Analysis And Prediction Of Defects In The Data Set Obtained In Stored Grain Ecology

Grain storage safety is very important.The temperature is one of the main factors affecting the safe storage of grain,and the insect is one of the direct causes of the loss of stored grain.This paper focuses on the application of the defective temperature data and the image data of grain insects which can be obtained in large quantities in the monitoring of stored grain.In this paper,a spatial visualization model and a spatial-temporal estimation prediction model based on the characteristics of the stored grain ecosystem were established based on the temperature data to reduce the estimation error caused by the loss,error and detection error of temperature data,and realize the function of early warning of abnormal temperature regions.And image recognition technology was used to realize the rapid multi-objective classification detection of grain insects and provide scientific basis for the safe storage of grain.The main research contents of this paper are as follows:1.Three-dimensional spatial visualization of missing grain temperature data and regional warning of abnormal temperatureBased on the characteristics of three-dimensional grain stack temperature data,a method of Universal Kriging solution and application was proposed to describe the spatial relationship between grain stack temperature,which solved the problem of accurate estimation when the three-dimensional grain stack temperature data was missing up to 60%.Non-Stationary Variance Universal Kriging model was proposed by considering the different distribution and amplitude of different parts of grain pile affected by external climate.In this model,different variances are used to represent the temperature variation law and variation degree of different parts of grain pile,which not only inherits the advantages of Universal Kriging,but also gives the temperature distribution characteristics under different grain temperatures(typical seasonal temperature distribution,empty storage,new grain,etc.).At the same time,the model gives the estimation and deviation of grain pile temperature value at the detection point by leaving one estimate.By estimating the confidence interval(95%),the potential internal temperature anomaly area can be warned.2.Flexible Universal Kriging state-space model(FUKSS).FUKSS was proposed to solve the problems of random missing,abnormal error and noise of grain temperature time series data at different time in the storage process.In this work,a recursive method similar to Kalman filter is used to estimate the time-series,avoiding the problem of increasing data caused by Kriging space-time extension.Based on the statistical characteristics of Kriging,this method introduces a spatial selection matrix to make the different observation data and state vectors identical at different times,which solves the problem of missing data.This model is suitable for dynamic problems,and compared with other spatiotemporal models,it has the following characteristics:no large amount of prior calculation;low data consistency requirements;small memory requirements(extra storage space is constant as a single moment of complete data O(N~2);relatively fast computing speed(calculation complexity is constant as a single moment of detection data O(N_t~3));can realize the estimation of any time and any location,etc.In addition,a dynamic linear model is introduced to solve the problem that the Universal Kriging state-space model cannot predict.Finally,the validity of the model was verified by using simulated data and actual granary temperature detection data,and the maximum RMSE of grain heap temperature prediction for consecutive weeks was 0.6 when the three-dimensional grain stack temperature data was missing up to 40%.3.A hybrid model containing thermodynamic and FUKSS.Hybrid model was proposed to further solve the loss in the training time series of the FUKSS model,which contained thermodynamic model and FUKSS model.In this model,the simplified heat transfer model is used to calculate the main change trend of grain stack temperature,and the detection data is used as a supplement to the error term of local temperature change.Therefore,the hybrid model complements the advantages of the heat transfer model and FUKSS model,and reduces the influence of missing data in time on the model.Although the hybrid model is more complex and has more parameters,compared with FUKSS model,its calculation speed only adds fixed heat transfer differential equation solution(determined by time and warehouse modeling size).Finally,the validity of the model is verified by the temperature detection data of actual granaries.When the 3d grain pile temperature data was missing up to 40%and the grain pile temperature data was missing for two weeks in the training sequence,the maximum RMSE of the prediction of grain pile temperature for three consecutive weeks was about 0.67℃.4.Multi-target detection of multi-food insects.In this paper,deep learn-based image recognition method was studied for the detection and recognition of 8 common stored grain pests,and improvement was proposed to solve the problems of small size,complex posture and shelter of stored grain pests in the image.The identification of stored grain pests was based on R-FCN algorithm,and a multi-scale training strategy was introduced to improve the relative small background of grain insects.Meanwhile,image enhancement and deep detachable convolution were used to enrich and optimize the feature extraction of the complex attitude of grain insects,and soft-NMS algorithm was introduced to improve the occlusion in the image of grain insects.Finally,the improved model achieved the accuracy of 88.06%and detection speed of 0.118s for 8 common stored grain pests with multiple targets and multiple classification,and verified the effectiveness of the method under white board background and actual background.This paper mainly focuses on the analysis of temperature data and image data of grain insects which can be obtained in the grain storage system.For temperature data,data-driven modeling is mainly used to reduce manual checking and preprocessing of missing,abnormal and wrong data,providing reference for automatic and graphical intelligent analysis of stored grain temperature.At the same time,time loss,spatial random loss and detection system error are the main characteristics for the analysis and prediction of grain stack time series data,which can provide reference for the layout of intelligent sensor network in grain stack for long-term storage and random sleep awakening.Finally,several common problems and corresponding solutions were proposed for target detection of food insect images,which can effectively reduce the collection operation of custodians and the recognition requirements of similar insect species.