Infrared thermal feature perception-based intelligent prevention and control system and method for grain storage insect pests

Abstract

The application discloses a grain depot insect pest intelligent prevention and control system and method based on infrared thermal feature sensing, and belongs to the technical field of grain storage prevention and control, and comprises the following steps: collecting grain pile thermal radiation data through an indoor infrared thermal imager array, extracting insect pest thermal spot features and temperature difference gradient features through thermal feature analysis to generate insect pest feature data; inputting the insect pest feature data into a transfer learning model to analyze insect pest breeding rules, matching a prevention and control rule base to generate a targeted prevention and control trigger instruction; and finally, responding to the instruction to control the targeted prevention and control equipment, performing precise prevention and control work on the insect pest thermal spot corresponding grain pile area according to target area coordinates, prevention and control operation parameters and the like; the application realizes all-around collection of grain pile surface layer and internal thermal radiation through the infrared thermal imager array, accurately analyzes insect pest breeding rules in combination with the transfer learning model, realizes deep insect pest sensing and precise pesticide application in the aid of the targeted prevention and control, effectively reduces the pesticide usage amount, reduces grain quality loss, and improves the intelligent and precise level of grain depot insect pest prevention and control.

Description

Technical Field

[0001] This invention belongs to the field of grain storage pest control technology, specifically an intelligent pest control system and method for grain storage based on infrared thermal feature sensing. Background Technology

[0002] Traditional pest control in grain depots relies heavily on manual inspections, sampling tests, and comprehensive pesticide spraying. This approach suffers from technical drawbacks such as delayed detection and low precision. Manual inspections and sampling tests can only detect pests on the surface of the grain pile, making it difficult to detect pest activity in deeper layers, easily missing early control opportunities. Furthermore, test results are greatly affected by human factors, making it impossible to accurately grasp pest reproduction patterns. Comprehensive pesticide spraying not only results in significant pesticide waste and increases storage and pest control costs, but also easily leads to excessive pesticide residues in grain, reducing grain quality. Additionally, excessive pesticide application can damage the ecological environment within the grain depot.

[0003] Existing pest detection technologies often rely on single infrared thermal imagers to collect grain pile temperature data, providing only two-dimensional surface temperature information. This lacks the ability to perceive the internal temperature of the grain pile and effectively extracts pest-related thermal characteristics, hindering accurate pest identification and location. Furthermore, existing pest analysis models are mostly general-purpose models, not optimized for specific grain storage scenarios. Their accuracy in predicting pest development stages, population density, and outbreak probability is insufficient, making it difficult to support precise control decisions. In addition, existing pesticide application equipment lacks intelligent integration with the detection system, preventing targeted and quantitative application based on pest distribution characteristics, resulting in low levels of automation and intelligence in pest control operations. Therefore, there is an urgent need for a technological solution that enables deep perception and precise analysis of grain pile pests, coupled with intelligent targeted control. This solution would address the problems of incomplete detection, inaccurate analysis, and unscientific control methods associated with traditional methods, reducing pesticide use and grain storage losses, and improving the intelligence and precision of pest control in grain storage facilities. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an intelligent pest control system and method for grain storage based on infrared thermal feature perception. The system collects thermal radiation data from grain piles using an array of infrared thermal imagers inside the storage facility. Through thermal feature analysis, pest hotspot features and temperature gradient features are extracted to generate pest feature data. This data is then input into a transfer learning model to analyze pest reproduction patterns and match them with a control rule base to generate targeted control trigger commands. Finally, the system responds to these commands to control targeted control equipment, performing precise control operations on the grain pile area corresponding to the pest hotspots based on the target area coordinates and control operation parameters. This invention achieves comprehensive collection of thermal radiation from the surface and interior of grain piles using an infrared thermal imager array, accurately analyzes pest reproduction patterns using a transfer learning model, and achieves deep pest perception and precise pesticide application through targeted control. This effectively reduces pesticide usage, lowers grain quality losses, and improves the intelligence and precision of pest control in grain storage facilities.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Intelligent pest control methods for grain depots based on infrared thermal signature sensing include:

[0007] Acquire thermal radiation data of the grain pile collected by an array of infrared thermal imagers inside the warehouse;

[0008] Based on the heat radiation data of the grain pile, thermal feature analysis is performed to extract the characteristics of insect pest hot spots and temperature difference gradient features, and insect pest feature data is generated.

[0009] The pest characteristic data is input into a preset transfer learning model to analyze the pest reproduction pattern, and a targeted control trigger command is generated based on the analysis results.

[0010] In response to receiving the targeted control trigger command, the targeted control equipment is controlled to perform control operations on the grain pile area corresponding to the insect infestation hot spots.

[0011] Specifically, acquiring the thermal radiation data of the grain pile collected by the infrared thermal imager array inside the warehouse includes:

[0012] A first infrared thermal imager group is set up along a preset grid below the grain silo roof to collect thermal radiation from the surface of the grain pile. At the same time, at least one detection rod is set up vertically at a preset depth inside the grain pile. A second infrared thermal imager group is set up at equal intervals along the direction of the rod on each detection rod to collect thermal radiation from different depth layers inside the grain pile.

