Tobacco worm quantity prediction method and device
Through the multi-level knowledge inheritance graph network model, variable rating and iterative training of tobaccoworm number distribution data is solved, which solves the problem that time series prediction methods are difficult to capture spatial correlation and data sparseness, and improves the accuracy of tobaccoworm distribution prediction.
Patent Information
- Application Number
- CN202410604766.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-15
- Publication Date
- 2024-11-08
AI Technical Summary
Existing timing prediction methods are difficult to capture the spatial correlation and processing data sparseness in the distribution data of tobacco insects, resulting in low accuracy in the prediction of tobacco insect distribution and inability to effectively deal with changes in the number of tobacco insects and migration.
The multi-level knowledge inheritance graph network model is used to hierarchical variables to form input data at the measurement point level, region level and workshop level, and iterative training is carried out through the multi-level knowledge inheritance strategy fusion. The spatial characteristics under different particle sizes are used to predict the number of tobacco insects.
It improves the accuracy of tobacco insect distribution prediction, can capture the tobacco insect distribution in tobacco plants more comprehensively and accurately, overcomes the problem of difficult data sparsity and spatial correlation, and achieves high-accuracy prediction.
Smart Images

Figure CN118917451A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of tobacco pest control, and more specifically, to a method and device for predicting the number of tobacco pests. Background Art
[0002] Cigarettes are a kind of smokable food, and the decisive factor affecting their product quality is the quality of cigarette raw materials. If the quality of raw materials cannot meet the standards, the quality of cigarettes will be greatly reduced. However, in the production and processing of raw materials, in order to ensure the operating performance of the machine, the production workshop needs to maintain a constant temperature and humidity environment all year round. At the same time, a large amount of tobacco shreds and tobacco dust will be produced during the production of cigarettes, providing a suitable environment and food source for the growth and reproduction of tobacco insects. When the cigarette raw materials are attacked by tobacco insects during the cigarette production stage, the quality of cigarette raw materials will have a great impact, which will not only affect the intrinsic quality of cigarette products, but also lead to a decrease in consumer satisfaction and market rate. In addition, when the packaging materials of cigarettes are eaten by insects, the appearance quality of cigarette products will also be affected. Therefore, in order to reduce economic losses, it is very important to master the laws of insects, establish an insect control mechanism, and monitor and effectively control the insects of tobacco insects. However, the current insect monitoring method is mainly based on manual visual identification and counting, which is not only inefficient, but also has serious lag. Faced with the possible migration of the tobacco insect's activity range and the increase in its number in a short period of time, manual identification methods are unable to capture the dynamic changes in the distribution of tobacco insects, making it difficult to take countermeasures in advance to reduce economic losses.
[0003] With the rise of artificial intelligence and big data technology in recent years, data-driven multivariate time series prediction methods, such as Autoregressive Integrated Moving Average (ARIMA), Support Vector Machines Regression (SVR), Deep Neural Network (DNN), etc., have gradually emerged in time series prediction tasks such as traffic flow prediction, sales prediction, and stock trend prediction. Among them, the multivariate time series prediction method based on deep neural network extracts deep features of the input signal through multi-layer nonlinear mapping. It does not require a deep understanding of the model mechanism of the source of time series data, and can also achieve high-accuracy time series prediction. It has been widely used in recent years. Representative methods include Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), etc.
[0004] However, these time series prediction methods face two challenges that are difficult to solve in the task scenario of tobacco insect distribution prediction. First, there is a significant spatial correlation between the distribution data of tobacco insect populations at different measurement points, and traditional time series prediction methods find it difficult to capture and model the spatial correlation contained in the data.
[0005] Secondly, tobacco insects reproduce and migrate slowly, and cigarette factories regularly carry out tobacco insect extermination work. Therefore, in the tobacco insect statistical data with measurement points as variables and days as sample units, the distribution of tobacco insects is very sparse, and traditional time series prediction methods are difficult to achieve a high prediction accuracy. Summary of the invention
[0006] The present application provides a method and device for predicting the number of tobacco insects, which classifies the collected historical data on the number of tobacco insects for variables, so as to characterize the spatial characteristics of the distribution of tobacco insects at different granularities, and based on the classified input data, uses a multi-level knowledge inheritance graph network model to obtain the predicted values of the number of tobacco insects at the measuring point level, the regional level and the workshop level, so as to more specifically deal with the strong spatial characteristics of the distribution data of the number of tobacco insects, and effectively improve the accuracy of the prediction of the distribution of tobacco insects.
