A grain warehouse intelligent camera and data management system based on the Internet of Things

Abstract

The present application relates to the technical field of grain storehouse pest control, and particularly relates to a grain storehouse intelligent camera and data management system based on the Internet of Things, which comprises the following steps: combining grain surface spectrum data, grain pile three-dimensional data and grain pile temperature and humidity change data, quantitatively analyzing the horizontal and vertical distribution uniformity of the pesticide on the grain surface, and analyzing the longitudinal penetration degree of the pesticide in the grain pile; thereby analyzing the effect duration of the pesticide in the grain pile; combining the grain pile three-dimensional data and the insect activity trajectory data, analyzing the path interception degree of the pests during the movement on the grain surface; based on the effect duration and the path interception degree, analyzing the pest spread retardation effect; thereby realizing the accurate management of the grain pile in the grain storehouse; and improving the reliability of the grain storehouse pest control.

Description

Technical Field

[0001] This invention relates to the field of grain warehouse pest control technology, and in particular to an intelligent grain warehouse camera and data management system based on the Internet of Things. Background Technology

[0002] As a strategic national resource and the foundation of people's livelihood, the safety of grain storage directly affects the stability of food supply and the health of the people. During long-term storage, grain piles are continuously affected by environmental factors such as temperature and humidity changes and microbial activity, gradually disrupting the internal micro-ecological environment and leading to problems like localized heating and mold. Furthermore, the reproduction and activity of pests in grain piles can progress from initial sporadic distribution to regional spread, and finally to large-scale outbreaks, causing grain quality deterioration, quantity loss, and even the production of fungal toxins. Failure to promptly detect pest activity patterns and assess the effectiveness of control measures can trigger large-scale pest outbreaks, causing not only huge economic losses but also jeopardizing food safety. Traditional grain pest monitoring relies heavily on manual sampling, which is inefficient and struggles to cover deep layers of grain piles and complex areas such as surface ditches, making it difficult to detect pests in their early stages. By using smart cameras and IoT technology for monitoring, the continuous data collection and real-time transmission capabilities can easily cover all areas of the grain warehouse, including grain surface ditches, the inside of grain piles and other areas that are difficult for people to reach, thus overcoming the limitations of manual inspection and greatly improving monitoring efficiency and data density.

[0003] However, current technologies for intelligent management of grain warehouses fail to correlate the uniformity of pesticide distribution after application, the vertical penetration effect of pesticides in grain piles, and the movement paths of pests in the grain surface ditches. For example, after spraying pesticides on the grain surface, without considering the impact of the depth and density of the grain surface ditches on pesticide distribution and the obstruction effect of the ditches' terrain on pest movement, it is difficult to accurately determine the actual control effect of the pesticides and the degree of pest spread inhibition, leading to misjudgments of pest risk. Furthermore, current technologies lack a comprehensive consideration of the duration of pesticide effects and the degree of pest movement interception, making it difficult for monitoring results to accurately reflect the pest spread inhibition effect after pesticide application. This often results in overuse of pesticides or delayed control measures, increasing storage costs and reducing the reliability of pest control.

[0004] To address these issues, this application presents an IoT-based intelligent camera and data management system for grain warehouses. Summary of the Invention

[0005] The purpose of this invention is to provide an IoT-based intelligent camera and data management system for grain warehouses. By quantifying the uniformity and penetration of pesticide distribution, the system assesses the duration of the effect. It also analyzes the degree to which pest movement is hindered by terrain and integrates the duration of the effect and the degree to which pest movement is hindered by terrain to obtain the pest spread inhibition effect. This enables precise assessment and intelligent decision-making for pest control in grain warehouses.

[0006] This invention is implemented as follows: In a first aspect, the present invention provides a grain warehouse data management system based on the Internet of Things, including a multi-source acquisition module, a pesticide effect monitoring module, an insect path monitoring module, an insect pest blocking analysis module, and a grain pile management module. Among them, the multi-source acquisition module is used to collect spectral data of the grain surface, three-dimensional data of the grain pile, and insect activity trajectory data of the grain pile after pest control operations by using intelligent multispectral cameras deployed in the grain warehouse; and simultaneously collect data on temperature and humidity changes in the grain pile. The pesticide effect monitoring module is used to combine grain surface spectral data, grain pile three-dimensional data, and grain pile temperature and humidity change data to quantitatively analyze the uniformity of pesticide distribution on the grain surface and the degree of pesticide penetration in the grain pile; based on the uniformity of horizontal and vertical distribution and the degree of vertical penetration, the module comprehensively analyzes the duration of the pesticide effect in the grain pile. The insect path monitoring module is used to analyze the degree of path interception of pests during their movement on the grain surface by combining three-dimensional data of the grain pile and insect activity trajectory data. The pest inhibition analysis module is used to analyze the pest spread inhibition effect of the grain pile after pest control operations based on the duration of the pesticide's effect in the grain pile and the degree of path interception of pests during their movement on the grain surface. The grain pile management module is used to manage grain piles in grain warehouses based on the analysis results of the pest spread inhibition effect after pest control operations.

[0007] In one implementation of the present invention, the uniformity of the horizontal and vertical distribution of the agent on the grain surface is quantitatively analyzed by combining grain surface spectral data, three-dimensional data of the grain pile, and temperature and humidity change data of the grain pile, and the vertical penetration degree of the agent in the grain pile is analyzed, including the following specific steps: S21. Acquire spectral data of grain surface and three-dimensional data of grain pile: After the pest control operation is completed, the spectral data of grain surface at multiple sampling points on the grain pile is collected by intelligent multispectral cameras deployed in the grain warehouse. This includes the multispectral reflectance curve of each sampling point. The multispectral reflectance curve of each sampling point is standardized. Based on the partial least squares regression model established in advance through laboratory calibration, the standardized multispectral reflectance curves of each sampling point are converted into the pesticide concentration index of each sampling point to form an initial pesticide concentration distribution map of the grain surface. S22. Collect three-dimensional data of the grain pile surface, including three-dimensional point cloud data of the grain pile surface, grid the three-dimensional point cloud data, calculate the ditch depth and ditch density of each grid unit, and construct a topographic feature map of the grain surface. S23. Spatially align the grain surface topographic feature map with the initial grain surface pesticide concentration distribution map, and use the topographic features to perform topographic correction on the pesticide concentration to eliminate the pesticide pooling effect caused by ditches, thereby obtaining a topographically corrected pesticide concentration distribution map; the topographic correction is achieved by dividing the pesticide concentration of each grid cell by the topographic influence coefficient of the corresponding grid cell. S24. Using the terrain-corrected drug concentration distribution map as training samples, construct a convolutional autoencoder network; the convolutional autoencoder network consists of an encoder and a decoder. The encoder is used to compress the input image into a low-dimensional feature vector, and the decoder is used to reconstruct the low-dimensional feature vector into an output image; the training objective is to minimize the root mean square error between the input image and the reconstructed image, so that the network learns the normal pattern of drug concentration distribution. S25. Obtain pesticide concentration distribution maps collected after multiple pesticide application operations on the grain pile and corrected for terrain, and label them according to the actual uniformity of pesticide distribution to obtain labeled historical pesticide distribution data. The uniformity of actual distribution is obtained by measuring pesticide residue through gridded sampling and calculating the coefficient of variation. Divide the labeled historical pesticide distribution data into training set and validation set. Use the training set to train a convolutional autoencoder and use the validation set to adjust the network hyperparameters until the reconstruction error stabilizes and converges, thus obtaining a trained model for analyzing the uniformity of the horizontal and vertical distribution. S26. Input the pesticide concentration distribution map obtained after the current monitoring grain pile has been treated with pesticides and corrected for terrain into the trained horizontal and vertical distribution uniformity analysis model, and calculate the reconstruction error corresponding to the current monitoring grain pile after pesticide treatment; divide the reconstruction error corresponding to the current monitoring grain pile after pesticide treatment by the maximum reconstruction error corresponding to multiple pesticide treatment operations in the grain warehouse in history to obtain the horizontal and vertical distribution uniformity of pesticides on the grain surface after pesticide treatment in the current monitoring grain pile.