[0013] According to the preset synchronous scanning cycle, the first infrared thermal imager group and the second infrared thermal imager group are controlled to synchronously collect raw thermal radiation intensity values. The raw thermal radiation intensity values ​​collected by each infrared thermal imager are converted into temperature values. Based on the three-dimensional spatial coordinates and pixel field of view parameters of each infrared thermal imager in the grain silo, spatial coordinate registration and data fusion are performed on all temperature values ​​to generate three-dimensional temperature field distribution data of the grain pile that characterizes the temperature spatial distribution of the entire grain pile, which serves as the thermal radiation data of the grain pile.

[0014] Specifically, based on the thermal radiation data of the grain pile, thermal feature analysis is performed to extract insect pest heat spot features and temperature difference gradient features to generate insect pest feature data, including:

[0015] Based on the three-dimensional temperature field distribution data of the grain pile, a three-dimensional local maximum detection algorithm is used for traversal scanning. When the temperature value of any spatial point is detected to be higher than the average temperature of all points in its three-dimensional 26-neighborhood and the temperature difference exceeds the first preset threshold, the corresponding point is marked as a suspected hot spot seed point. Three-dimensional region growth is performed on the suspected hot spot seed points. Points with temperature differences within the second preset threshold are merged into the same connected region. Then, the geometric center coordinates, volume, surface area, and peak temperature of the connected region are calculated, and the geometric center coordinates, volume, surface area, and peak temperature are used as the characteristics of insect-infested hot spots.

[0016] Simultaneously, based on the three-dimensional temperature field distribution data of the grain pile, the Sobel operator is used to calculate the temperature gradient components of each spatial point in the horizontal, vertical and depth directions. The gradient vector magnitude of the corresponding point is synthesized according to the temperature gradient components, and the rate of change of the gradient vector magnitude and the consistency index of the gradient direction within the preset spatial scale are statistically analyzed to generate temperature difference gradient features.

[0017] The insect pest hot spot features and the temperature difference gradient features are concatenated to form a multi-dimensional feature vector describing the current insect pest activity status of the grain pile, i.e., insect pest feature data.

[0018] Specifically, the pest characteristic data is input into a preset transfer learning model to analyze the pest reproduction patterns, and targeted control trigger commands are generated based on the analysis results, including:

[0019] A pre-stored deep convolutional neural network model trained on a large dataset from the source domain is loaded as the base network; the base network includes at least an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer;

[0020] The network parameters of the first convolutional layer in the basic network are fixed, and the network parameters of the second convolutional layer and the fully connected layer in the basic network are fine-tuned using the historical pest feature data labeled in the target domain to form a transfer learning model adapted to the current grain depot scenario.

[0021] The real-time extracted pest feature data is input into the transfer learning model with fine-tuned parameters. After layer-by-layer feature transformation and nonlinear mapping within the model, the classification result of the current developmental stage of the pest population, the pest population density level, and the predicted value of the pest outbreak probability within a preset time period are output.

[0022] Based on the developmental stage classification results, insect population density levels, and predicted insect outbreak probability, a preset prevention and control rule library is matched. When the triggering conditions are determined to be met, a targeted prevention and control triggering instruction containing the target area coordinates, prevention and control operation parameters, and execution time period is generated.

[0023] Specifically, the deep convolutional neural network model adopts the Densenet121 architecture. Its input layer receives normalized pest feature data. Its multiple densely connected blocks concatenate all feature maps of the previous layer along the channel dimension as input to the next layer. Its transition layer connects adjacent densely connected blocks to reduce the feature map size. Its pooling layer converts the feature map output by the last densely connected block into a feature vector. Its fully connected layer uses the Adam optimizer combined with the cross-entropy loss function for parameter fine-tuning. Its output layer uses the Softmax activation function to output the pest development stage classification result, wherein the pest development stage includes at least the egg stage, larval stage, pupal stage, and adult stage.

[0024] Specifically, the source domain big data set is a publicly available grain insect image dataset or thermal imaging dataset collected from grain depots in multiple different ecological zones; the historical pest feature data labeled in the target domain is data collected and labeled by infrared thermal imager arrays during the historical storage period of the current grain depot, showing the actual occurrence and reproduction stages of pests. The general feature representation capabilities learned in the source domain are transferred to the specified scenario of the current grain depot through transfer learning.

[0025] Specifically, the matching preset prevention and control rule base, when determining that the triggering condition is met, generates a targeted prevention and control triggering instruction containing the target area coordinates, prevention and control operation parameters, and execution time period, including:

[0026] Query the control rule base to obtain the minimum insect population density trigger threshold corresponding to the classification result of the current insect development stage;

[0027] The insect population density level is converted into an insect population density value, and it is determined whether the insect population density value is greater than or equal to the minimum insect population density trigger threshold.

[0028] If the judgment result is yes, then obtain the control operation parameter template that is jointly associated with the current pest development stage classification result and the pest outbreak probability prediction value within the preset time period; the control operation parameter template includes at least the pesticide type, pesticide concentration, pesticide dosage and operation mode.

[0029] The coordinates of the target area are determined based on the geometric center coordinates of the pest hot spot features. The control operation parameters are filled in according to the control operation parameter template. The execution period is determined according to the pest outbreak probability prediction value. The targeted control trigger command is then assembled and generated.