[0007] The present application provides a method for predicting the number of tobacco insects, comprising:
[0008] The historical data of tobacco insect population distribution collected within the recent preset time are classified into variable levels to form classified input data, which includes measurement point level input data, regional level input data and workshop level input data;
[0009] The hierarchical input data is input into the multi-level knowledge inheritance graph network model to obtain the multi-level tobacco insect population prediction value within the preset future time. The multi-level tobacco insect population prediction value includes the measurement point level tobacco insect population prediction value, the regional level tobacco insect population prediction value and the workshop level tobacco insect population prediction value.
[0010] Preferably, the multi-level knowledge inheritance graph network model includes a workshop-level graph network model, a workshop+region-level graph network model, and a workshop+region+measurement point-level graph network model.
[0011] Preferably, training the multi-level knowledge inheritance graph network model includes:
[0012] The tobacco insect population distribution data is classified into variables to obtain a workshop-level dummy variable data set, a regional-level dummy variable data set, and a measuring point-level variable data set. The workshop-level dummy variable data set is composed of the tobacco insect population distribution data of all regions in the workshop, the regional-level dummy variable data set is composed of the tobacco insect population distribution data of all measuring points in the region, and the measuring point-level variable data set is composed of the tobacco insect population distribution data of all measuring points.
[0013] The workshop-level dummy variable dataset, the regional-level dummy variable dataset and the measuring point-level variable dataset are used to iteratively train the workshop-level graph network model, the workshop+region-level graph network model and the workshop+region+measuring point-level graph network model.
[0014] Preferably, when iteratively training the graph network model of each level, the adjacency matrix of the graph structure learning module in the graph network model of that level and the quantity prediction module in the graph network model of that level are updated simultaneously, and the product of the adjacency matrix and the graph node embedding matrix of that level is the input data of the quantity prediction module.
[0015] Preferably, in the graph structure learning module of the graph network model at each level, the information of each node of the graph structure is a feature mapped from each variable of each sample data in the data set of the level.
[0016] The present application also provides a tobacco insect population prediction device, comprising a first classification module and a prediction module;
[0017] The first classification module is used to classify variables of tobacco insect population distribution historical data collected within the recent preset time to form classified input data, which includes measurement point level input data, regional level input data and workshop level input data;
[0018] The prediction module is used to input the hierarchical input data into the multi-level knowledge inheritance graph network model to obtain the multi-level tobacco insect population prediction value within the future preset time. The multi-level tobacco insect population prediction value includes the measurement point level tobacco insect population prediction value, the regional level tobacco insect population prediction value and the workshop level tobacco insect population prediction value.
[0019] Preferably, the multi-level knowledge inheritance graph network model includes a workshop-level graph network model, a workshop+region-level graph network model, and a workshop+region+measurement point-level graph network model.
[0020] Preferably, the tobacco insect quantity prediction device further comprises a training module, and the training module comprises a second classification module and an iterative training module;
[0021] The second classification module is used to classify the tobacco insect population distribution data, and obtain a workshop-level virtual variable data set, a regional-level virtual variable data set, and a measuring point-level variable data set. The workshop-level virtual variable data set is composed of the tobacco insect population distribution data of all areas in the workshop, the regional-level virtual variable data set is composed of the tobacco insect population distribution data of all measuring points in the area, and the measuring point-level variable data set is composed of the tobacco insect population distribution data of all measuring points;
[0022] The iterative training module is used to iteratively train the workshop-level graph network model, the workshop+regional-level graph network model and the workshop+regional+measurement-point-level graph network model in turn using the workshop-level dummy variable data set, the region-level dummy variable data set and the measurement point-level variable data set.