[0008] In one implementation of the present invention, the uniformity of the horizontal and vertical distribution of the agent on the grain surface is quantitatively analyzed by combining grain surface spectral data, grain pile three-dimensional data, and grain pile temperature and humidity change data, and the vertical penetration degree of the agent in the grain pile is analyzed. Specifically, the method further includes the following steps: S27. Collect three-dimensional point cloud data of the grain pile surface, calculate the ditch depth and ditch density at each sampling point, and form a grain pile feature vector; embed temperature and humidity sensors at equal depths along the vertical direction in the grain pile, and continuously collect temperature and humidity change data of the grain pile at each depth after the pest control operation is completed, including the temperature and humidity change rate at each depth within a fixed monitoring period after pesticide application; align the ditch feature vector with the temperature and humidity change rate at each depth according to spatial location to form a spatiotemporal sequence data containing surface topographic features; S28. Based on the spatiotemporal sequence data containing surface topographic features, a spatiotemporal graph convolutional network integrating topographic features is constructed. The spatiotemporal graph convolutional network is used to simultaneously extract the change patterns in the time dimension and the transmission relationship in the spatial dimension, and incorporates the initial influence of surface topography on the longitudinal penetration of the agent. The spatiotemporal graph convolutional network outputs the agent penetration intensity at each depth position. S29. Obtain the actual pesticide penetration intensity at each depth location, determined by stratified sampling and actual measurement of pesticide residue concentration during multiple pesticide application operations on the grain pile. Combine the actual pesticide penetration intensity at each depth location with the corresponding grain pile feature vector and the temperature and humidity change rate at each depth location to form a longitudinal penetration analysis dataset. Divide the longitudinal penetration analysis dataset into a training set and a validation set. Use the training set to train a spatiotemporal graph convolutional network that integrates terrain features, so that the spatiotemporal graph convolutional network that integrates terrain features can predict the pesticide penetration intensity at each depth location based on the grain pile feature vector and the temperature and humidity change rate at each depth location, thus obtaining a pesticide penetration intensity analysis model. S210. Set weight coefficients for different depth positions on the trained agent penetration intensity analysis model, wherein the weight of shallow depth positions is lower than that of deep depth positions; calculate the weighted average of the agent penetration intensity at each depth position output by the agent penetration intensity analysis model according to the weight coefficient of the corresponding depth position to obtain the longitudinal penetration degree of the agent in the grain pile. S211. Input the feature vector of the grain pile after pest control and the temperature and humidity change rate at each depth within a fixed monitoring period after pest control into the trained pesticide penetration intensity analysis model. Output the pesticide penetration intensity at each depth in the current pest control operation. Sum the products of the pesticide penetration intensity at all depths and the corresponding weight coefficients at the depths to obtain the longitudinal penetration degree of the pesticide in the grain pile after pest control.

[0009] In one implementation of the present invention, the duration of the agent's effect in the grain pile is comprehensively analyzed based on the uniformity of horizontal and vertical distribution and the degree of vertical penetration, including the following specific steps: S212. Extract the uniformity of the horizontal and vertical distribution of pesticide on the grain surface and the degree of vertical penetration of pesticide in the grain pile after the current monitoring grain pile has been treated. S213. The uniformity of the horizontal and vertical distribution of the pesticide on the grain surface and the degree of vertical penetration of the pesticide in the grain pile after the pesticide application are weighted and summed to obtain the duration of the pesticide effect in the grain pile after the pesticide application.

[0010] In one implementation of the present invention, the degree of path interception of pests during their movement on the grain surface is analyzed by combining three-dimensional data of the grain pile and insect activity trajectory data, including the following specific steps: S31. Obtain the three-dimensional point cloud data of the grain pile surface from the three-dimensional data of the grain pile, and project the three-dimensional point cloud data of the grain pile surface onto a two-dimensional plane to generate a ditch distribution grid map; wherein, each grid cell in the ditch distribution grid map stores the ditch depth value and slope value at that location. S32. From the insect activity trajectory data, obtain the temporal sequence of individual insect locations detected by continuous frame images, and treat each location sequence as an insect movement trajectory; for each insect movement trajectory, calculate the detour distance ratio, ditch crossing frequency, and velocity variation variance; wherein, the detour distance ratio is the ratio of the actual movement path length of the corresponding trajectory to the straight-line distance from the start point to the end point of the corresponding trajectory; the ditch crossing frequency is the number of grid cells in the ditch distribution grid map traversed by the corresponding trajectory; the velocity variation variance is the degree of fluctuation in the movement speed of the corresponding trajectory in the dense ditch area; S33. Obtain pest movement trajectory data from multiple historical experiments in the grain pile. The pest movement trajectory data includes the detour distance ratio, ditch crossing frequency, speed variation variance, and obstruction degree label for each trajectory, calibrated experimentally. The obstruction degree label is obtained by setting experimental scenarios with different obstacle densities, observing the proportion of pest passage time, and normalizing the results. The pest movement trajectory data is used to construct a path interception analysis dataset, which is then divided into a path interception analysis training set and a path interception analysis validation set.

[0011] In one implementation of the present invention, the degree of path interception of pests during their movement on the grain surface is analyzed by combining three-dimensional data of the grain pile and insect activity trajectory data, and the following specific steps are also included: S34. Construct a graph neural network model. The graph neural network model uses each position point on the pest's movement trajectory as a node, and the connection between adjacent position points as an edge. Each node is appended with the ditch depth and slope value of the corresponding grid cell as node features. The graph neural network model is trained using the path interception analysis training set. The node features of each pest movement trajectory are used as input, and the obstruction degree label corresponding to the trajectory is used as the output target to obtain an initial path interception degree prediction model. The initial path interception degree prediction model is verified using the path interception analysis validation set. The model with a prediction error lower than a preset error threshold is selected as the path interception degree prediction model. S35. Input the movement trajectory of each individual pest in the currently monitored grain pile into the path interception degree prediction model, and output the obstruction probability value of each pest movement trajectory. S36. Obtain the actual path length of each pest movement trajectory within the currently monitored grain pile; sum the actual path lengths of all trajectories to obtain the total path length; use the proportion of the actual path length of each trajectory to the total path length as the weight coefficient of the corresponding trajectory, and sum the products of the obstruction probability values ​​of all pest movement trajectories and the weight coefficients of the corresponding trajectories to obtain the degree of path interception of pests during their movement on the grain surface of the currently monitored grain pile.

[0012] In one implementation of the present invention, the effect of pest control on the spread of pests in the grain pile after the pest control operation is analyzed based on the duration of the pesticide's effect in the grain pile and the degree of path interception during the movement of pests on the grain surface. The specific steps include the following: S41. Extract the duration of the pesticide's effect on the grain pile after insecticide application and the degree of path interception of pests during their movement on the grain surface of the current monitored grain pile. S42. The effect of pesticide on the grain pile after pest control and the degree of path interception of pests during their movement on the grain surface of the current monitored grain pile are weighted and summed to obtain the pest spread inhibition effect of the current monitored grain pile after pest control.

[0013] In one implementation of the present invention, the grain piles in the grain warehouse are managed based on the analysis results of the pest spread inhibition effect of the grain piles after pest control operations, including the following specific contents: S51. Obtain the pest spread inhibition effect of the monitored grain pile after the pest control operation. S52. Preset the threshold for the inhibition effect. When the inhibition effect of pest spread on the currently monitored grain pile after the pest control operation is greater than the threshold, normal monitoring of the currently monitored grain pile is carried out to maintain the current grain pile management status. When the inhibition effect of pest spread on the currently monitored grain pile after the pest control operation is less than or equal to the threshold, the grain discharge operation of the currently monitored grain pile is suspended and the full fumigation and pest control program is started.

