Dual-track self-supervised caterpillar quantity prediction method and device guided by scatter plot

Abstract

The application discloses a kind of dispersing graph guided double-track self-supervision caterpillar quantity prediction method and device, method includes: receiving current caterpillar distribution historical data;Using the first graph feature vector of trained graph feature extraction model to extract current caterpillar distribution historical data;First graph feature vector is input into trained fine-grained caterpillar quantity prediction model, and the caterpillar distribution quantity prediction value of future one time step is obtained;Wherein, graph feature extraction model is assisted training using coarse-grained self-supervision auxiliary prediction model.The application learns the distribution rule of caterpillar from caterpillar distribution data using graph feature extraction model, and the graph feature of graph feature extraction model is input into fine-grained caterpillar quantity prediction model to predict the caterpillar distribution quantity of future one time step, improve the accuracy of caterpillar distribution prediction.

Description

Technical Field

[0001] This application relates to the technical field, and more specifically, to a scatter plot-guided dual-track self-supervised method and apparatus for predicting the number of tobacco insects. Background Technology

[0002] Cigarettes are a smoky commodity, and the quality of the tobacco raw materials plays a decisive role in the intrinsic quality of the product. However, in the tobacco raw material production and processing stage, the production workshop needs to maintain constant temperature and humidity conditions, and the cigarette production process generates a large amount of tobacco shreds and dust, providing a suitable growth and reproduction environment and an ample food source for tobacco insects. These insects feed on tobacco, and when raw materials are infested with them, their quality is severely affected. This not only impacts the taste and other aspects of the cigarette product's quality but also leads to a decline in consumer satisfaction and market share. Furthermore, insect damage to cigarette packaging materials also affects the appearance of the tobacco product. Therefore, in order to reduce economic losses, understanding the patterns of insect infestation and establishing a management and control mechanism for insect infestation are extremely important for the monitoring and effective control of tobacco insects.

[0003] Currently, tobacco beetle monitoring mainly relies on manual visual identification and counting, which is inefficient and suffers from significant time lag. Faced with potential changes in the range of tobacco beetle activity and rapid increases in numbers, manual identification methods cannot capture the dispersal patterns and dynamic changes in their distribution, making it difficult to implement preventative measures to reduce economic losses.

[0004] In recent years, with the rapid development of big data technology, data-driven multivariate time series forecasting methods have demonstrated powerful predictive capabilities across various fields. In time series forecasting tasks such as traffic flow forecasting, sales forecasting, and stock trend forecasting, methods such as Moving Average (MA), Autoregressive Integrated Moving Average (ARIMA), and Artificial Neural Network (ANN) have achieved significant results. In recent years, multivariate time series forecasting methods based on deep neural networks have been widely used. These methods extract deep features of the input signal through multi-layer nonlinear mapping, without requiring an in-depth understanding of the underlying model mechanisms of the time series data. By mining the potential characteristics of the data, they can achieve high-accuracy time series forecasting tasks. Network structures such as Recurrent Neural Networks (RNN) and Long Short-Term Memory Networks (LSTM) have been widely applied in this field.

[0005] However, these time-series prediction methods face difficulties in predicting the distribution of tobacco insects in actual cigarette factories. This is because tobacco insects may migrate and reproduce between regions within the factory, which is related to factors such as cigarette production processes and distances between regions. The amount of information in the hidden features learned by the model is insufficient to characterize the distribution patterns of tobacco insects in fine-grained data such as regional variables and daily sampling frequencies. Consequently, the model cannot accurately predict the distribution of tobacco insects. Therefore, improving the model's ability to capture the distribution patterns of tobacco insects and obtaining model features that better characterize the fine-grained distribution characteristics of tobacco insects is a very meaningful and urgent problem to be solved for insect infestation monitoring and control in cigarette factories. Summary of the Invention

[0006] This application provides a scatter graph-guided dual-track self-supervised tobacco insect population prediction method and apparatus. It utilizes a graph feature extraction model to learn the distribution patterns of tobacco insects from tobacco insect distribution data, obtains graph features that better characterize the distribution characteristics of tobacco insects, and inputs the graph features extracted by the graph feature extraction model into a fine-grained tobacco insect population prediction model to predict the distribution number of tobacco insects in the next time step, thereby improving the accuracy of tobacco insect distribution prediction.