[0023] Preferably, when the iterative training module is used to iteratively train the graph network model of each level, the adjacency matrix of the graph structure learning module in the graph network model of that level and the quantity prediction module in the graph network model of that level are updated simultaneously, and the product of the adjacency matrix and the graph node embedding matrix of that level is the input data of the quantity prediction module.
[0024] Preferably, in the graph structure learning module of the graph network model at each level, the information of each node of the graph structure is a feature mapped from each variable of each sample data in the data set of the level.
[0025] Other features and advantages of the present application will become apparent from the following detailed description of exemplary embodiments of the present application with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the application and, together with the description, serve to explain the principles of the application.
[0027] Figure 1 A flowchart of the method for predicting the number of tobacco insects provided for this application;
[0028] Figure 2 A flowchart for training a multi-level knowledge inheritance graph network model provided for this application;
[0029] Figure 3 This is a structural diagram of the tobacco insect population prediction device provided in this application. DETAILED DESCRIPTION
[0030] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present application.
[0031] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present application, its application, or uses.
[0032] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.
[0033] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.
[0034] The present application provides a method and device for predicting the number of tobacco insects, which classifies the collected historical data on the number of tobacco insects for variables, and is used to characterize the spatial characteristics of tobacco insect distribution at different granularities. Based on the classified input data, a multi-level knowledge inheritance graph network model is used to obtain the predicted values of the number of tobacco insects at the measuring point level, the regional level, and the workshop level, which can more specifically deal with the strong spatial characteristics of the tobacco insect number distribution data and effectively improve the prediction accuracy of tobacco insect distribution. In addition, in the multi-level knowledge inheritance graph network model of the present application, the graph network models of the workshop level, the workshop + regional level, and the workshop + regional + measuring point level are trained in sequence, and the inter-layer knowledge inheritance strategy is used to integrate the spatial characteristics at different granularities, so as to overcome the sparse distribution of the measuring point level data, and realize the comprehensive and accurate prediction of the distribution of tobacco insects at multiple granularity perspectives in the cigarette factory.
[0035] like Figure 1 As shown, the tobacco insect population prediction method provided in this application includes:
[0036] S110: Variable classification is performed on the historical data on the distribution of tobacco insect populations collected within a recent preset time to form classified input data, where the classified input data includes measurement point-level input data, area-level input data, and workshop-level input data.
[0037] Specifically, each workshop is equipped with multiple tobacco insect quantity detection areas, and each tobacco insect quantity detection area is equipped with multiple measuring points. Therefore, the tobacco insect quantity distribution history data of all measuring points form the measuring point-level input data, and the regional-level variables of the area are formed by accumulating the tobacco insect quantity distribution history data of all measuring points in each area, and the regional-level variables of all areas form the regional-level input data (the construction process is the same as the construction process of the regional-level dummy variable data set described below, please see the detailed description below), and the workshop-level variables of the workshop are formed by accumulating the tobacco insect quantity distribution history data of all areas in the same workshop, and the workshop-level variables of all workshops form the workshop-level input data (the construction process is the same as the construction process of the workshop-level dummy variable data set described below, please see the detailed description below).
[0038] S120: Input the hierarchical input data into the multi-level knowledge inheritance graph network model to obtain the multi-level tobacco insect population prediction value within the preset future time, the multi-level tobacco insect population prediction value includes the measurement point level tobacco insect population prediction value, the regional level tobacco insect population prediction value and the workshop level tobacco insect population prediction value, thereby realizing the tobacco insect population prediction task.
[0039] Specifically, the multi-level knowledge inheritance graph network model includes a workshop-level graph network model, a workshop+region-level graph network model, and a workshop+region+measurement point-level graph network model.
[0040] like Figure 2 As shown, training the multi-level knowledge inheritance graph network model includes:
[0041] S210: Variable classification is performed on the tobacco insect population distribution data to obtain a workshop-level dummy variable data set, a regional-level dummy variable data set, and a measuring point-level variable data set. The workshop-level dummy variable data set is composed of the tobacco insect population distribution data of all areas in the workshop, the regional-level dummy variable data set is composed of the tobacco insect population distribution data of all measuring points in the area, and the measuring point-level variable data set is composed of the tobacco insect population distribution data of all measuring points.