[0014] Secondly, the present invention provides an intelligent grain warehouse camera based on the Internet of Things, including the following specific contents: In response to the collection command after the pest control operation, the spectral data of the grain surface on the grain pile inside the grain warehouse is collected, and the three-dimensional data of the grain pile on the grain pile surface is collected simultaneously. Based on multi-frame spectral data of grain surface in continuous time series, the activity area of ​​insects on the surface of grain pile is identified, and the activity trajectory data of insects is extracted based on the positional changes of insects in different time series frames. Simultaneously collect data on temperature and humidity changes inside the grain pile; The spectral data of the grain surface, the three-dimensional data of the grain pile, the insect activity trajectory data, and the temperature and humidity change data of the grain pile are packaged and uploaded to the aforementioned Internet of Things-based grain warehouse data management system.

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. This invention reduces the risk of control failure due to uneven distribution or insufficient penetration of pesticides by quantitatively analyzing the uniformity of the horizontal and vertical distribution of pesticides on the grain surface and the degree of vertical penetration in the grain pile, thereby improving the reliability of pest control in grain warehouses. 2. This invention analyzes the degree of path interception of pests during their movement on the grain surface, accurately identifies the obstructive effect of the terrain of the grain surface ditches on pest activities, thereby optimizing the synergistic effect of physical and chemical control and reducing the risk of pest spread. 3. This invention comprehensively evaluates the effect of pest control on the spread of pests in grain piles after pest control operations, thereby managing grain piles, reducing grain loss and quality deterioration caused by uncontrolled pests, and ensuring the long-term safety and quality stability of stored grain. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the workflow of a grain warehouse data management system based on the Internet of Things according to the present invention. Figure 2 This is a schematic diagram of the structure of a grain warehouse data management system based on the Internet of Things according to the present invention; Figure 3This is a flowchart illustrating the process of the pesticide effect monitoring module in a grain warehouse data management system based on the Internet of Things according to the present invention. Figure 4 This is a flowchart illustrating the insect path monitoring module in a grain warehouse data management system based on the Internet of Things (IoT) of the present invention. Figure 5 This is a flowchart illustrating the process of the pest inhibition analysis module in a grain warehouse data management system based on the Internet of Things, according to the present invention. Figure 6 This is a flowchart illustrating the operation of an IoT-based smart camera for grain warehouses according to the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0018] Example 1 like Figure 2 As shown, this embodiment provides a grain warehouse data management system based on the Internet of Things, including a multi-source acquisition module, a pesticide effect monitoring module, an insect path monitoring module, an insect pest blockage analysis module, and a grain pile management module; Among them, such as Figure 1As shown, the multi-source acquisition module is used to collect spectral data of the grain surface, 3D data of the grain pile, and insect activity trajectory data of the grain pile after pest control operations, using intelligent multispectral cameras deployed in the grain warehouse; and simultaneously collects temperature and humidity change data of the grain pile. Specifically, the intelligent multispectral camera is activated after the pest control operation is completed, and performs multispectral imaging of the grain pile surface according to a preset sampling grid, acquiring the reflectance curve of each sampling point in the visible to near-infrared band (e.g., 400nm-1000nm). In this embodiment, the sampling point spacing can be set to 0.5m × 0.5m according to the grain warehouse area to ensure full coverage of the grain surface; at the same time, through the 3D structured light module integrated on the camera, a structured light pattern is projected and 3D point cloud data of the grain pile surface is collected, with a point cloud density of not less than 1000 points per square meter, thereby accurately describing the depth and width of the grain surface ditches and the number of insect activity trajectories. According to the analysis of continuous frame images, specifically, the camera continuously captures images of the grain surface at a frequency of 2 frames per second. The YOLOv8 target detection model is used to identify individual pests in the images, and the same pest in the previous and subsequent frames is associated with the Kalman filter tracking algorithm to generate a temporal location point sequence for each individual pest. Each trajectory contains at least 10 consecutive location points to ensure the validity of the trajectory. Tracking the movement trajectory of individual pests through the camera is an existing technology, so it will not be described in detail. The temperature and humidity change data of the grain pile are collected by wireless temperature and humidity sensor nodes pre-buried at different equidistant depths (e.g., 10cm, 30cm, 50cm) in the grain pile. The sensor sampling frequency is once every 10 minutes. After the pest control operation, the monitoring is continuously carried out for a fixed monitoring period (in this embodiment, the fixed monitoring period is set to 72 hours by default) to record the temperature and humidity change curves of each depth position over time.

[0019] The pesticide effect monitoring module is used to combine grain surface spectral data, grain pile three-dimensional data, and grain pile temperature and humidity change data to quantitatively analyze the uniformity of pesticide distribution on the grain surface and the degree of pesticide penetration in the grain pile; based on the uniformity of horizontal and vertical distribution and the degree of vertical penetration, the module comprehensively analyzes the duration of the pesticide effect in the grain pile. The insect path monitoring module is used to analyze the degree of path interception of pests during their movement on the grain surface by combining three-dimensional data of the grain pile and insect activity trajectory data. The pest inhibition analysis module is used to analyze the pest spread inhibition effect of the grain pile after pest control operations based on the duration of the pesticide's effect in the grain pile and the degree of path interception of pests during their movement on the grain surface. The grain pile management module is used to manage grain piles in grain warehouses based on the analysis results of the pest spread inhibition effect after pest control operations.

[0020] In this embodiment, as Figure 3As shown, by combining spectral data of the grain surface, three-dimensional data of the grain pile, and temperature and humidity variation data of the grain pile, the uniformity of the horizontal and vertical distribution of the pesticide on the grain surface is quantitatively analyzed, and the vertical penetration degree of the pesticide in the grain pile is analyzed, including the following specific steps: S21. Acquiring spectral data of grain surface and three-dimensional data of grain pile: After the pest control operation is completed, intelligent multispectral cameras deployed in the grain warehouse collect spectral data of grain surface at multiple sampling points on the grain pile surface, including the multispectral reflectance curve of each sampling point; the multispectral reflectance curve of each sampling point is standardized; the specific standardization method is: the difference between the multispectral reflectance of each sampling point and the mean of the multispectral reflectance of all sampling points is divided by the standard deviation of the multispectral reflectance of all sampling points; based on the partial least squares regression model established in advance through laboratory calibration, the standardized multispectral reflectance curves of each sampling point are converted into the pesticide concentration index of each sampling point to form an initial pesticide concentration distribution map of the grain surface; in this embodiment, the training set of the partial least squares regression model consists of different concentrations of pesticide samples prepared in the laboratory and their corresponding spectral data. In this embodiment, the optimal number of principal components is determined to be 5 through 10-fold cross-validation, and the determination coefficient of the model reaches above 0.95. The pesticide concentration indices of all sampling points are arranged according to spatial location to form an initial pesticide concentration distribution map of the grain surface; Specifically, the construction and application process of the partial least squares regression model is as follows: Grain particles of the same variety as those in the grain warehouse were taken and evenly spread in a simulated grain bin (1m x 1m x 0.5m) with a smooth surface. The same pesticide used in the field (e.g., phosphine or methyl pyrimidine) was prepared according to a preset concentration gradient: 0 mg / kg, 0.5 mg / kg, 1 mg / kg, 2 mg / kg, 5 mg / kg, 10 mg / kg, 20 mg / kg, and 50 mg / kg, for a total of eight concentration levels. Five parallel samples were prepared for each concentration level, for a total of 40 samples. The prepared pesticide solution was evenly sprayed onto the grain surface using a sprayer. After standing for 24 hours to allow the pesticide to fully adsorb onto the grain surface, nine sampling points (3×3 grid) were selected on each sample surface. Multispectral reflectance data were collected at each sampling point using a handheld multispectral probe, with a spectral acquisition range of 400 nm to 1000 nm and a sampling interval of 5 nm, for a total of 121 bands. During sampling, the probe was held perpendicular to the grain surface, 10 cm away. Each sampling point was sampled three times, and the average value was taken to reduce random errors. Simultaneously, a 10g sample was taken from the corresponding location at each sampling point, and the actual residual concentration of the pesticide was determined using high-performance liquid chromatography-mass spectrometry (HPLC-MS / MS) to obtain the true concentration value for that sampling point.