[0007] This application provides a scatter plot-guided, dual-track self-supervised method for predicting the number of tobacco insects, including:

[0008] Receive historical data on the current distribution of tobacco beetles;

[0009] The first graph feature vector of the current tobacco insect distribution historical data is extracted using a trained graph feature extraction model.

[0010] Input the feature vector of the first image into the trained fine-grained tobacco beetle quantity prediction model to obtain the predicted distribution quantity of tobacco beetles at the next time step.

[0011] Among them, a coarse-grained self-supervised auxiliary prediction model is used to assist in training the graph feature extraction model.

[0012] Preferably, a coarse-grained self-supervised prediction model is used to assist in training the image feature extraction model, specifically including:

[0013] Data cleaning was performed on the original tobacco insect distribution dataset to obtain a complete tobacco insect distribution dataset. Each data point in the complete tobacco insect distribution dataset includes a sample unit and its corresponding label unit.

[0014] The original graph feature extraction model is used to extract the second graph feature vector from each individual sample in the complete tobacco insect distribution dataset;

[0015] Input all feature vectors of the second graph into the coarse-grained self-supervised auxiliary prediction model, and backpropagate the prediction results of the coarse-grained self-supervised auxiliary prediction model to optimize the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model, thereby obtaining the trained graph feature extraction model.

[0016] Preferably, training the fine-grained tobacco beetle population prediction model specifically includes:

[0017] The trained graph feature extraction model is used to extract the third graph feature vector from each individual sample in the complete tobacco insect distribution dataset.

[0018] The fine-grained tobacco insect population prediction model was trained using all the feature vectors of the third graph to obtain the trained fine-grained tobacco insect population prediction model.

[0019] Preferably, the graph feature extraction model extracts graph features based on a comprehensive adjacency matrix guided by a scatter graph. The comprehensive adjacency matrix is ​​determined by the exogenous tobacco insect scatter adjacency matrix and the endogenous tobacco insect scatter adjacency matrix, which are determined by the cigarette production process flowchart and the geographical location relationship graph between each region.

[0020] Preferably, the original tobacco worm distribution dataset is cleaned to obtain a complete tobacco worm distribution dataset, specifically including:

[0021] The original tobacco insect distribution dataset is divided into multiple sample units and corresponding label units by a sliding window according to the product of the preset time step and the preset number of time steps. Each pair of corresponding sample units and label units forms a data point.

[0022] Combine all the data to form a complete dataset of tobacco insect distribution.

[0023] Preferably, the coarse-grained self-supervised prediction model includes independent spatial and temporal channels. When using the coarse-grained self-supervised prediction model to assist in training the graph feature extraction model, the spatial and temporal channels are trained simultaneously.

[0024] This application provides a scatter plot-guided dual-track self-supervised tobacco insect population prediction device, including a data receiving module, a first feature extraction module, a prediction module, and a training module;

[0025] The data receiving module is used to receive historical data on the current distribution of tobacco insects;

[0026] The first feature extraction module is used to extract the first graph feature vector of the current tobacco insect distribution historical data using a trained graph feature extraction model;

[0027] The prediction module is used to input the feature vector of the first image into the trained fine-grained tobacco insect quantity prediction model to obtain the predicted value of the tobacco insect distribution quantity at the next time step.

[0028] The training module includes a first model training module, which is used to train a graph feature extraction model by using a coarse-grained self-supervised auxiliary prediction model.

[0029] Preferably, the first model training module includes a data cleaning module, a second feature extraction module, and a coarse-grained training module;

[0030] The data cleaning module is used to clean the original tobacco insect distribution dataset to obtain a complete tobacco insect distribution dataset. Each data point in the complete tobacco insect distribution dataset includes a sample unit and its corresponding label unit.