[0042] As an example, according to the spatial structure of the tobacco factory, there are three workshops in the factory that need to carry out tobacco insect control work, including the rolling workshop, the silk workshop and the power workshop. Each workshop is divided into different areas, and each area contains different measurement points, thus forming a variable hierarchical relationship between workshop-area-measurement point.
[0043] Preferably, in each level of the data set, each data monomer is composed of paired sample data and label data, the sample data is tobacco insect population distribution data within a preset number of time steps, and the label data is tobacco insect population distribution data within the next time step.
[0044] As an example, one time step is one week.
[0045] On the basis of the above, preferably, before the variable classification, the tobacco insect population distribution data is also cleaned. Specifically, it includes:
[0046] P1: The original tobacco insect population distribution data collected at the measuring point level Perform sliding window summation with a window width of k to obtain a sliding window data set of tobacco insect measurement point-level variables Where n is the number of samples, m is a is the number of measurement point-level variables (i.e. the total number of measurement points), is a sample in the tobacco insect population distribution data, which includes the tobacco insect populations of all measurement points collected at the same time. The number of tobacco insects at each measurement point is A measurement point level variable). r=n-k+1, Depend on Sum by variable dimension (i.e. , by summing up all variables at the same position in the .
[0047] As an example, the original tobacco insect population distribution data at the measuring point level is the tobacco insect population distribution data with the measuring point as the variable collected from January 2017 to January 2023. The sampling step is once a day, with a total of 1,800 samples. Each sample is collected at the same time by tobacco insect collection devices distributed in different areas of the three workshops of winding, silk making and power, and obtained by manual counting.
[0048] P2: Sliding window dataset of tobacco insect observation point-level variables Divide the data set, set the input historical time step to l, and slide the window according to the input historical time step to obtain the point-level variable data set Where t = rl, is the input historical tobacco insect distribution sequence, i.e. the sample data mentioned above; It is the tobacco insect population distribution data at the next moment, that is, the label data mentioned above.
[0049] As an embodiment, in the obtained data set, 70% of the samples are divided into a training set, and the remaining 30% of the samples are divided into a test set. train =t×70% is the number of training set samples, t test =t×30% is the number of test set samples. That is, the workshop-level dummy variable data set, the regional-level dummy variable data set, and the measurement point-level variable data set contain data subsets for training and testing, respectively.
[0050] When classifying variables, we target the point-level variable dataset D a For each sample data and label data, extract the variable data of all measuring points in each region in units of regions, and then accumulate the variable data of all measuring points in the same region in the same sample data / label data to obtain the regional-level variable of the region corresponding to the sample data / label data. The regional-level variables of all regions corresponding to the same sample data / label data form the regional-level virtual variable dataset D b A regional variable Sample data for Label data That is, the regional dummy variable dataset in Depend on Accumulate the same area to get m b is the number of regional-level variables (i.e., the total number of regions).
[0051] For the regional dummy variable dataset D bFor each sample data and label data, take the workshop as the unit, extract the virtual variable data of all areas in each workshop, and then accumulate the virtual variable data of all areas in the same workshop in the same sample data / label data to obtain the workshop-level variables of the workshop corresponding to the sample data / label data. The workshop-level variables of all workshops corresponding to the same sample data / label data form the workshop-level virtual variable dataset D c A shop-level variable Sample data for / Tag data That is, the workshop-level dummy variable dataset in Depend on It is obtained by accumulating the variables in the same workshop, and mc is the number of workshop-level variables (i.e. the total number of workshops).
[0052] Therefore, the data are divided into tobacco insect distribution data representing different particle sizes through variable classification, reflecting the spatial characteristics of tobacco insect distribution in the tobacco factory, and using this spatial characteristic to improve the prediction accuracy of the model.
[0053] S220: Iteratively train the workshop-level graph network model, the workshop+region-level graph network model, and the workshop+region+measurement point-level graph network model in sequence using the workshop-level dummy variable dataset, the region-level dummy variable dataset, and the measurement point-level variable dataset.