[0021] Data from 360 sampling points (40 samples × 9 sampling points) across 40 samples were compiled. Each sampling point included spectral reflectance values ​​across 121 bands as the independent variable and the residual drug concentration determined by chromatography as the dependent variable. All 360 data sets were stratified into a training set and a test set. The training set contained 288 data sets (80% of the total data), and the test set contained 72 data sets (20% of the total data). Stratified sampling was performed based on concentration levels to ensure a consistent proportion of each concentration level in the training and test sets.

[0022] The model is built using training data. The core principle of the model is to extract the latent components (principal components) from the spectral data that are most strongly correlated with drug concentration. The original 121 spectral bands are dimensionality-reduced into several mutually orthogonal principal components, and then a linear regression relationship is established between the principal components and drug concentration. The training set is randomly divided into 10 parts. One part is used as the validation set, and the remaining 9 parts are used as the training set. Validation is performed on principal component numbers from 1 to 20, and the root mean square error (RMSE) of prediction is calculated for each validation. The average RMSE of 10 validations is used as the performance index for that principal component number. The principal component number corresponding to the minimum average RMSE is selected, and the optimal principal component number is determined to be 5. The final model is retrained on the entire training set using the optimal principal component number. The trained model is then applied to the test set for validation, and the coefficient of determination is calculated. In this embodiment, the coefficient of determination of the model reaches above 0.95 after validation on the test set, indicating that the model in this embodiment has high prediction accuracy.

[0023] The trained partial least squares regression model was used to calculate the pesticide concentration index on the actual grain surface. Specifically, after the pest control operation was completed, a smart multispectral camera collected multispectral reflectance curves of each sampling point on the grain pile surface according to a preset sampling grid, collecting reflectance values ​​of 121 bands at each sampling point. The reflectance value of each sampling point was used as input to the trained partial least squares regression model. The model automatically extracted the five principal components most relevant to the pesticide concentration and calculated the pesticide concentration index (in milligrams per kilogram) for that sampling point through a regression equation. The pesticide concentration indices of all sampling points were arranged according to their spatial coordinates to form an initial pesticide concentration distribution map on the grain surface. This initial pesticide concentration distribution map was a two-dimensional matrix, where rows corresponded to the Y-coordinates of the sampling points, columns corresponded to the X-coordinates, and matrix elements represented the pesticide concentration index at the corresponding coordinate position.

[0024] S22. Collect three-dimensional data of the grain pile surface, including three-dimensional point cloud data of the grain pile surface. Mesh the three-dimensional point cloud data and calculate the ditch depth (the difference between the minimum and mean Z-values ​​of the point cloud within the grid cell) and ditch density (the total length of the ditch boundary lines per unit area) of each grid cell to construct a grain surface topographic feature map. Further, in this embodiment, Gaussian process regression is used to fit the mapping relationship between ditch depth, ditch density, and topographic influence coefficient. In this embodiment, the topographic influence coefficient is defined as the ratio of the actual residual concentration of the pesticide in the ditch area to the concentration in the flat area without ditches. The mapping relationship between ditch depth, ditch density, and topographic influence coefficient is obtained through an indoor simulated grain surface ditch experiment. Specifically, the implementation process of the indoor simulated grain surface ditch experiment is as follows: Thirty experimental grain bins were constructed, with five levels of ditch depth (0-30mm) and six levels of ditch density (0-2.5m / m²), resulting in 30 possible combinations of factors. After constructing the pre-designed ditches on the grain surface using a shaped scraper, a fixed amount of pesticide was sprayed according to on-site operational standards. After 24 hours of settling, samples were collected at a 0.3m grid density, totaling 900 sampling points. The ditch depth and density were recorded at each sampling point, and the pesticide residue concentration was determined using chromatography. Using the average concentration in flat areas as a benchmark, the topographic influence coefficient (the ratio of concentration in the ditch area to that in the flat area) was calculated for each region. 876 sets of valid data were divided into training and validation sets. Gaussian process regression was used to fit the mapping relationship between ditch depth, ditch density, and topographic influence coefficient. A squared exponential kernel was selected as the kernel function, and hyperparameters were optimized by maximizing the marginal log-likelihood. The validation set determination coefficient reached 0.92. In practical implementation of this example, the ditch depth and density of each grid cell can be input into the model to obtain the topographic influence coefficient, which can be used for topographic correction of pesticide concentration.

[0025] S23. Spatially align the grain surface topographic feature map with the initial grain surface pesticide concentration distribution map, and use the topographic features to perform topographic correction on the pesticide concentration to eliminate the pesticide pooling effect caused by ditches, thus obtaining a topographically corrected pesticide concentration distribution map. Topographic correction is achieved by dividing the pesticide concentration of each grid cell by the topographic influence coefficient of the corresponding grid cell. In this embodiment, the topographically corrected pesticide concentration distribution map can eliminate the phenomenon of artificially high concentration caused by pesticide pooling in ditches, so that the distribution map can truly reflect the uniformity of pesticide spread on the grain surface. S24. Using the terrain-corrected drug concentration distribution map as training samples, a convolutional autoencoder network is constructed. The convolutional autoencoder network consists of an encoder and a decoder. The encoder compresses the input image into a low-dimensional feature vector, and the decoder reconstructs the low-dimensional feature vector into an output image. The training objective is to minimize the root mean square error between the input image and the reconstructed image, enabling the network to learn the normal pattern of drug concentration distribution. In this embodiment, a convolutional autoencoder network is constructed to quantify the uniformity of the horizontal and vertical distribution. The specific network structure is as follows: the input layer receives a terrain-corrected drug concentration distribution map with a size of 64×64 pixels; the encoder part contains three convolutional layers, with the first layer having a kernel size of 5×5, a stride of 2, and 32 output channels. The second convolutional layer has a kernel size of 3×3, a stride of 2, and 64 output channels. The third convolutional layer has a kernel size of 3×3, a stride of 2, and 128 output channels. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The encoder outputs a 128-dimensional feature vector. The decoder contains three deconvolutional layers, symmetrical to the encoder structure, and finally outputs a reconstructed image with the same size as the input. Furthermore, the network uses the Adam optimizer with an initial learning rate of 0.001 and a loss function of root mean square error. During training, a Dropout layer (dropout rate of 0.3) is used to prevent overfitting, and an EarlyStopping strategy is used, stopping training when the validation set loss does not decrease for 10 consecutive epochs.