[0031] The second feature extraction module is used to extract the second graph feature vector from each individual sample in the complete tobacco insect distribution dataset using the original graph feature extraction model;

[0032] The coarse-grained training module is used to input all the feature vectors of the second graph into the coarse-grained self-supervised auxiliary prediction model. The prediction results of the coarse-grained self-supervised auxiliary prediction model are backpropagated to optimize the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model, so as to obtain the trained graph feature extraction model.

[0033] Preferably, the training module further includes a second model training module, which includes a third feature extraction module and a fine-grained training module.

[0034] The third feature extraction module is used to extract the third graph feature vector from each individual sample in the complete tobacco insect distribution dataset using the trained graph feature extraction model.

[0035] The fine-grained training module is used to train the fine-grained tobacco insect population prediction model using all the feature vectors of the third graph, and obtain the trained fine-grained tobacco insect population prediction model.

[0036] Preferably, the graph feature extraction model extracts graph features based on a comprehensive adjacency matrix guided by a scatter graph. The comprehensive adjacency matrix is ​​determined by the exogenous tobacco insect scatter adjacency matrix and the endogenous tobacco insect scatter adjacency matrix, which are determined by the cigarette production process flowchart and the geographical location relationship graph between each region.

[0037] Preferably, the data cleaning module includes a partitioning module and a combination module;

[0038] The partitioning module is used to perform sliding window partitioning on the original tobacco insect distribution dataset according to the product of the preset time step and the preset number of time steps, to obtain multiple sample units and corresponding label units. Each pair of corresponding sample units and label units forms a data entry.

[0039] The combination module is used to combine all the data to form a complete tobacco insect distribution dataset.

[0040] Other features and advantages of this application will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the present application and, together with their description, serve to explain the principles of the present application.

[0042] Figure 1 A flowchart of the scatter plot-guided dual-track self-supervised tobacco insect population prediction method provided in this application;

[0043] Figure 2 A flowchart of the training graph feature extraction model provided in this application;

[0044] Figure 3 A structural diagram of the scatter plot-guided dual-track self-supervised tobacco insect population prediction device provided in this application:

[0045] Figure 4 This is a schematic diagram for collecting data on the distribution of tobacco insects. Detailed Implementation

[0046] 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, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0047] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0048] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0049] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0050] This application provides a scatter graph-guided dual-track self-supervised tobacco insect population prediction method and apparatus. It utilizes a graph feature extraction model to learn the distribution patterns of tobacco insects from tobacco insect distribution data, obtains graph features that better characterize the distribution characteristics of tobacco insects, and inputs the graph features extracted by the graph feature extraction model into a fine-grained tobacco insect population prediction model to predict the distribution number of tobacco insects in the next time step, thereby improving the accuracy of tobacco insect distribution prediction.

[0051] like Figure 1As shown, the scatter plot-guided dual-track self-supervised tobacco worm population prediction method provided in this application includes:

[0052] S110: Receive historical data on the current distribution of tobacco insects.

[0053] The current historical data on tobacco beetle distribution includes the distribution data of tobacco beetles at the current time and several previous time steps.

[0054] As an example, the tobacco insect distribution data is obtained using regional variables and a daily sampling frequency, meaning that sampling is performed once a day to obtain one data point. During sampling, each tobacco insect detection point within a preset area is treated as a variable, meaning the number of variables in each data point is equal to the number of tobacco insect detection points within the preset area.

[0055] As one example, the area includes a cigarette-making workshop, a tobacco processing workshop, and a power workshop. Tobacco beetle distribution data is collected by tobacco beetle collection devices installed at each tobacco beetle detection point, and obtained through device identification and counting methods, such as... Figure 4 As shown.

[0056] It should be noted that the current historical data on tobacco beetle distribution is obtained after data preprocessing measures such as outlier handling and missing value imputation.

[0057] S120: Use the trained graph feature extraction model to extract the first graph feature vector of the current tobacco insect distribution historical data.

[0058] S130: Input the feature vector of the first image into the trained fine-grained tobacco insect quantity prediction model to obtain the predicted distribution quantity of tobacco insects at the next time step.

[0059] It should be noted that, in this application, the training of the model includes the training of the graph feature extraction model and the training of the fine-grained tobacco insect quantity prediction model.