[0054] Specifically, each level of the graph neural network consists of a graph structure learning module and a quantity prediction module, wherein the dimension of the adjacency matrix of the graph structure in the graph structure learning module is set according to the number of variables in the data sets of different levels. The information of each node of the graph structure is the feature mapped by each variable in each sample data in the data set of this level. In graph network models of different levels, the number of variables in the corresponding data sets is different, and thus the number of nodes is also different.
[0055] The adjacency matrix can represent the correlation between different nodes in the graph network model of the corresponding level. It is regarded as the carrier of graph structure knowledge and is set as a learnable parameter matrix in the model, which can be updated with model training.
[0056] As an embodiment, the quantity prediction module is a fully connected layer for prediction tasks, and the fully connected layer is set as a learnable parameter matrix that can be updated with model training.
[0057] When training the graph network model at each level, the data is input into the graph network model, the corresponding adjacency matrix is obtained through the graph structure learning module, and the prediction data is obtained through the quantity prediction module. The graph network model is trained based on the loss function until the model loss converges. During the iterative training of the graph network model at each level, the adjacency matrix of the level and the quantity prediction module in the graph network model at the level are updated simultaneously. The product of the adjacency matrix and the graph node embedding matrix of the level is the input data of the quantity prediction module.
[0058] Preferably, during the training process, when initializing the graph network model of each level, the adjacency matrix learned from the graph network model of the previous level is used as the graph structure knowledge to be inherited; the nodes shared by the graph network models of the current level and the previous level are regarded as common nodes, and the positions representing the correlation between common nodes in the adjacency matrix of the graph network model of the current level are regarded as areas to be inherited, and the remaining positions are non-inherited areas; the values of the adjacency matrix of the graph network model of the previous level are assigned to the areas to be inherited in the adjacency matrix of the graph network model of the current level, and the values of the non-inherited areas are randomly initialized. Among them, the workshop-level graph network model is the graph network model of the previous level of the workshop + regional level graph network model, and the workshop + regional level graph network model is the graph network model of the previous level of the workshop + regional + measurement point level graph network model, such as Figure 2 Therefore, the inter-layer knowledge inheritance strategy is only enabled when training the workshop+region level graph network model and the workshop+region+measurement point level graph network model.
[0059] Specifically, during training, the following steps are performed in sequence:
[0060] S2201: Train the workshop-level graph network model.
[0061] Q1: Calculate the workshop-level adjacency matrix.
[0062] The workshop-level dummy variable dataset D c The training subset sample data in is input into the workshop-level graph network model. For each sample data input into the workshop-level graph network model Re-partition the network according to the variable dimension to obtain the graph node embedding Calculate the corresponding workshop-level adjacency matrix in the graph structure learning module The calculation method is:
[0063]
[0064] in, is the adjacency matrix S c The element in the pth row and qth column of represents the vector concatenation operation, f c(*) is a learnable fully connected layer mapping function, sigmoid(*) is a nonlinear mapping function used for numerical normalization, Embedding for graph nodes The p-th and q-th vectors in .
[0065] Q2: The calculated workshop-level adjacency matrix and graph node embedding matrix Multiply and input into the quantity prediction module to calculate the predicted data of tobacco insect quantity in the next time step
[0066]
[0067] Among them, g c (*) is a concatenated set of mapping functions of the multi-layer learnable graph network fully connected layers of the workshop-level graph network model.
[0068] Q3: According to the forecast data With label data Calculate the model loss, train the workshop-level graph network model based on the loss function until the model loss converges, and update the graph structure during the training process.
[0069] The loss function uses cross entropy loss, mean square error loss, etc. Here, mean square error loss is selected as the loss function of the model for training:
[0070]
[0071] in‖·‖ F represents the Frobenius norm. The model parameters are continuously updated during the training process, and finally the workshop-level adjacency matrix S is obtained. c .
[0072] S2202: Train the workshop + region level graph network model.
[0073] R1: Calculate the workshop + region level adjacency matrix.