[0030] Specifically, the step of responding to the targeted control trigger command and controlling the targeted control equipment to perform control operations on the grain pile area corresponding to the insect infestation hotspots includes:

[0031] The targeted prevention and control trigger command is parsed to obtain the target area coordinates and prevention and control operation parameters;

[0032] Based on the coordinates of the target area, plan the movement path of the targeted control equipment in the grain warehouse, and calculate the attitude angle of the end effector of the equipment when it reaches the target position;

[0033] At the start of the execution period, a movement control command is sent to the targeted control device to drive it to move along the planned path and adjust its posture to the target position;

[0034] Once the targeted control equipment is in place, a pesticide application control command is generated based on the control operation parameters. This command controls the pesticide pipeline, proportional valve, and atomizing nozzle of the targeted control equipment to work together to apply pesticides to the target area.

[0035] Specifically, the targeted control device is a track-mounted mobile spraying robot, whose end effector is a multi-degree-of-freedom robotic arm, with at least one independently controlled atomizing nozzle mounted at the end of the robotic arm; the generation of spraying control instructions specifically involves: selecting the corresponding pesticide pipeline according to the pesticide identification in the control operation parameters; controlling the opening degree of the ratio valve between the mother liquor and dilution water according to the pesticide concentration in the control operation parameters; and controlling the opening and closing sequence and flow rate of the atomizing nozzle according to the spraying rate and operation mode in the control operation parameters.

[0036] The intelligent pest control system for grain depots based on infrared thermal signature sensing includes:

[0037] The data acquisition module synchronously collects the original thermal radiation intensity values ​​of the surface layer and different depth layers of the grain pile. After temperature conversion, spatial coordinate registration and data fusion, it generates three-dimensional temperature field distribution data of the grain pile.

[0038] The insect pest thermal feature analysis module extracts insect pest hot spot features and temperature difference gradient features based on the three-dimensional temperature field distribution data of the grain pile, and concatenates the insect pest hot spot features and temperature difference gradient features into insect pest feature data.

[0039] The transfer learning model analysis module is used to build a transfer learning model adapted to the current grain depot scenario. Real-time pest feature data is input into the transfer learning model, and after feature transformation and nonlinear mapping, the model outputs the classification results of pest development stage, pest population density level, and the predicted value of pest outbreak probability within a preset time period.

[0040] The targeted prevention and control instruction generation module, based on the analysis results of the transfer learning model, matches the preset prevention and control rule library and assembles and generates targeted prevention and control trigger instructions containing target area coordinates, prevention and control operation parameters, and execution time period;

[0041] The targeted prevention and control equipment execution module receives and parses the targeted prevention and control trigger command, extracts the prevention and control operation parameters, drives the equipment to move and adjust its posture to the target position during the specified execution period, and controls the equipment's agent pipeline, proportional valve, and atomizing nozzle to work together.

[0042] The prevention and control rule base and parameter management module is used to store and manage various rules, thresholds and parameter templates for the entire prevention and control process.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] 1. This invention proposes an intelligent pest control system for grain depots based on infrared thermal feature sensing, and optimizes and improves its architecture, operation steps and processes. The system has the advantages of simple process, low investment and operating costs and low production costs.

[0045] 2. This invention proposes an intelligent pest control method for grain storage based on infrared thermal feature perception. This method utilizes a layered deployment and synchronous acquisition of an infrared thermal imager array within the storage facility. Combined with spatial coordinate registration and data fusion, it generates three-dimensional temperature field distribution data of the grain pile. Then, it accurately extracts the thermal spot features and temperature gradient features of pests. Coupled with a transfer learning model optimized based on the Densenet121 architecture, it achieves precise analysis of pest development stages, population density, and outbreak probability. This method overcomes the pain points of traditional grain storage pest detection, which struggles to detect deep-seated pests and has low identification accuracy. It enables early identification of pests and accurate prediction of their reproduction patterns, providing a scientific and accurate decision-making basis for pest control operations and significantly improving the intelligence and precision of grain storage pest detection.

[0046] 3. This invention proposes an intelligent pest control method for grain depots based on infrared thermal feature sensing. This method generates targeted control instructions containing target areas and control operation parameters based on pest analysis results. It controls a track-mounted spraying robot to perform directional and quantitative precise spraying in adaptive modes such as point, strip, and area spraying, replacing the traditional whole-area spraying method. This effectively reduces the amount of pesticides used, lowers pesticide residues and quality loss in grain, and ensures grain storage safety. Moreover, the entire control process is automated and intelligently linked, requiring minimal manual intervention, improving the operational efficiency of pest control in grain depots, and reducing the labor and material costs of warehousing control. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the intelligent pest control method for grain depots based on infrared thermal feature sensing according to the present invention.

[0048] Figure 2 This is a flowchart illustrating the principle of the intelligent pest control method for grain depots based on infrared thermal feature sensing, as described in this invention.