[0060] As an example, a coarse-grained self-supervised prediction model is used to assist in training the graph feature extraction model.

[0061] As an example, such as Figure 2 As shown, a coarse-grained self-supervised prediction model is used to assist in training the image feature extraction model, specifically including:

[0062] S210: Clean the original tobacco insect distribution dataset to obtain a complete tobacco insect distribution dataset. Each data point in the complete tobacco insect distribution dataset includes a sample unit and its corresponding label unit.

[0063] As an example, data cleaning is performed on the original tobacco worm distribution dataset to obtain a complete tobacco worm distribution dataset, specifically including:

[0064] S2101: The original tobacco insect distribution dataset is divided into multiple sample units and corresponding label units by a sliding window according to the product of the preset time step and the preset number of time steps. Each pair of corresponding sample units and label units forms a data point.

[0065] It should be noted that the original tobacco insect distribution dataset is a dataset obtained after data preprocessing measures such as outlier handling and missing value imputation, denoted as X=[x1,x2,x3,...,x n ] T ∈R n×m , where n is the number of data points in the original tobacco insect distribution dataset, and m is the number of variables in each data point.

[0066] When partitioning the original tobacco worm distribution dataset, a preset time step of k is set, and a preset number of time steps in each tobacco worm data sequence is set to b. A sliding window is used to partition the dataset by multiplying the preset time step by the preset number of time steps, resulting in multiple individual samples. (i.e., the historical distribution and quantity sequence of tobacco insects) and the corresponding tag monomers (i.e., the number of tobacco insects in the next time step to be predicted).

[0067] S2102: Combine all data to form a complete tobacco insect distribution dataset, denoted as E = [e1, e2, e3, ..., e l ] T ∈R l×m Where l = n - bk,

[0068] As an example, in the obtained complete tobacco insect distribution dataset, 70% of the samples are divided into a training set and the remaining 30% into a test set. Let l train =l×70% is the number of training set samples, l test =l×30% is the number of test set samples.

[0069] S220: Use the original graph feature extraction model to extract the second graph feature vector from each individual sample in the complete tobacco insect distribution dataset.

[0070] First, before inputting the sample individual into the original graph feature extraction model, a fully connected layer function maps the sample individual to the corresponding graph embedding g. i ∈R (b-1)k×m The calculation method is as follows:

[0071]

[0072] Here, the function f(·) is a learnable nonlinear mapping function.

[0073] As an example, the original graph feature extraction model is a multi-layer learnable graph convolutional network (GCN).

[0074] Preferably, in this application, the graph feature extraction model extracts graph features based on a comprehensive adjacency matrix guided by a scatter graph. The comprehensive adjacency matrix is ​​the endogenous tobacco insect scatter adjacency matrix A determined by the cigarette production process flowchart and the geographical location relationship graph between each region. i And the adjacency matrix A of exogenous tobacco insects e Determine (A) i A e ∈R m×m This helps to capture the endogenous and exogenous distribution patterns of tobacco insects.

[0075] Specifically, in the endogenous tobacco insect dispersal adjacency matrix A i In the adjacency matrix A, the adjacency relationship between regions with upstream and downstream process relationships is set to 1, while the adjacency relationship between regions without process relationships is set to 0. e In the model, the adjacency relationship between variables in regions less than 10 meters apart is set to 1, and the adjacency relationship is set to 0 otherwise. Furthermore, the endogenous tobacco insect distribution adjacency matrix A is... i And the adjacency matrix A of exogenous tobacco insects e Weighted integration yields the scatter-map-guided comprehensive adjacency matrix A. d The calculation method is as follows:

[0076] A d =μA i +ρA e (2)

[0077] Where μ and ρ are weighting coefficient hyperparameters.

[0078] Embed the graph of each sample individual into g i ∈R (b-1)k×m The input image feature extraction model extracts the endogenous and exogenous dispersion features of tobacco insects, obtaining the feature vector h of the second image. i ∈R m×(b-1)k The calculation method is as follows:

[0079]

[0080] S230: Input all the feature vectors of the second graph into the coarse-grained self-supervised auxiliary prediction model, backpropagate the prediction results of the coarse-grained self-supervised auxiliary prediction model, optimize the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model, and obtain the trained graph feature extraction model.