[0074] The workshop-level dummy variable dataset D c and the regional dummy variable dataset D b Join in the variable dimension to obtain the workshop + regional level variable data set m cb =m c +m b For each training subset sample data input to the workshop + region level graph network model Re-partition the network according to the variable dimension to obtain the graph node embedding For the workshop + region level adjacency matrix Note the set U cb This is the inheritance area for the network at this level to inherit knowledge between layers, and the set V is recorded cb To set the zero zone, record the set W cb For non-inherited areas, the definition method is:
[0075] U cb ={(p,q)|p,q∈[1,m c ]} (4)
[0076]
[0077]
[0078] Based on this division, the corresponding adjacency matrix is calculated in the graph structure learning module The calculation method is:
[0079]
[0080] in, is the adjacency matrix S cb The element in the pth row and qth column of represents the vector concatenation operation, f cb (*) is a learnable fully connected layer mapping function, and sigmoid (*) is a nonlinear mapping function used for numerical normalization.
[0081] R2: The calculated workshop + region level adjacency matrix S cb and graph node embedding matrix Multiply and input into the quantity forecast module to calculate the forecast data for the next time step
[0082]
[0083] Among them, g cb (*) is a concatenated set of mapping functions of the multi-layer learnable graph network fully connected layers of the workshop + region-level graph network model.
[0084] R3: Based on the predicted data With label data Calculate the model loss, train the workshop + region level graph network model based on the loss function until the model loss converges, and update the graph structure during the training process.
[0085] Here we choose mean square error loss as the loss function of the model for training:
[0086]
[0087] in‖·‖ F represents the Frobenius norm. The model parameters are continuously updated during the training process, and finally the workshop + region level adjacency matrix S is obtained. cb .
[0088] S2203: Train the workshop+area+measurement point level graph network model.
[0089] S1: Calculate the workshop+region+measurement point level adjacency matrix.
[0090] The workshop + region level variable dataset D cb With the measurement point level variable data set D a Combine in variable dimension to get workshop + region + measurement point level variable data set m cba =m cb +m a For each input workshop + area + measurement point level graph network model training subset sample data Re-partition the network according to the variable dimension to obtain the graph node embedding For the workshop + area + measurement point level adjacency matrix Note the set U cba This is the inheritance area for the network at this level to inherit knowledge between layers, and the set V is recorded cba To set the zero zone, record the set W cba For non-inherited areas, the definition method is:
[0091] U cba ={(p,q)|p,q∈[1,m cb ]} (10)
[0092]
[0093]
[0094] Based on this division, the corresponding adjacency matrix is calculated in the graph structure learning module The calculation method is:
[0095]
[0096] in, is the adjacency matrix S cba The element in the pth row and qth column of represents the vector concatenation operation, f cba (*) is a learnable fully connected layer mapping function, and sigmoid (*) is a nonlinear mapping function used for numerical normalization.
[0097] S2: The calculated workshop + area + measurement point level adjacency matrix S cba Multiply it with the graph node embedding matrix and input it into the quantity prediction module to calculate the prediction data for the next time step
[0098]
[0099] Among them, g cba (*) is a concatenated set of mapping functions of the fully connected layers of a multi-layer learnable graph network.
[0100] S3: Based on the predicted data With label data Calculate the model loss and train the workshop + region + measurement point level graph network model based on the loss function until the model loss converges.
[0101] Here we choose mean square error loss as the loss function of the model for training:
[0102]
[0103] in‖·‖ F Indicates the Frobenius norm. The model parameters are continuously updated during the training process to obtain a trained workshop + area + measurement point level graph network model.
[0104] S2204: Use workshop + area + measurement point level variable data set The model is tested on the test set in .
[0105] For each sample data input in the test set For the prediction data output by the model, first round the integer part and convert the resulting output With label data The values of each variable in are compared. If they are the same, the prediction is considered accurate; otherwise, it is considered an error. After obtaining the prediction data of all the samples to be tested, the prediction accuracy of the model is calculated as follows:
[0106]
[0107] Where N correct Indicates the number of samples with accurate prediction, N total =t test Indicates the total number of samples in the test set.