[0049] Figure 3 This is a diagram illustrating the architecture of the intelligent pest control system for grain depots based on infrared thermal feature sensing, as described in this invention. Detailed Implementation

[0050] Example 1:

[0051] Please see Figure 1 and Figure 2 This invention provides an embodiment of an intelligent pest control method for grain depots based on infrared thermal feature sensing. The method includes steps S1 to S4. The following detailed explanation is based on a real pest control case from a grain depot. This grain depot has 12 standardized flat warehouses. This embodiment uses warehouse No. 5, which stores 3000 tons of corn, as the specific application object, and includes the following steps:

[0052] S1: Acquire the thermal radiation data of the grain pile collected by the infrared thermal imager array inside the warehouse;

[0053] In this embodiment, 12 first infrared thermal imagers were evenly deployed in a 6m x 6m grid beneath the ceiling of warehouse No. 5. Their lenses pointed vertically downwards to collect thermal radiation from the surface of the grain pile. Simultaneously, to detect insect activity inside the grain pile, six specially designed detection rods were pre-embedded within it. These rods were evenly distributed along the central axis and sides of the warehouse, vertically inserted into the grain pile to its bottom. On each rod, a second infrared thermal imager was installed at 1.5m intervals along the rod's direction, for a total of four rods, or 24 in total. These probes were placed close to the grain to collect thermal radiation data from different depths within the grain pile. The system operated according to a preset synchronous scanning cycle, which in this embodiment was every 30 minutes. At 2:00 AM on July 15, 2024, the control system simultaneously issued a collection command to all 36 infrared thermal imagers. All imagers synchronously collected the raw thermal radiation intensity values ​​within their field of view and immediately transmitted the data back to the edge computing server on-site via the warehouse's high-speed industrial Ethernet. After receiving the data, the server first converts the raw thermal radiation intensity values ​​into temperature values ​​with an accuracy of 0.1℃, based on the calibration curves built into each thermal imager. Then, the server calls a pre-stored 3D spatial model of the grain silo. This model contains the precise 3D spatial coordinates (X, Y, Z) of each thermal imager and the field of view parameters corresponding to each pixel. X, Y, and Z represent the coordinate values ​​of the horizontal, vertical, and axial axes, respectively. Through spatial coordinate transformation and interpolation algorithms, the server performs spatial coordinate registration and data fusion on 36 sets of 2D temperature image data from different angles and depths. This process is similar to stitching together 36 photos from different angles into a complete, three-dimensional temperature sculpture. Finally, a 3D temperature field distribution data of the grain pile that can truly reflect the spatial temperature distribution of the entire grain pile in Warehouse No. 5 is generated, which is the required grain pile thermal radiation data. At this point, the data shows that the temperature in most areas of the grain pile is stable at 22-24℃, but in the southeast corner, 2 meters deep from the grain surface, there is an abnormally high temperature zone with a diameter of about 0.8 meters, and the center temperature reaches 28.5℃.

[0054] S2: Based on the heat radiation data of the grain pile, perform thermal feature analysis, extract the characteristics of insect pest hot spots and temperature difference gradient features, and generate insect pest feature data.

[0055] In this embodiment, the edge computing server begins scanning and analyzing the three-dimensional temperature field distribution data of Warehouse No. 5. First, a three-dimensional local maximum detection algorithm is used to traverse the data, scanning each spatial point one by one. When the temperature of the center point of an abnormally high-temperature area in the southeast corner is detected (e.g., 28.5℃), significantly higher than the average temperature of all points within its 26-neighborhood (approximately 23.5℃), and the temperature difference exceeds a preset first threshold (set to 2.5℃ in this embodiment), the system marks this point as a suspected hotspot seed point. Then, starting from this seed point, three-dimensional region growth is performed, continuously absorbing and interacting with suspected hotspots. The hot spot seed points are connected in temperature, and all points with temperature differences within a second preset threshold (set to 1.0℃ in this embodiment) are merged into a single connected region. Finally, this abnormal region is accurately delineated, and the system automatically calculates the geometric center coordinates, volume, surface area, and peak temperature of the connected region. The geometric center coordinates (X=12.3m, Y=4.8m, Z=2.1m) are 12.3 meters from the southeast wall, 4.8 meters from the east wall, 2.1 meters deep, with a volume of approximately 0.5 cubic meters and a peak temperature of 28.5℃. These parameters are encapsulated as the characteristics of the insect pest hot spot.

[0056] Simultaneously, the server invokes the Sobel operator, including convolution kernels in the horizontal, vertical, and depth directions, to perform three-dimensional convolution operations on the three-dimensional temperature field distribution data. This is equivalent to calculating the rate of change and direction of change of the temperature field at each point. For example, at the edge of the newly delineated hot spot, the temperature drops sharply from 28.5℃ inside to 24℃ outside. Therefore, the gradient vector magnitude in this region is very large. The algorithm statistically analyzes the rate of change of gradient magnitude and the consistency of gradient direction within a preset spatial scale of the hot spot and its surroundings. The results show that the gradient directions at the edge of the hot spot are highly consistent, all pointing towards the center of the hot spot, forming a clear heat source outline. These calculation results are integrated into a temperature difference gradient feature.

[0057] Finally, the system concatenates the extracted pest hot spot features with the temperature gradient features to form a multi-dimensional feature vector that can comprehensively describe the current pest activity status in Warehouse No. 5, namely pest feature data. This vector is equivalent to drawing a digital portrait of the current pest situation.