[0081] Preferably, the coarse-grained self-supervised auxiliary prediction model is a dual-track coarse-grained self-supervised auxiliary prediction model. Specifically, the coarse-grained self-supervised auxiliary prediction model includes mutually independent spatial and temporal channels. When using the coarse-grained self-supervised auxiliary prediction model to assist in training the graph feature extraction model, the spatial and temporal channels are trained simultaneously.

[0082] Specifically, for coarse-grained self-supervised auxiliary tasks at the spatial level, each sample unit... Corresponding tag unit By performing segmented accumulation at the variable level, we obtain individual tags with workshop-level variables and daily sampling frequencies. p represents the number of workshops. Specifically, the operation involves assigning individual tags to regional variables and daily sampling frequencies. In this process, data from all tobacco insect detection points within the same workshop are summed to obtain a workshop-level variable and a daily sampling frequency for each tag unit. The feature vector h of the second image i The data is input into the spatial channel of a coarse-grained self-supervised auxiliary prediction model to obtain the predicted number of tobacco insects for the first stage of the sampling frequency at the workshop level and the day level. The calculation method is as follows:

[0083]

[0084] Among them, GNN space (·) represents a multi-layer learnable graph network layer applied to coarse-grained self-supervised auxiliary tasks at the spatial level.

[0085] Predicted number of tobacco insects based on workshop-level variables and daily sampling frequency. Corresponding to workshop-level variables and daily-level sampling frequency of individual tags as supervisory information. Calculate the model loss. Loss functions include cross-entropy loss and mean squared error loss.

[0086] As an example, mean squared error loss is chosen as the first loss function for training the model:

[0087]

[0088] Among them ||·|| F This represents the Frobenius norm.

[0089] For coarse-grained self-supervised auxiliary tasks at the time level, each sample unit... Corresponding regional variables and daily sampling frequency of individual tags By performing segmented accumulation at the sample level, we obtain the label unit with regional variables and weekly sampling frequency. Where k > r, k = 7r. Specifically, the operation involves assigning individual tags to regional variables and daily sampling frequencies. In the process, tobacco insect data are accumulated on a weekly basis to obtain regional-level variables and weekly-level sampling frequency labels. The feature vector h of the second image i The data is input into the time channel of a coarse-grained self-supervised auxiliary prediction model to obtain the predicted number of second tobacco insects for regional variables and weekly sampling frequency. The calculation method is as follows:

[0090]

[0091] Among them, GNN time ((·) represents a multi-layer learnable graph network layer applied to coarse-grained self-supervised auxiliary tasks at the time level.

[0092] The predicted number of tobacco insects based on regional variables and weekly sampling frequency. This corresponds to the regional-level variables and the weekly-level sampling frequency of the individual tags, which serve as supervisory information. To calculate the model loss, we choose the mean squared error loss as the second loss function for training:

[0093]

[0094] Among them ||·|| F This represents the Frobenius norm.

[0095] During backpropagation, the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model are optimized to obtain a trained graph feature extraction model. Specifically, the network parameters of the spatial channel are updated based on the spatial channel loss, the network parameters of the temporal channel are updated based on the temporal channel loss, and the network parameters of the graph feature extraction model are updated by combining the spatial channel loss and the temporal channel loss (e.g., the arithmetic sum or weighted sum of the two) until the loss converges and training is complete.

[0096] It should be noted that the network parameters of the fine-grained tobacco insect quantity prediction model were frozen when training the graph feature extraction model.

[0097] As an example, training a fine-grained tobacco beetle population prediction model specifically includes:

[0098] P1: Use the trained graph feature extraction model to extract the third graph feature vector h from each individual sample in the complete tobacco insect distribution dataset. i '.