[0026] S25. Obtain pesticide concentration distribution maps collected after multiple historical pest control operations on the grain pile and corrected for topographical distribution (500 maps in this embodiment). Label the maps according to the uniformity of the actual pesticide distribution to obtain the labeled horizontal and vertical distribution data of pesticides from historical pest control operations. The uniformity of the actual distribution is obtained by measuring the pesticide residue through grid sampling and calculating the coefficient of variation. Specifically, on the actual grain surface corresponding to each map, samples are taken in a 2m × 2m grid, and the pesticide concentration is determined by high performance liquid chromatography. The coefficient of variation of pesticide concentration for all samples is calculated. The smaller the coefficient of variation, the higher the uniformity. The coefficient of variation is used as a label. The labeled historical pesticide distribution data was divided into training and validation sets, with 80% and 20% of the data respectively. A convolutional autoencoder was trained using the training set to reconstruct a uniformly distributed normal pattern, defined as a sample with a coefficient of variation below 0.2. The network hyperparameters were adjusted using the validation set until the reconstruction error stabilized and converged. In this embodiment, the hidden layer dimension was ultimately determined to be 128, and the learning rate decay factor was 0.5 (every 20 epochs), thus obtaining a trained model for analyzing the uniformity of the horizontal and vertical distribution. S26. Input the pesticide concentration distribution map obtained after pest control and corrected for terrain into the trained horizontal and vertical distribution uniformity analysis model of the current monitored grain pile. Calculate the reconstruction error corresponding to the current monitored grain pile after pest control. The reconstruction error reflects the degree of deviation between the current distribution and the normal pattern learned by the network. Divide the reconstruction error corresponding to the current monitored grain pile after pest control by the maximum reconstruction error corresponding to multiple historical pest control operations in the grain warehouse (i.e., the reconstruction error corresponding to the sample with the lowest coefficient of variation among all historical samples) to obtain the horizontal and vertical distribution uniformity of the pesticide on the grain surface after pest control in the current monitored grain pile.

[0027] In this embodiment, by combining spectral data of the grain surface, three-dimensional data of the grain pile, and temperature and humidity change data of the grain pile, the uniformity of the horizontal and vertical distribution of the agent on the grain surface is quantitatively analyzed, and the vertical penetration degree of the agent in the grain pile is analyzed. Specifically, the following steps are also included: S27. Collect three-dimensional point cloud data of the grain pile surface, calculate the ditch depth and ditch density at each sampling point, and form a grain pile feature vector; pre-embed temperature and humidity sensors at equidistant depths along the vertical direction within the grain pile, and continuously collect temperature and humidity change data at each depth after the pest control operation, including the temperature and humidity change rate at each depth within a fixed monitoring period after pesticide application; align the ditch feature vector with the temperature and humidity change rate at each depth according to spatial location to form a spatiotemporal sequence data containing surface topographic features; in this embodiment, temperature and humidity sensors at different depths (e.g., 10cm, 30cm, 50cm) of the grain pile are regarded as spatial nodes, and edges are established between adjacent depth nodes to form a chain-like spatiotemporal graph. Each node has a temperature and humidity change rate sequence within a fixed monitoring period (72 hours) in the time dimension, and the change rate is calculated every 6 hours, for a total of 12 time steps. Furthermore, in this embodiment, the grain pile characteristics (average ditch depth and ditch density at each sampling point) are used as global static features and embedded into the initial state of each node through a fully connected layer to realize the initial influence of surface topography on deep infiltration. S28. Based on spatiotemporal sequence data containing surface topographic features, a spatiotemporal graph convolutional network integrating topographic features is constructed. The spatiotemporal graph convolutional network is used to simultaneously extract the variation patterns in the time dimension and the transmission relationships in the spatial dimension, and incorporates the initial influence of surface topography on the longitudinal penetration of the agent; the spatiotemporal graph convolutional network outputs the agent penetration intensity at each depth location. S29. Obtain the actual pesticide penetration intensity at each depth location, determined by stratified sampling and actual measurement of pesticide residue concentration during multiple historical pesticide application operations on the grain pile; construct a longitudinal penetration analysis dataset by combining the actual pesticide penetration intensity at each depth location with the corresponding grain pile feature vector and the temperature and humidity change rate at each depth location; divide the longitudinal penetration analysis dataset into a training set and a validation set; train a spatiotemporal graph convolutional network that integrates terrain features using the training set, enabling the spatiotemporal graph convolutional network that integrates terrain features to predict the pesticide penetration intensity at each depth location based on the grain pile feature vector and the temperature and humidity change rate at each depth location, thus obtaining a pesticide penetration intensity analysis model; S210. Set weight coefficients for different depth positions on the trained pesticide penetration intensity analysis model. The weight of shallow depth positions is lower than that of deep depth positions. The pesticide penetration intensity at each depth position output by the pesticide penetration intensity analysis model is weighted and averaged according to the weight coefficients of the corresponding depth positions to obtain the longitudinal penetration degree of the pesticide in the grain pile. In this embodiment, the specific structure of the spatiotemporal graph convolutional network is as follows: The input layer receives spatiotemporal graph data, with a shape of (number of nodes × time step × feature dimension); the spatiotemporal graph convolutional layer uses ST-GCN modules, each module containing a spatial graph convolutional layer (using the Laplacian matrix to aggregate neighbor node information) and a temporal convolutional layer (kernel size of 3, stride of 2); the network stacks a total of 3 ST-GCN modules, with 64, 128, and 256 channels respectively; each ST-GCN module is followed by a Dropout layer (dropout rate of 0.2); finally, a global average pooling layer and a fully connected layer are connected to output the predicted value of drug penetration intensity at each depth location.

[0028] In this embodiment, the network training data comes from pesticide residue concentrations measured through stratified sampling during multiple historical pest control operations. After each pest control operation, samples were collected at the same depth in the grain pile using a sampler, and the pesticide concentration was determined using liquid chromatography-mass spectrometry. These measured pesticide concentration values ​​served as labels for the actual penetration intensity at each depth. A total of 200 sets of historical data were collected, each set including ditch characteristics, temperature and humidity change rates at each depth, and the corresponding actual penetration intensity. These were divided into training and validation sets at 80% and 20% respectively. During training, the Adam optimizer was used with an initial learning rate of 0.0005 and a loss function of mean squared error, with a learning rate decay (decreasing by 0.5 every 30 epochs). Furthermore, to prevent overfitting, an L2 regularization term (with a coefficient of 0.0001) was added to the loss function. The validation set was used for early stopping (stopping if there was no improvement after 15 epochs).

[0029] After the model training is completed, weight coefficients are set for each depth location. In this embodiment, the weight coefficients are learned from historical data through linear regression. Specifically, the predicted infiltration intensity at each depth is used as the independent variable, and the actual infiltration intensity is used as the dependent variable. The weights at each depth location are solved by the least squares method to ensure that the weights of shallow layers are lower than those of deep layers (for example, 0.1 at 10cm, 0.3 at 30cm, and 0.6 at 50cm). The current grain pile characteristics and the rate of change of temperature and humidity at each depth in the current monitoring period are input into the network to obtain the predicted infiltration intensity at each depth. Then, the weighted sum of the predicted values ​​at each depth location is obtained to get the longitudinal infiltration degree.

[0030] This embodiment eliminates the interference of grain surface ditches on the judgment of pesticide distribution uniformity through terrain correction, making the quantification of the uniformity of horizontal and vertical distribution more objective. The spatiotemporal graph convolutional network simultaneously considers the temperature and humidity change patterns in the time dimension and the vertical transmission relationship in the spatial dimension, and incorporates the initial influence of surface topography on the infiltration path, thereby accurately predicting the infiltration intensity of pesticides at different depths, providing a reliable basis for subsequent analysis of the duration of effect.

[0031] S211. Input the feature vector of the grain pile after pest control and the temperature and humidity change rate at each depth within a fixed monitoring period after pest control into the trained pesticide penetration intensity analysis model. Output the pesticide penetration intensity at each depth in the current pest control operation. Sum the products of the pesticide penetration intensity at all depths and the corresponding weight coefficients at the depths to obtain the longitudinal penetration degree of the pesticide in the grain pile after pest control.