[0108] In order to more clearly reflect the superiority of the multi-level knowledge inheritance graph network model of this application in the task of predicting the distribution of tobacco insect population. Linear regression, convolutional neural network CNN, and long short-term memory network LSTM were selected as comparative methods for comparison of prediction performance, and the results are shown in Table 1. For the prediction task of each method in Table 1, the results of 5 parallel experiments were averaged. It can be found that the method proposed in this application is significantly better than other methods in prediction accuracy, which can basically reach about 90%. This proves the feasibility and superiority of this application.
[0109] Table 1 Comparison of prediction accuracy of this application and other time series prediction methods in tobacco insect distribution prediction tasks
[0110]
[0111] Based on the above, the present application also provides a tobacco insect population prediction device. Figure 3 As shown, the tobacco insect quantity prediction device includes a first classification module 310 and a prediction module 320 .
[0112] The first classification module 310 is used to classify variables of tobacco insect population distribution historical data collected within a recent preset time to form classified input data, which includes measurement point level input data, area level input data and workshop level input data.
[0113] The prediction module 320 is used to input the hierarchical input data into the multi-level knowledge inheritance graph network model to obtain the multi-level tobacco insect population prediction value within a preset time in the future. The multi-level tobacco insect population prediction value includes the measurement point level tobacco insect population prediction value, the regional level tobacco insect population prediction value and the workshop level tobacco insect population prediction value.
[0114] Preferably, the multi-level knowledge inheritance graph network model includes a workshop-level graph network model, a workshop+region-level graph network model, and a workshop+region+measurement point-level graph network model.
[0115] Preferably, the tobacco insect population prediction device further includes a training module 330 , and the training module 330 includes a second classification module 3301 and an iterative training module 3302 .
[0116] The second grading module 3301 is used to perform variable grading on the tobacco insect population distribution data to obtain a workshop-level virtual variable data set, a regional-level virtual variable data set and a measuring point-level variable data set. The workshop-level virtual variable data set is composed of the tobacco insect population distribution data of all areas in the workshop, the regional-level virtual variable data set is composed of the tobacco insect population distribution data of all measuring points in the area, and the measuring point-level variable data set is composed of the tobacco insect population distribution data of all measuring points.
[0117] The iterative training module 3302 is used to iteratively train the workshop-level graph network model, the workshop+region-level graph network model and the workshop+region+measurement point-level graph network model in sequence using the workshop-level virtual variable data set, the region-level virtual variable data set and the measurement point-level variable data set.
[0118] Preferably, when the iterative training module 3302 is used to iteratively train the graph network model of each level, the adjacency matrix of the graph structure learning module in the graph network model of that level and the quantity prediction module in the graph network model of that level are updated simultaneously, and the product of the adjacency matrix and the graph node embedding matrix of that level is the input data of the quantity prediction module.
[0119] Preferably, in the graph structure learning module of the graph network model at each level, the information of each node of the graph structure is a feature mapped from each variable of each sample data in the data set of the level.
[0120] This application integrates the original point-level data and divides it into workshop-level and regional-level virtual variable data sets by constructing a virtual variable data set, and designs a multi-level graph network structure and an inter-level knowledge inheritance strategy to train a tobacco insect distribution prediction model. The model can learn the tobacco insect distribution information contained in variables at different levels in order from coarse to fine granularity, while overcoming the prediction inaccuracy problem caused by the sparsity of the point-level data, and can effectively and comprehensively capture the spatial characteristics of the tobacco insect distribution in the entire tobacco factory, thereby improving the prediction accuracy of the model.
[0121] Although some specific embodiments of the present application have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are only for illustration, not for limiting the scope of the present application. It should be understood by those skilled in the art that the above embodiments may be modified without departing from the scope and spirit of the present application. The scope of the present application is defined by the appended claims.
Claims
1. A method for predicting the number of tobacco insects, characterized in that: include: Performing variable classification on historical data on tobacco insect population distribution collected within a recent preset time to form classified input data, wherein the classified input data includes measurement point-level input data, regional-level input data, and workshop-level input data; The hierarchical input data is input into a multi-level knowledge inheritance graph network model to obtain a multi-level tobacco insect population prediction value within a preset future time, wherein the multi-level tobacco insect population prediction value includes a measurement point-level tobacco insect population prediction value, a regional-level tobacco insect population prediction value, and a workshop-level tobacco insect population prediction value.