[0099] Using all the eigenvectors h of the third graph i'Train the fine-grained tobacco beetle population prediction model to obtain the trained fine-grained tobacco beetle population prediction model. Specifically, the feature vector h in the third figure...' i Input the fine-grained tobacco beetle population prediction model to obtain the third tobacco beetle population prediction value with regional variables and daily sampling frequency. Corresponding to the regional-level variables and daily-level sampling frequency labels used as supervisory information To calculate the model loss, we choose the mean squared error loss as the third loss function for training:

[0100]

[0101] Among them ||·|| F This represents the Frobenius norm.

[0102] The training of the fine-grained tobacco beetle population prediction model serves as a fine-tuning part of the model, and the number of training rounds should be relatively fewer than that of the coarse-grained tobacco beetle population prediction model.

[0103] It should be noted that when training the fine-grained tobacco insect population prediction model, the network parameters of the graph feature extraction model are frozen.

[0104] After training, use the complete tobacco insect distribution dataset E = [e1, e2, e3, ..., e l ] T ∈R l×m The model was tested on the test set, with each regional variable and daily sampling frequency sample in the test set being tested. The corresponding model output is obtained, namely, the predicted number of fourth tobacco insects based on regional variables and daily sampling frequency. Then round the integer part of the output to get the further output. Labels with regional variables and daily sampling frequencies The values ​​of each variable are compared; if all variables are the same, the prediction is considered accurate; otherwise, it is considered an incorrect prediction. After obtaining the prediction results for all individual samples in the test set, the model's prediction accuracy is calculated as follows:

[0105]

[0106] Where L correct L represents the number of samples that are accurately predicted. total =l test This indicates the total number of samples in the test set.

[0107] It should be noted that during the prediction phase, only the fine-grained tobacco insect population prediction model is used to predict the distribution and number of tobacco insects, see S110-S130.

[0108] To more comprehensively and reliably demonstrate the superior performance of the scatter plot-guided dual-track self-supervised tobacco insect population prediction method proposed in this application on tobacco insect population prediction tasks with regional variables and daily sampling frequencies, linear regression, graph convolutional neural network (GCN), and long short-term memory network (LSTM) were selected as comparative methods for prediction performance comparison. The results are shown in Table 1. For each method's prediction task, four parallel experiments were conducted in Table 1, and the results from the four experiments were averaged. It can be seen that the prediction accuracy of this application is significantly better than other methods, which proves the feasibility and superiority of the proposed method.

[0109] Table 1 Accuracy Comparison

[0110]

[0111] Based on the above, this application also provides a scatter plot-guided dual-track self-supervised tobacco insect population prediction device. For example... Figure 3 As shown, the dual-track self-supervised tobacco insect population prediction device includes a data receiving module 310, a first feature extraction module 320, a prediction module 330, and a training module 340.

[0112] The data receiving module 310 is used to receive historical data on the current distribution of tobacco insects.

[0113] The first feature extraction module 320 is used to extract the first graph feature vector of the current tobacco insect distribution historical data using a trained graph feature extraction model.

[0114] The prediction module 330 is used to input the feature vector of the first image into the trained fine-grained tobacco insect quantity prediction model to obtain the predicted value of the tobacco insect distribution quantity at the next time step.

[0115] The training module 340 includes a first model training module 3401, which is used to train a graph feature extraction model by using a coarse-grained self-supervised auxiliary prediction model.

[0116] Preferably, the first model training module 3401 includes a data cleaning module, a second feature extraction module, and a coarse-grained training module.

[0117] The data cleaning module is used to clean the original tobacco insect distribution dataset to obtain a complete tobacco insect distribution dataset. Each data point in the complete tobacco insect distribution dataset includes a sample unit and its corresponding label unit.

[0118] The second feature extraction module is used to extract the second graph feature vector from each individual sample in the complete tobacco insect distribution dataset using the original graph feature extraction model.

[0119] The coarse-grained training module is used to input all the feature vectors of the second graph into the coarse-grained self-supervised auxiliary prediction model. The prediction results of the coarse-grained self-supervised auxiliary prediction model are backpropagated to optimize the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model, so as to obtain the trained graph feature extraction model.

[0120] Preferably, the training module 340 further includes a second model training module 3402, which includes a third feature extraction module and a fine-grained training module.