[0032] In this embodiment, the duration of the agent's effect in the grain pile is comprehensively analyzed based on the uniformity of its horizontal and vertical distribution and the degree of vertical penetration, including the following specific steps: S212. Extract the uniformity of the horizontal and vertical distribution of pesticide on the grain surface and the degree of vertical penetration of pesticide in the grain pile after the current monitoring grain pile has been treated. S213. The uniformity of the horizontal and vertical distribution of the pesticide on the grain surface and the degree of vertical penetration of the pesticide in the grain pile after the pesticide application are weighted and summed to obtain the duration of the pesticide effect in the grain pile after the pesticide application.

[0033] In this embodiment, as Figure 4 As shown, by combining three-dimensional data of the grain pile and insect activity trajectory data, the degree of path interception of pests during their movement on the grain surface is analyzed, including the following specific steps: S31. Obtain the three-dimensional point cloud data of the grain pile surface from the three-dimensional data of the grain pile, project the three-dimensional point cloud data of the grain pile surface onto a two-dimensional plane, and generate a ditch distribution raster map with a resolution of 0.05 m × 0.05 m; wherein, each raster cell in the ditch distribution raster map stores the ditch depth value (the difference between the point cloud Z value and the average height of the grain surface) and the slope value (the gradient magnitude is calculated by the Sobel operator) at that location. S32. From the insect activity trajectory data, obtain the temporal sequence of individual insect locations obtained through continuous frame image detection, and treat each location sequence as an insect movement trajectory; for each insect movement trajectory, calculate the detour distance ratio, ditch crossing frequency, and speed variation variance; where, the detour distance ratio is the ratio of the actual movement path length (cumulative Euclidean distance) of the corresponding trajectory to the straight-line distance from the starting point to the ending point of the corresponding trajectory, the larger the ratio, the more severe the detour; the ditch crossing frequency is the number of grid cells in the ditch distribution grid map crossed by the corresponding trajectory, and whether the corresponding trajectory crossed any of the grid cells in the ditch distribution grid map is determined by a preset ditch depth threshold (5mm by default in this embodiment); the speed variation variance is the degree of fluctuation of the movement speed of the corresponding trajectory in the ditch-dense area (in this embodiment, the area where the proportion of grid cells with a ditch depth > 3mm exceeds 30% by default) (i.e., the variance of the instantaneous movement speed), the speed is calculated by removing adjacent frame bits at frame intervals, the larger the variance, the more unstable the movement. S33. Obtain pest movement trajectory data from multiple historical experiments in the grain pile. The pest movement trajectory data includes the detour distance ratio, ditch crossing frequency, speed variation variance, and obstruction degree label for each trajectory, which is calibrated through experiments. The obstruction degree label is obtained by setting experimental scenarios with different obstacle densities, observing the proportion of pest passage time, and normalizing the result. The pest movement trajectory data is used to construct a path interception analysis dataset, which is then divided into a path interception analysis training set and a path interception analysis validation set.

[0034] In this embodiment, the degree of path interception of pests during their movement on the grain surface is analyzed by combining three-dimensional data of the grain pile and insect activity trajectory data. The analysis also includes the following specific steps: S34. Construct a graph neural network model. The graph neural network model uses each location point on the pest's movement trajectory as a node, and the connection between adjacent location points as edges. Each node is appended with the ditch depth and slope value of the corresponding grid cell as node features. The graph neural network model is trained using a path interception analysis training set. The node features of each pest movement trajectory are used as input, and the obstruction degree label corresponding to the trajectory is used as the output target to obtain an initial path interception degree prediction model. The initial path interception degree prediction model is validated using a path interception analysis validation set. The model with a prediction error lower than a preset error threshold is selected as the path interception degree prediction model. In this embodiment, a graph neural network model is constructed for trajectory obstruction degree prediction. The specific network structure is as follows: Each location point on each trajectory is used as a node, and edges are established between adjacent locations to form a chain graph; the initial features of each node include the trench depth and slope of the grid where the point is located, as well as the instantaneous velocity of the point; the graph neural network adopts a 3-layer GraphSAGE architecture, the first layer aggregates the information of neighboring nodes and outputs 128-dimensional features, the second layer outputs 64-dimensional features, the third layer outputs 32-dimensional features, and finally a global average pooling layer and a fully connected layer are connected to output a blocked probability value between 0 and 1; each GraphSAGE layer is followed by the LeakyReLU activation function and Dropout (dropout rate 0.2). In this embodiment, the training dataset is derived from the movement trajectories of pests obtained through multiple historical experiments in grain piles. The experiments were conducted in a simulated grain warehouse, where a certain number of pests were released onto the grain surface with artificially set obstacle densities (ditch depth and density), and their movement trajectories were recorded using a camera. The degree of obstruction for each trajectory was calibrated experimentally: the proportion of time delay from the release point to the endpoint (the ratio of the actual passage time to the passage time on an unobstructed, flat grain surface) was measured and normalized to the [0,1] interval. A total of 1000 trajectory data points were collected and divided into training and validation sets at 80% and 20% respectively. During training, the Adam optimizer was used with an initial learning rate of 0.001 and the mean squared error loss function was adopted. EarlyStopping was used, and the training was stopped if the loss did not decrease after 10 rounds.

[0035] S35. Input the movement trajectory of each individual pest in the currently monitored grain pile into the path interception degree prediction model, and output the obstruction probability value of each pest movement trajectory. S36. In this embodiment, since the lengths of different trajectories vary greatly, longer trajectories are more representative of the overall movement. Therefore, a weighted average is used to calculate the overall path interception degree. Specifically, the actual movement path length of each pest movement trajectory in the currently monitored grain pile is obtained; the actual movement path lengths of all trajectories are summed to obtain the total movement path length; the proportion of the actual movement path length of each trajectory to the total movement path length is used as the weight coefficient of the corresponding trajectory; the product of the obstruction probability value of all pest movement trajectories and the weight coefficient of the corresponding trajectory is summed to obtain the path interception degree of the pests during the movement of grain on the surface of the currently monitored grain pile.

[0036] This embodiment uses a graph neural network to deeply couple the pest movement trajectory with the grain surface topography. Features such as detour distance ratio and ditch crossing frequency comprehensively characterize the obstruction effect of the terrain on pest movement. The weighted average method takes into account the differences in trajectory representativeness, making the assessment of path interception degree more accurate.

[0037] In this embodiment, as Figure 5 As shown, based on the duration of the pesticide's effect in the grain pile and the degree of path interception during the movement of pests on the grain surface, the effect of pest control on the spread of pests in the grain pile after pesticide application is analyzed, including the following specific steps: S41. Extract the duration of the pesticide's effect on the grain pile after insecticide application and the degree of path interception of pests during their movement on the grain surface of the current monitored grain pile. S42. The effect of pesticide on the grain pile after pest control and the degree of path interception of pests during their movement on the grain surface of the current monitored grain pile are weighted and summed to obtain the pest spread inhibition effect of the current monitored grain pile after pest control.

[0038] In this embodiment, based on the analysis results of the pest spread inhibition effect on grain piles after pest control operations, the management of grain piles in grain warehouses is carried out, including the following specific contents: S51. Obtain the pest spread inhibition effect of the monitored grain pile after the pest control operation. S52. A preset threshold for the inhibition effect is established. When the inhibition effect of pest spread on the currently monitored grain pile after pest control operations is greater than the threshold, normal monitoring is maintained on the current grain pile, and the current management status of the grain pile is upheld. When the inhibition effect of pest spread on the currently monitored grain pile after pest control operations is less than or equal to the threshold, the grain discharge operation of the currently monitored grain pile is suspended, and a comprehensive fumigation and pest control program is initiated. It should be noted that in this embodiment, all preset thresholds, weight coefficients, model hyperparameters, etc., are obtained through historical data fitting. The specific fitting method is as follows: collect all monitoring data (including grain surface spectrum, three-dimensional ditch data, temperature and humidity, insect trajectory, actual pest control effect, etc.) of no less than 50 pest control operations in the past three years of the grain warehouse. The monitoring data is imported into the calculation process of each step in this embodiment. The optimal parameter combination is determined through cross-validation and grid search to minimize the error between the final predicted value of pest spread inhibition effect and the actual observed value, thereby obtaining the values ​​of the preset threshold, weight coefficient, and model hyperparameter. For example, hyperparameters such as the hidden layer dimension of the convolutional autoencoder with uniform horizontal and vertical distribution, the number of layers in the spatiotemporal graph convolutional network, and the aggregation method of the graph neural network are all optimized using the coefficient of determination on the validation set. All models are trained using an early stopping mechanism to avoid overfitting, and the generalization ability of the models is verified on the test set to ensure the accuracy and stability of the output results of each module. Through the scheme in this embodiment, the distribution and persistence of pesticides in grain piles can be comprehensively quantified, the hindering effect of terrain on pest movement can be accurately assessed, and the overall pest spread inhibition effect can be derived. Based on this, targeted management measures can be taken, thereby significantly improving the scientificity and effectiveness of pest control in grain warehouses, reducing the overuse of chemical pesticides, and lowering storage losses.

