters by an alternate optimization algorithm based on mirror descent algorithm.Then,the MTMKL model was used to predict the paddy storage character.The experimental results showed that the prediction correlation coefficient of MTMKL model for the fatty acid value of paddy reached 0.885,and the prediction correlation coefficient of MTMKL model for the tasting assessment value reached 0.933.Compared with the MKSVR model,the results of MAE,RMSE and MAPE of fatty acid value was decreased by 9.48%,6.05%and 9.60%,respectively,and R~2 increased by 0.009.The results of MAE,RMSE and MAPE of tasting assessment value was decreased by 11.66%,12.39%and 11.97%,respectively,and R~2increased by 0.005,which indicated that MTMKL model can effectively improve the prediction accuracy of paddy storage character,and can be used as a new method for the prediction of stored grain qulity.(5)Development and simulation application of stored grain safety early warning system.Through systematic demand analysis and function design,the early warning rules for the quality of stored grains have been formulated.The early warning level can be divided into:LevelⅠ,LevelⅡ,LevelⅢ,and LevelⅣ.On this basis,the design framework of stored grain safety early warning system was proposed.Based on Lab VIEW language and machine learning model,a set of stored grain safety warning and quality prediction software was designed and developed,which realized the prediction of stored grain quality.The actual granary of a grain depot in Yushu City,Jilin Province was taken as an example to verify the system.The result showed that the system worked well,and the prediction error of fatty acid value was within±1.5 mg/100g,and the prediction error of tasting assessment value was within±1 point.Based on big grain condition data and machine learning methods,the system can accurately predict the quality of grain stored and greatly reduce the economic cost in the process of grain quality inspection,and reduce the risk of serious deterioration in the process of grain storage.It provides technical support for the precise control of the stored grain quality and has a guiding significance for the safety management of grain storage. |