2. The method for predicting the number of tobacco insects according to claim 1, characterized in that: The multi-level knowledge inheritance graph network model includes a workshop-level graph network model, a workshop+region-level graph network model, and a workshop+region+measurement point-level graph network model.
3. The method for predicting the number of tobacco insects according to claim 2, characterized in that: Training the multi-level knowledge inheritance graph network model includes: The tobacco insect population distribution data is subjected to variable classification to obtain a workshop-level dummy variable data set, a regional-level dummy variable data set, and a measuring point-level variable data set, wherein the workshop-level dummy variable data set is composed of tobacco insect population distribution data of all regions in the workshop, the regional-level dummy variable data set is composed of tobacco insect population distribution data of all measuring points in the region, and the measuring point-level variable data set is composed of tobacco insect population distribution data of all measuring points; The workshop-level virtual variable data set, the region-level virtual variable data set and the measuring point-level variable data set are used to iteratively train the workshop-level graph network model, the workshop+region-level graph network model and the workshop+region+measuring point-level graph network model in sequence.
4. The method for predicting the number of tobacco insects according to claim 3, characterized in that: When iteratively training the graph network model at each level, the adjacency matrix of the graph structure learning module in the graph network model at that level and the quantity prediction module in the graph network model at that level are updated simultaneously, and the product of the adjacency matrix and the graph node embedding matrix at that level is the input data of the quantity prediction module.
5. The method for predicting tobacco insect numbers according to claim 4, characterized in that: In the graph structure learning module of the graph network model at each level, the information of each node in the graph structure is the feature mapped by each variable of each sample data in the data set of that level.
6. A tobacco insect population prediction device, characterized in that: including a first classification module and a prediction module; The first classification module is used to classify variables of tobacco insect population distribution historical data collected within a recent preset time to form classified input data, wherein the classified input data includes measurement point level input data, area level input data and workshop level input data; The prediction module is used to input the hierarchical input data into a multi-level knowledge inheritance graph network model to obtain a multi-level tobacco insect population prediction value within a preset time in the future, wherein the multi-level tobacco insect population prediction value includes a measurement point-level tobacco insect population prediction value, a regional-level tobacco insect population prediction value, and a workshop-level tobacco insect population prediction value.
7. The tobacco insect population prediction device according to claim 6, characterized in that: The multi-level knowledge inheritance graph network model includes a workshop-level graph network model, a workshop+region-level graph network model, and a workshop+region+measurement point-level graph network model.
8. The tobacco insect population prediction device according to claim 7, characterized in that: It also includes a training module, which includes a second classification module and an iterative training module; The second classification module is used to perform variable classification on the tobacco insect population distribution data to obtain a workshop-level virtual variable data set, a regional-level virtual variable data set, and a measuring point-level variable data set, wherein the workshop-level virtual variable data set is composed of tobacco insect population distribution data of all regions in the workshop, the regional-level virtual variable data set is composed of tobacco insect population distribution data of all measuring points in the region, and the measuring point-level variable data set is composed of tobacco insect population distribution data of all measuring points; The iterative training module is used to iteratively train the workshop-level graph network model, the workshop+region-level graph network model and the workshop+region+measurement point-level graph network model in sequence using the workshop-level virtual variable data set, the region-level virtual variable data set and the measurement point-level variable data set.
9. The tobacco insect population prediction device according to claim 8, characterized in that: The iterative training module is used to iteratively train the graph network model of each level, while updating the adjacency matrix of the graph structure learning module in the graph network model of that level and the quantity prediction module in the graph network model of that level. The product of the adjacency matrix and the graph node embedding matrix of that level is the input data of the quantity prediction module.
10. The tobacco insect population prediction device according to claim 9, characterized in that: In the graph structure learning module of the graph network model at each level, the information of each node in the graph structure is the feature mapped by each variable of each sample data in the data set of that level.
Citation Information
Cited By
In-vehicle intelligent mosquito repelling method
CN120910790A