[0039] Example 2 like Figure 6 As shown in the figure, this embodiment of the present invention provides a smart camera for grain warehouses based on the Internet of Things, including the following specific contents: In response to the collection command after the pest control operation, the spectral data of the grain surface on the grain pile inside the grain warehouse is collected, and the three-dimensional data of the grain pile on the grain pile surface is collected simultaneously. Based on multi-frame spectral data of grain surface in continuous time series, the activity area of ​​insects on the surface of grain pile is identified, and the activity trajectory data of insects is extracted based on the positional changes of insects in different time series frames. Simultaneously collect data on temperature and humidity changes inside the grain pile; The spectral data of the grain surface, the three-dimensional data of the grain pile, the data on the activity trajectory of insects, and the data on the temperature and humidity changes of the grain pile are packaged and uploaded to the aforementioned Internet of Things-based grain warehouse data management system.

[0040] The parameters and steps for implementing the corresponding functions of each unit module in the IoT-based smart camera for grain warehouses described above can be found in the embodiments of the IoT-based grain warehouse data management system described above, and will not be repeated here.

[0041] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0042] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A grain warehouse data management system based on the Internet of Things, characterized in that, It includes a multi-source acquisition module, a pesticide effect monitoring module, an insect path monitoring module, an insect pest blockage analysis module, and a grain pile management module; Among them, the multi-source acquisition module is used to collect spectral data of the grain surface, three-dimensional data of the grain pile, and insect activity trajectory data of the grain pile after pest control operations by using intelligent multispectral cameras deployed in the grain warehouse; and simultaneously collect data on temperature and humidity changes in the grain pile. The pesticide effect monitoring module is used to combine grain surface spectral data, grain pile three-dimensional data, and grain pile temperature and humidity change data to quantitatively analyze the uniformity of pesticide distribution on the grain surface and the degree of pesticide penetration in the grain pile; based on the uniformity of horizontal and vertical distribution and the degree of vertical penetration, the module comprehensively analyzes the duration of the pesticide effect in the grain pile. The insect path monitoring module is used to analyze the degree of path interception of pests during their movement on the grain surface by combining three-dimensional data of the grain pile and insect activity trajectory data. The pest inhibition analysis module is used to analyze the pest spread inhibition effect of the grain pile after pest control operations based on the duration of the pesticide's effect in the grain pile and the degree of path interception of pests during their movement on the grain surface. The grain pile management module is used to manage grain piles in grain warehouses based on the analysis results of the pest spread inhibition effect after pest control operations.

2. The grain warehouse data management system based on the Internet of Things according to claim 1, characterized in that, The method combines spectral data of the grain surface, three-dimensional data of the grain pile, and temperature and humidity variation data of the grain pile to quantitatively analyze the uniformity of the horizontal and vertical distribution of the pesticide on the grain surface, and analyzes the vertical penetration degree of the pesticide in the grain pile. This includes the following specific steps: S21. Acquire grain surface spectral data and grain pile three-dimensional data: After the pest control operation is completed, collect grain surface spectral data from multiple sampling points on the grain pile surface using intelligent multispectral cameras deployed in the grain warehouse. This includes the multispectral reflectance curve of each sampling point. Standardize the multispectral reflectance curve of each sampling point. Based on the partial least squares regression model established in advance through laboratory calibration, the standardized multispectral reflectance curves of each sampling point are converted into the pesticide concentration index of each sampling point to form an initial pesticide concentration distribution map of the grain surface. S22. Collect three-dimensional data of the grain pile surface, including three-dimensional point cloud data of the grain pile surface, grid the three-dimensional point cloud data, calculate the ditch depth and ditch density of each grid unit, and construct a topographic feature map of the grain surface. S23. Spatially align the grain surface topographic feature map with the initial grain surface pesticide concentration distribution map, and use the topographic features to perform topographic correction on the pesticide concentration to eliminate the pesticide pooling effect caused by ditches, thereby obtaining a topographically corrected pesticide concentration distribution map; the topographic correction is achieved by dividing the pesticide concentration of each grid cell by the topographic influence coefficient of the corresponding grid cell. S24. Using the terrain-corrected drug concentration distribution map as training samples, construct a convolutional autoencoder network; the convolutional autoencoder network consists of an encoder and a decoder. The encoder is used to compress the input image into a low-dimensional feature vector, and the decoder is used to reconstruct the low-dimensional feature vector into an output image; the training objective is to minimize the root mean square error between the input image and the reconstructed image, so that the network learns the normal pattern of drug concentration distribution. S25. Obtain pesticide concentration distribution maps collected after multiple pesticide application operations on the grain pile and corrected for terrain, and label them according to the actual uniformity of pesticide distribution to obtain labeled historical pesticide distribution data. The uniformity of actual distribution is obtained by measuring pesticide residue through gridded sampling and calculating the coefficient of variation. Divide the labeled historical pesticide distribution data into training set and validation set. Use the training set to train a convolutional autoencoder and use the validation set to adjust the network hyperparameters until the reconstruction error stabilizes and converges, thus obtaining a trained model for analyzing the uniformity of the horizontal and vertical distribution. S26. Input the pesticide concentration distribution map obtained after the current monitoring grain pile has been treated with pesticides and corrected for terrain into the trained horizontal and vertical distribution uniformity analysis model, and calculate the reconstruction error corresponding to the current monitoring grain pile after pesticide treatment; divide the reconstruction error corresponding to the current monitoring grain pile after pesticide treatment by the maximum reconstruction error corresponding to multiple pesticide treatment operations in the grain warehouse in history to obtain the horizontal and vertical distribution uniformity of pesticides on the grain surface after pesticide treatment in the current monitoring grain pile.