[0121] The third feature extraction module is used to extract the third graph feature vector from each individual sample in the complete tobacco insect distribution dataset using a trained graph feature extraction model.

[0122] The fine-grained training module is used to train the fine-grained tobacco insect population prediction model using all the feature vectors of the third graph, and obtain the trained fine-grained tobacco insect population prediction model.

[0123] Preferably, the graph feature extraction model extracts graph features based on a comprehensive adjacency matrix guided by a scatter graph. The comprehensive adjacency matrix is ​​determined by the exogenous tobacco insect scatter adjacency matrix and the endogenous tobacco insect scatter adjacency matrix, which are determined by the cigarette production process flowchart and the geographical location relationship graph between each region.

[0124] Preferably, the data cleaning module includes a partitioning module and a combination module.

[0125] The partitioning module is used to perform sliding window partitioning on the original tobacco insect distribution dataset according to the product of a preset time step and a preset number of time steps, to obtain multiple sample individuals and corresponding label individuals. Each pair of corresponding sample individuals and label individuals forms a data entry.

[0126] The combination module is used to combine all the data to form a complete tobacco insect distribution dataset.

[0127] This application addresses the problem of traditional time-series prediction methods struggling to capture fine-grained distribution characteristics and exhibiting poor predictive performance in tobacco insect distribution prediction tasks. It proposes a scatter graph-guided dual-track self-supervised tobacco insect quantity prediction method. Based on considerations of the intrinsic and extrinsic distribution patterns of tobacco insects and the concept of self-supervised auxiliary tasks, this method uses a graph convolutional neural network layer combining intrinsic and extrinsic scatter graphs for feature extraction. Two self-supervised tracks are designed to learn coarse-grained distribution characteristics at both the spatial and temporal levels of tobacco insect data, enabling the model to possess more comprehensive and accurate feature representation capabilities. Furthermore, this application designs a two-stage training strategy, allowing the model to optimize feature representation capabilities on the initial self-supervised auxiliary task while simultaneously fine-tuning on subsequent fine-grained tobacco insect distribution data. Through the proposed strategy, the model can fully exploit the distribution characteristics of tobacco insects, achieving high prediction accuracy in fine-grained tobacco insect distribution data prediction tasks.

[0128] While specific embodiments of this application have been described in detail by way of examples, those skilled in the art should understand that the above examples are for illustrative purposes only and are not intended to limit the scope of this application. Those skilled in the art should understand that modifications can be made to the above embodiments without departing from the scope and spirit of this application. The scope of this application is defined by the appended claims.

Claims

1. A scatter plot-guided dual-track self-supervised method for predicting the number of tobacco insects, characterized in that, The dual-track self-supervision refers to designing two self-supervised tracks to learn the coarse-grained distribution characteristics of tobacco worms at both the spatial and temporal levels, based on the spatial and temporal distribution characteristics of the data. The prediction methods include: Receive historical data on the current distribution of tobacco insects; The first graph feature vector of the current tobacco insect distribution historical data is extracted using a trained graph feature extraction model. Input the feature vector of the first image into the trained fine-grained tobacco beetle quantity prediction model to obtain the predicted distribution quantity of tobacco beetles at the next time step. When using a coarse-grained self-supervised auxiliary prediction model to assist in training a graph feature extraction model, both spatial and temporal channels are trained simultaneously. The coarse-grained self-supervised auxiliary prediction model includes independent spatial and temporal channels. The auxiliary training specifically includes: cleaning the original tobacco worm distribution dataset to obtain a complete tobacco worm distribution dataset, where each data point includes a sample unit and its corresponding label unit; extracting a second graph feature vector from each sample unit in the complete tobacco worm distribution dataset using the original graph feature extraction model; inputting all second graph feature vectors into the coarse-grained self-supervised auxiliary prediction model; backpropagating the prediction results of the coarse-grained self-supervised auxiliary prediction model to optimize both the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model, thus obtaining a trained graph feature extraction model. The graph feature extraction model extracts graph features based on a comprehensive adjacency matrix guided by a scatter graph. The comprehensive adjacency matrix is ​​determined by the exogenous tobacco insect scatter adjacency matrix and the endogenous tobacco insect scatter adjacency matrix, which are determined by the cigarette production process flowchart and the geographical location relationship graph between each region.