[0058] S3: Input the pest characteristic data into a preset transfer learning model, analyze the pest reproduction pattern, and generate a targeted control trigger command based on the analysis results;

[0059] In this embodiment, a deep convolutional neural network model trained on a large source domain dataset is pre-loaded on the server as the base network. This large source domain dataset contains millions of thermal images and labeled data of grain pests collected from hundreds of grain depots in different ecological zones over the past five years. The model has learned to identify the general thermal feature patterns generated by the activities of various grain pests, such as corn weevils, grain borers, and wheat moths. In this embodiment, the base network adopts the powerful Densenet121 architecture. Since each grain depot has its own unique characteristics in terms of grain varieties, warehouse structure, and local climate, directly using a general model may not be accurate enough. Therefore, the parameters of the first convolutional layer in the Densenet121 network, which is responsible for extracting general features such as basic edges and textures, are fixed. Then, using the target domain data, namely, the data collected by Warehouse No. 5 during its storage period last year through the same infrared system and manually verified and labeled with the actual occurrence and breeding stages of pests, such as the larval stage - corn weevil - low-density historical data, the parameters of the second convolutional layer and fully connected layer responsible for high-level semantic understanding in the network are fine-tuned. After several rounds of iterative training, a transfer learning model adapted to the scenario of Warehouse No. 5 is generated. Now, the pest feature data extracted in real time in step S2 is input into this fine-tuned model. The model undergoes layer-by-layer feature transformation and nonlinear mapping, ultimately providing prediction results at the output layer. For the hot spot in the southeast corner, the model outputs developmental stage classification results, such as larval stage: probability 92%, insect population density level: level 3, corresponding to 30-50 insects per kilogram of grain, and predicted outbreak probability within 7 days: 85%. The classification results indicate that a larval population in its rapid feeding growth phase has a fairly high density and is highly likely to break out within a week, entering the pupal stage and emerging as adults, causing wider damage. The system immediately matches these analysis results with the preset control rule base. One rule in the rule base states: if the pest is in the larval stage and the insect population density is greater than or equal to level 3 and the outbreak probability is greater than 80%, a level 1 targeted control warning is triggered, and a point-based precision injection operation mode is recommended. After a successful match, the system automatically generates a targeted control trigger command. This command includes the target area coordinates (X=12.3m, Y=4.8m, Z=2.1m) and the calculated three-dimensional spatial range, as well as control operation parameters, such as pesticide type: grain insecticide microcapsule suspension; pesticide concentration: 2%; dosage: 50ml per cubic meter; operation mode: point-like precision injection; and suggested execution time, such as 2:30 AM, which meets the requirements for pesticide application, so execute immediately.

[0060] S4: In response to receiving the targeted prevention and control trigger command, control the targeted prevention and control equipment to perform precise prevention and control operations on the grain pile area corresponding to the insect infestation hot spots.

[0061] The targeted control equipment is a track-mounted mobile spraying robot, whose end effector is a multi-degree-of-freedom robotic arm. At the end of the robotic arm is at least one independently controlled atomizing nozzle. The specific spraying control instructions are as follows: select the corresponding pesticide pipeline according to the pesticide identification in the control operation parameters; control the opening degree of the ratio valve between the mother liquor and dilution water according to the pesticide concentration in the control operation parameters; and control the opening and closing sequence and flow rate of the atomizing nozzle according to the spraying rate and operation mode in the control operation parameters, so as to achieve directional, quantitative, and fixed-mode spraying of the pesticide solution.

[0062] In this embodiment, the targeted control equipment in Warehouse No. 5 is a track-mounted mobile spraying robot. Upon receiving the command, it immediately starts. The robot's control system first parses the command, accurately acquiring the target area coordinates, pesticide label (Grain Insect Killer), recommended pesticide concentration (2%), spraying rate (2 liters / minute), operating mode (precise point injection), and execution time period. Based on the target area coordinates (Z=2.1m), the robot calculates that it cannot reach the target area by the ceiling track alone. Therefore, it first moves along the ceiling track above the target area (X=12.3m, Y=4.8m), and then slowly lowers its multi-degree-of-freedom robotic arm. The atomizing nozzle at the end of the robotic arm precisely probes into the grain pile, reaching a depth of 2.1 meters. During the descent, the robot's control system calculates and adjusts the attitude angles of each joint of the robotic arm in real time based on feedback from the depth sensor, ensuring that the nozzle accurately reaches the predetermined three-dimensional coordinate point. Once the nozzle is in place, the control system generates detailed spraying control commands according to the precise point injection operating mode. First, open the corresponding pesticide pipeline for "Grain Insect Killer". Then, according to the recommended pesticide concentration of 2%, precisely control the opening of the ratio valve between the mother liquor and dilution water. Next, control the opening and closing sequence and flow rate of the atomizing nozzle at a spraying rate of 2 liters / minute, and begin directional and quantitative precise spraying on the target hot spot area. The pesticide is atomized into tiny particles, which are evenly covered on the densely infested grain grains. The entire process lasts about 3 minutes, and the spraying area is precisely controlled within a spherical area with a diameter of about 1 meter, minimizing pesticide waste and contamination of grain in other areas.