3. A grain warehouse data management system based on the Internet of Things according to claim 2, characterized in that, The method combines spectral data of the grain surface, three-dimensional data of the grain pile, and temperature and humidity variation data of the grain pile to quantitatively analyze the uniformity of the horizontal and vertical distribution of the pesticide on the grain surface, and analyzes the vertical penetration degree of the pesticide in the grain pile. Specifically, it also includes the following steps: S27. Collect three-dimensional point cloud data of the grain pile surface, calculate the ditch depth and ditch density at each sampling point, and form a grain pile feature vector; embed temperature and humidity sensors at equal depths along the vertical direction in the grain pile, and continuously collect temperature and humidity change data of the grain pile at each depth after the pest control operation is completed, including the temperature and humidity change rate at each depth within a fixed monitoring period after pesticide application; align the ditch feature vector with the temperature and humidity change rate at each depth according to spatial location to form a spatiotemporal sequence data containing surface topographic features; S28. Based on the spatiotemporal sequence data containing surface topographic features, a spatiotemporal graph convolutional network integrating topographic features is constructed. The spatiotemporal graph convolutional network is used to simultaneously extract the change patterns in the time dimension and the transmission relationship in the spatial dimension, and incorporates the initial influence of surface topography on the longitudinal penetration of the agent. The spatiotemporal graph convolutional network outputs the agent penetration intensity at each depth position. S29. Obtain the actual pesticide penetration intensity at each depth location, determined by stratified sampling and actual measurement of pesticide residue concentration during multiple pesticide application operations on the grain pile. Combine the actual pesticide penetration intensity at each depth location with the corresponding grain pile feature vector and the temperature and humidity change rate at each depth location to form a longitudinal penetration analysis dataset. Divide the longitudinal penetration analysis dataset into a training set and a validation set. Use the training set to train a spatiotemporal graph convolutional network that integrates terrain features, so that the spatiotemporal graph convolutional network that integrates terrain features can predict the pesticide penetration intensity at each depth location based on the grain pile feature vector and the temperature and humidity change rate at each depth location, thus obtaining a pesticide penetration intensity analysis model. S210. Set weight coefficients for different depth positions on the trained agent penetration intensity analysis model, wherein the weight of shallow depth positions is lower than that of deep depth positions; calculate the weighted average of the agent penetration intensity at each depth position output by the agent penetration intensity analysis model according to the weight coefficient of the corresponding depth position to obtain the longitudinal penetration degree of the agent in the grain pile. S211. Input the feature vector of the grain pile after pest control and the temperature and humidity change rate at each depth within a fixed monitoring period after pest control into the trained pesticide penetration intensity analysis model. Output the pesticide penetration intensity at each depth in the current pest control operation. Sum the products of the pesticide penetration intensity at all depths and the corresponding weight coefficients at the depths to obtain the longitudinal penetration degree of the pesticide in the grain pile after pest control.

4. A grain warehouse data management system based on the Internet of Things according to claim 3, characterized in that, The analysis of the duration of the agent's effect in the grain pile, based on the uniformity of horizontal and vertical distribution and the degree of vertical penetration, includes the following specific steps: S212. Extract the uniformity of the horizontal and vertical distribution of pesticide on the grain surface and the degree of vertical penetration of pesticide in the grain pile after the current monitoring grain pile has been treated. S213. The uniformity of the horizontal and vertical distribution of the pesticide on the grain surface and the degree of vertical penetration of the pesticide in the grain pile after the pesticide application are weighted and summed to obtain the duration of the pesticide effect in the grain pile after the pesticide application.

5. A grain warehouse data management system based on the Internet of Things according to claim 4, characterized in that, The analysis of the degree of path interception of pests during their movement on the grain surface, combining three-dimensional data of the grain pile and insect activity trajectory data, includes the following specific steps: S31. Obtain the three-dimensional point cloud data of the grain pile surface from the three-dimensional data of the grain pile, and project the three-dimensional point cloud data of the grain pile surface onto a two-dimensional plane to generate a ditch distribution grid map; wherein, each grid cell in the ditch distribution grid map stores the ditch depth value and slope value at that location. S32. From the insect activity trajectory data, obtain the temporal sequence of individual insect locations detected by continuous frame images, and treat each location sequence as an insect movement trajectory; for each insect movement trajectory, calculate the detour distance ratio, ditch crossing frequency, and velocity variation variance; wherein, the detour distance ratio is the ratio of the actual movement path length of the corresponding trajectory to the straight-line distance from the start point to the end point of the corresponding trajectory; the ditch crossing frequency is the number of grid cells in the ditch distribution grid map traversed by the corresponding trajectory; the velocity variation variance is the degree of fluctuation in the movement speed of the corresponding trajectory in the dense ditch area; S33. Obtain pest movement trajectory data from multiple historical experiments in the grain pile. The pest movement trajectory data includes the detour distance ratio, ditch crossing frequency, speed variation variance, and obstruction degree label for each trajectory, calibrated experimentally. The obstruction degree label is obtained by setting experimental scenarios with different obstacle densities, observing the proportion of pest passage time, and normalizing the results. The pest movement trajectory data is used to construct a path interception analysis dataset, which is then divided into a path interception analysis training set and a path interception analysis validation set.

6. A grain warehouse data management system based on the Internet of Things according to claim 5, characterized in that, The analysis of the degree of path interception of pests during their movement on the grain surface, combining three-dimensional data of the grain pile and insect activity trajectory data, also includes the following specific steps: S34. Construct a graph neural network model, wherein the graph neural network model uses each position point on the pest's movement trajectory as a node, the connection between adjacent position points as an edge, and each node is attached with the ditch depth value and slope value of the corresponding grid cell as node features. The graph neural network model is trained using the path interception analysis training set. The node features of each pest movement trajectory are used as input, and the obstruction degree label corresponding to the trajectory is used as the output target to obtain an initial path interception degree prediction model. The initial path interception degree prediction model is then verified using the path interception analysis validation set, and the model with a prediction error lower than a preset error threshold is selected as the path interception degree prediction model. S35. Input the movement trajectory of each individual pest in the currently monitored grain pile into the path interception degree prediction model, and output the obstruction probability value of each pest movement trajectory. S36. Obtain the actual movement path length of each pest movement trajectory in the currently monitored grain pile; sum the actual movement path lengths of all trajectories to obtain the total movement path length; The proportion of the actual movement path length of each trajectory to the total movement path length is used as the weight coefficient of the corresponding trajectory. The product of the obstruction probability value of all pest movement trajectories and the weight coefficient of the corresponding trajectory is summed to obtain the path interception degree of pests in the current monitoring grain pile surface movement process.

7. A grain warehouse data management system based on the Internet of Things according to claim 6, characterized in that, The analysis of the pest control effect on the grain pile after pest control operations, based on the duration of the pesticide's effect in the grain pile and the degree of path interception during the movement of pests on the grain surface, includes the following specific steps: S41. Extract the duration of the pesticide's effect on the grain pile after insecticide application and the degree of path interception of pests during their movement on the grain surface of the current monitored grain pile. S42. The effect of pesticide on the grain pile after pest control and the degree of path interception of pests during their movement on the grain surface of the current monitored grain pile are weighted and summed to obtain the pest spread inhibition effect of the current monitored grain pile after pest control.

8. A grain warehouse data management system based on the Internet of Things according to claim 7, characterized in that, Based on the analysis results of the pest control effect on grain piles after pest control operations, the management of grain piles in grain warehouses includes the following specific contents: S51. Obtain the pest spread inhibition effect of the monitored grain pile after the pest control operation. S52. Preset the threshold for the inhibition effect. When the inhibition effect of pest spread on the currently monitored grain pile after the pest control operation is greater than the threshold, normal monitoring of the currently monitored grain pile is carried out to maintain the current grain pile management status. When the inhibition effect of pest spread on the currently monitored grain pile after the pest control operation is less than or equal to the threshold, the grain discharge operation of the currently monitored grain pile is suspended and the full fumigation and pest control program is started.

9. A smart camera for grain warehouses based on the Internet of Things (IoT), applied to the IoT-based grain warehouse data management system according to any one of claims 1-8, characterized in that, Includes the following specific content: In response to the collection command after the pest control operation, the spectral data of the grain surface on the grain pile inside the grain warehouse is collected, and the three-dimensional data of the grain pile on the grain pile surface is collected simultaneously. Based on multi-frame spectral data of grain surface in continuous time series, the activity area of ​​insects on the surface of grain pile is identified, and the activity trajectory data of insects is extracted based on the positional changes of insects in different time series frames. Simultaneously collect data on temperature and humidity changes inside the grain pile; The spectral data of the grain surface, the three-dimensional data of the grain pile, the insect activity trajectory data, and the temperature and humidity change data of the grain pile are packaged and uploaded to the grain warehouse data management system based on the Internet of Things as described in any one of claims 1-8.

Images