2. The scatter plot-guided dual-track self-supervised tobacco insect population prediction method according to claim 1, characterized in that, Training the fine-grained tobacco beetle population prediction model specifically includes: The trained graph feature extraction model is used to extract the third graph feature vector from each individual sample in the complete tobacco insect distribution dataset. The fine-grained tobacco insect population prediction model was trained using all the feature vectors of the third graph to obtain the trained fine-grained tobacco insect population prediction model.

3. The scatter plot-guided dual-track self-supervised tobacco insect population prediction method according to claim 1, characterized in that, The original tobacco worm distribution dataset was cleaned to obtain a complete tobacco worm distribution dataset, specifically including: The original tobacco insect distribution dataset is divided into multiple sample units and corresponding label units by a sliding window according to the product of a preset time step and a preset number of time steps. Each pair of corresponding sample units and label units forms a data entry. All data are combined to form the complete tobacco insect distribution dataset.

4. A scatter plot-guided dual-track self-monitored tobacco insect population prediction device, characterized in that, The dual-track self-supervision refers to designing two self-supervised tracks to learn the coarse-grained distribution characteristics of tobacco worms at the spatial and temporal levels, respectively; the prediction device includes a data receiving module, a first feature extraction module, a prediction module, and a training module. The data receiving module is used to receive historical data on the current distribution of tobacco insects; The first feature extraction module is used to extract the first graph feature vector of the current tobacco insect distribution historical data using a trained graph feature extraction model; The prediction module is used to input the feature vector of the first image into the trained fine-grained tobacco insect quantity prediction model to obtain the predicted value of the tobacco insect distribution quantity at a future time step. The training module includes a first model training module, which is used to train both the spatial and temporal channels simultaneously when using a coarse-grained self-supervised auxiliary prediction model to assist in training the graph feature extraction model. The coarse-grained self-supervised auxiliary prediction model includes independent spatial and temporal channels. The auxiliary training specifically includes: cleaning the original tobacco worm distribution dataset to obtain a complete tobacco worm distribution dataset, where each data point includes a sample unit and its corresponding label unit; extracting a second graph feature vector from each sample unit in the complete tobacco worm distribution dataset using the original graph feature extraction model; inputting all second graph feature vectors into the coarse-grained self-supervised auxiliary prediction model; backpropagating the prediction results of the coarse-grained self-supervised auxiliary prediction model to optimize both the coarse-grained self-supervised auxiliary prediction model and the graph feature extraction model to obtain a trained graph feature extraction model. The graph feature extraction model extracts graph features based on a comprehensive adjacency matrix guided by a scatter graph. The comprehensive adjacency matrix is ​​determined by the exogenous tobacco insect scatter adjacency matrix and the endogenous tobacco insect scatter adjacency matrix, which are determined by the cigarette production process flowchart and the geographical location relationship graph between each region.

5. The scatter plot-guided dual-track self-monitored tobacco insect quantity prediction device according to claim 4, characterized in that, The training module further includes a second model training module, which includes a third feature extraction module and a fine-grained training module. The third feature extraction module is used to extract the third graph feature vector of each individual sample in the complete tobacco insect distribution dataset using the trained graph feature extraction model. The fine-grained training module is used to train the fine-grained tobacco insect quantity prediction model using all the feature vectors of the third graph, so as to obtain the trained fine-grained tobacco insect quantity prediction model.

6. The scatter plot-guided dual-track self-monitored tobacco insect quantity prediction device according to claim 4, characterized in that, The data cleaning module includes a partitioning module and a combination module; The partitioning module is used to perform sliding window partitioning on the original tobacco insect distribution dataset according to the product of a preset time step and a preset number of time steps, to obtain multiple sample units and corresponding label units, and each pair of corresponding sample units and label units forms a data entry. The combination module is used to combine all the data to form the complete tobacco insect distribution dataset.

Images