[0063] Furthermore, the targeted control device is a track-mounted mobile drug delivery robot. As shown in this embodiment, it selects the corresponding drug pipeline according to the drug label, controls the proportion valve according to the recommended drug concentration, and controls the atomizing nozzle to open intermittently in a pulse manner after reaching the target depth according to the drug application rate and the point-like precise injection mode, so as to achieve precise injection of the drug solution.

[0064] The acquisition of thermal radiation data of the grain pile collected by the infrared thermal imager array inside the warehouse includes:

[0065] S1.1: A first infrared thermal imager group is set up along a preset grid below the grain silo roof to collect thermal radiation from the surface of the grain pile. At the same time, at least one detection rod is set up vertically at a preset depth inside the grain pile. A second infrared thermal imager group is set up at equal intervals along the direction of the rod on each detection rod to collect thermal radiation from different depth layers inside the grain pile.

[0066] S1.2: According to the preset synchronous scanning cycle, which is once every 30 minutes in this embodiment, the first infrared thermal imager group and the second infrared thermal imager group are controlled to synchronously collect the raw thermal radiation intensity value. The raw thermal radiation intensity value collected by each infrared thermal imager is converted into temperature value. Based on the three-dimensional spatial coordinates and pixel field of view parameters of each infrared thermal imager in the grain warehouse, spatial coordinate registration and data fusion are performed on all temperature values ​​to generate three-dimensional temperature field distribution data of the grain pile that characterizes the spatial distribution of temperature of the entire grain pile, as the grain pile thermal radiation data.

[0067] In this embodiment, in warehouse 5, the first infrared thermal imager group collects the two-dimensional planar temperature distribution, like a top view of a grain pile, while the second infrared thermal imager group collects the temperature distribution on a vertical cross-section, like a side view of a grain pile. To fuse these two types of data, the edge computing server performs the following steps:

[0068] (1) For each pixel of the roof thermal imager, the three-dimensional coordinates (X, Y, 0) of the grain pile surface point corresponding to the pixel are calculated according to its installation position and field of view. For the thermal imager on the detection rod, the three-dimensional coordinates (X, Y, Z) of the grain pile interior point corresponding to each pixel are directly determined according to its installation height on the rod and the insertion position of the rod.

[0069] (2) Divide the entire grain warehouse space into a 10cm×10cm×10cm micro voxel grid;

[0070] (3) For each voxel, find all temperature sampling points with known coordinates within a certain radius, such as 20cm. Use the inverse distance weighted interpolation method, that is, the closer the sampling point is to the center of the voxel, the higher the weight of its temperature value. Calculate the temperature value of the voxel by weighted average. After all voxels are assigned values, a continuous and complete three-dimensional temperature field data volume is constructed. This three-dimensional temperature field data volume can be sliced ​​and observed from any angle.

[0071] Furthermore, the specific steps in S1.2 include:

[0072] (1) When deploying the system, establish a unified spatiotemporal reference for all infrared thermal imagers in the grain warehouse;

[0073] It is important to understand that before the system is officially put into operation, the geometric calibration and time synchronization of all infrared thermal imagers must be completed first. Technicians perform internal parameter calibration on each infrared thermal imager in a laboratory environment to determine its pixel field of view parameters. This parameter specifically refers to the spatial angle subtended by each pixel, which is used to convert pixel coordinates into actual spatial distance. Simultaneously, the optical distortion coefficients of the lens, including radial and tangential distortion coefficients, must be measured for geometric correction of the image. Subsequently, at the grain silo site, a high-precision total station is used to measure the precise position of the optical center of each infrared thermal imager in the grain silo's three-dimensional spatial coordinate system. This three-dimensional spatial coordinate system has the center of any pillar on the grain silo floor as the origin, the length of the silo as the X-axis, the width of the silo as the Y-axis, and the direction perpendicular to the ground upwards as the Z-axis. For the first infrared thermal imager group installed under the ceiling, technicians directly recorded the three-dimensional coordinates of the center point of its mounting bracket. For the second infrared thermal imager group embedded in the detection rods inside the grain pile, the planar coordinates of the insertion point at the bottom of the detection rod were first recorded. Then, combined with the installation height of each thermal imager on the rod, its precise three-dimensional spatial coordinates were calculated. After completing the spatial positioning, the control units of all infrared thermal imagers were connected to a time synchronization network based on the IEEE 1588v2 protocol. This protocol can synchronize the local clocks of all devices in the warehouse to microsecond-level accuracy, ensuring that the sampling time of each infrared thermal imager is strictly consistent when it receives the same trigger signal. It outputs the unique device number, precise three-dimensional spatial coordinates, pixel field of view parameters, optical distortion coefficient, and a unified microsecond-level time reference for the entire warehouse for each infrared thermal imager.

[0074] (2) After the system enters normal operation, the timer in the edge computing server broadcasts a synchronous trigger command to all 36 infrared thermal imagers in the warehouse according to the preset synchronous scanning cycle, which is set to once every 30 minutes in this embodiment. Since a time synchronization network has been established, all infrared thermal imagers simultaneously turn on their focal plane detectors at the same microsecond moment when they receive the command and begin to integrate the infrared radiation energy in the current field of view. Each thermal imager generates a two-dimensional original thermal radiation intensity value matrix according to its own physical resolution. In this embodiment, the resolution of the ceiling thermal imager is 640×480 pixels and the resolution of the probe thermal imager is 320×240 pixels. Each element in this matrix, that is, each pixel, records a dimensionless digital quantization value, which represents the sum of infrared radiation energy received by the tiny spatial area corresponding to the pixel during the detector integration time. After the data acquisition is completed, all thermal imagers immediately package and transmit the original thermal radiation intensity value matrix to the edge computing server via the high-speed industrial Ethernet in the warehouse. The data packet includes the precise timestamp of the image acquisition and the device number of the thermal imager itself. Finally, 36 original thermal radiation intensity value matrices with timestamps and device numbers from 36 thermal imagers are output.

[0075] (3) The original thermal radiation intensity matrix is ​​converted into a temperature value matrix with physical meaning, completing the initial transformation from digital signal to physical quantity. This includes: After receiving the original thermal radiation intensity matrix, the edge computing server first performs temperature conversion based on the radiation calibration curve of each infrared thermal imager calibrated at the factory and pre-stored in the server database. This radiation calibration curve is a complex functional relationship that establishes the correspondence between the digital quantization value output by the detector and the blackbody radiation temperature. However, the actual radiation transmission process is affected by ambient temperature, atmospheric attenuation, and the emissivity of the grain surface. Therefore, the conversion process is not a simple table lookup. The server first reads the ambient temperature sensor data in the current warehouse to correct the attenuation and interference of the atmospheric path on infrared radiation. Then, combining the emissivity of the grain, for the corn stored in this embodiment, this value is usually set to 0.95. The true radiative exitance of the target is calculated according to the basic law of thermal radiation. Finally, the corrected radiative exitance is accurately converted into a specific temperature value in degrees Celsius using the inverse function form of Planck's formula, accurate to one decimal place. This conversion process is performed on each pixel in each original thermal radiation intensity value matrix, thereby transforming 36 original radiation intensity images into 36 two-dimensional temperature value matrices filled with actual temperature data. The final output is 36 two-dimensional temperature value matrices with timestamps and device numbers. The inverse function form of Planck's formula is prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0076] (4) Based on the lens distortion coefficient obtained by each infrared thermal imager in geometric calibration, the two-dimensional temperature value matrix is ​​geometrically corrected to eliminate the pixel position error caused by optical lens distortion. Specifically, the correction process uses the nearest neighbor pixel resampling algorithm to find the non-integer coordinate position corresponding to each regular pixel coordinate in the output image in the original distorted image, and takes the temperature value of the nearest integer pixel around the position as the value of the output pixel, ensuring that each image becomes a distortion-free image that conforms to the ideal pinhole imaging model, so that the position of each pixel in the image is precisely corresponding to any unique spatial ray direction emitted from the optical center of the thermal imager, and finally outputs 36 geometrically accurate and distortion-free two-dimensional temperature value matrix images, which establish a strict geometric correspondence with the three-dimensional spatial coordinates of each thermal imager measured in (1). The nearest neighbor pixel resampling algorithm is the prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0077] (5) Based on the pixel field of view parameters, perform spatial coordinate mapping to convert each two-dimensional temperature value matrix image into a discrete point cloud with three-dimensional spatial coordinates. Taking a thermal imager installed on the ceiling as an example, it is installed at coordinates X=12.5m, Y=8.3m, Z=7.2m, with a horizontal field of view of 40 degrees and a vertical field of view of 30 degrees. For a pixel at the center of its image, based on the offset of its row and column coordinates from the center of the image and combined with the pixel field of view parameters, calculate the horizontal and vertical deflection angles of the pixel relative to the optical axis centerline of the thermal imager. Then, combined with the installation pitch angle and horizontal orientation angle of the thermal imager recorded in (1), calculate the direction vector of the spatial ray corresponding to the pixel in the unified coordinate system of the grain warehouse through three-dimensional spatial geometric transformation. Since the target observed by the ceiling thermal imager is the grain pile surface, the ray and the grain pile surface model, i.e., the Z coordinate, are equal to the grain surface. The intersection of the actual horizontal plane is the three-dimensional coordinate of the grain pile surface point corresponding to that pixel. For the thermal imager on the probe rod, its pixel corresponds to the intersection of the ray and the internal space of the grain pile within a preset radius centered on the probe rod. This radius is determined by the depth of field and focusing parameters of the thermal imager. Thus, each two-dimensional temperature value matrix image is converted into a discrete set of points with precise three-dimensional spatial coordinates. Each point contains information in four dimensions: X coordinate value, Y coordinate value, Z coordinate value, and temperature value T. After this step, the 36 images are converted into 36 sparse temperature point clouds, each point cloud being associated with the device number of its source thermal imager. Finally, 36 sparse temperature point clouds composed of four-dimensional data points with device numbers are output. The three-dimensional spatial geometric transformation operation is prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0078] Furthermore, the offset of the image center is calculated as follows: For any target pixel to be calculated, based on its known row and column coordinates, the offset is calculated by subtracting the coordinates of the image center; the offset in the row direction is equal to the row coordinates of the target pixel minus the row coordinates of the image center, and the offset in the column direction is equal to the column coordinates of the target pixel minus the column coordinates of the image center; assuming the target pixel is located at row coordinate 120 and column coordinate 480, the offset in the row direction is -120, indicating that the pixel is above the image center; the offset in the column direction is +160, indicating that the pixel is to the right of the image center. Finally, two values ​​with positive and negative signs are output, which accurately quantify the degree to which the target pixel deviates from the center point on the image plane, where the positive and negative signs indicate the direction of deviation, and the numerical value indicates the number of pixels that have deviated.

[0079] Furthermore, the process of obtaining the single-pixel field-of-view parameters of the thermal imager includes: When establishing a full-scale spatiotemporal reference for the system, the pixel field-of-view parameters of each thermal imager have already been obtained through laboratory calibration. To establish the conversion relationship between pixel offset and spatial angle offset, it is necessary to calculate the horizontal single-pixel field-of-view angle subtended by each pixel in the horizontal direction based on the total horizontal field-of-view angle and horizontal resolution of the thermal imager; simultaneously, based on the total vertical field-of-view angle and vertical resolution, the vertical single-pixel field-of-view angle subtended by each pixel in the vertical direction is calculated. Taking a thermal imager with a horizontal field-of-view angle of 40 degrees and a horizontal resolution of 640 pixels as an example, its horizontal single-pixel field-of-view angle is the total horizontal field-of-view angle divided by the number of effective pixels in the horizontal direction. Similarly, taking a thermal imager with a vertical field-of-view angle of 30 degrees and a vertical resolution of 480 pixels as an example, its vertical single-pixel field-of-view angle is the total vertical field-of-view angle divided by the number of effective pixels in the vertical direction. The horizontal and vertical single-pixel field-of-view angles are output, representing the spatial angle change corresponding to each pixel offset in the image.

[0080] Furthermore, the process of calculating the preliminary horizontal and vertical deflection angles using pixel offset and single-pixel field of view includes: multiplying the row offset by the vertical single-pixel field of view to obtain the preliminary deflection angle of the target pixel relative to the optical axis centerline in the vertical direction. The row offset is -120, and multiplying it by the vertical single-pixel field of view yields a negative angle value. This negative sign indicates that the deflection direction is upward, deviating from the optical axis. Multiplying the column offset by the horizontal single-pixel field of view to obtain the preliminary deflection angle of the target pixel relative to the optical axis centerline in the horizontal direction. The column offset is 160, and multiplying it by the horizontal single-pixel field of view yields a positive angle value. This positive sign indicates that the deflection direction is to the right, deviating from the optical axis. Finally, two angle values ​​with clear directional signs are output, which initially express the degree of horizontal and vertical deviation of the line of sight relative to the optical axis centerline of the thermal imager when looking from the optical center of the thermal imager towards the spatial point represented by the target pixel.

[0081] (6) Spatial convergence and coarse difference division are performed on the sparse temperature point cloud generated by all thermal imagers to eliminate contradictions and outliers among multi-source data. Coarse difference division refers to the process of eliminating obviously erroneous or unreliable outlier data points before fusing the temperature point cloud data generated by multiple thermal imagers.

[0082] It's important to understand that since 36 thermal imagers observe the same grain pile from different angles, the point clouds they generate all reside within the same defined overall coordinate system of the grain silo. Therefore, they naturally overlap and complement each other spatially. The server aggregates the data from all 36 sparse temperature point clouds, forming a massive point cloud dataset containing millions of four-dimensional data points. At this point, fine-tuning and coarse-grained differential division between the multi-source data are required. For the same small spatial area observed by different thermal imagers, such as any point on the surface of the grain pile, which may be simultaneously observed by two adjacent thermal imagers on the ceiling and a nearby probe thermal imager, theoretically their coordinates and temperature values ​​should be highly consistent. The server uses a nearest neighbor search algorithm to find point pairs from different thermal imagers that are extremely close in spatial distance. In this embodiment, a distance of less than 5 centimeters is defined as the observation of the same space. The system analyzes the regions and compares their temperature values. If the temperature difference is too large, exceeding a set threshold (e.g., 0.5℃), it indicates that there may be a significant measurement error at one of the points. This error may originate from changes in the surface emissivity of the grain with the observation angle or instantaneous infrared reflection interference. Instead of simply discarding these points, the system employs a confidence-based weighted processing strategy. Based on the observation angle of each thermal imager relative to the observation point, the distance between the thermal imager and the target, and the historical stability data of the thermal imager itself, a confidence weight value between 0.1 and 1.0 is assigned to each data point. The closer the observation angle is to vertical, the closer the distance, and the more stable the historical data of the thermal imager, the higher the confidence weight is assigned. The final output is a massive, spatially aligned four-dimensional point cloud dataset containing all the original data points, but with an additional confidence weight attribute added to each point.

[0083] (7) Perform three-dimensional spatial data fusion and voxel meshing interpolation to transform discrete point clouds into continuous and regular three-dimensional temperature field distribution data;

[0084] For example, the server first divides the entire grain silo space—a cubic region extending from X=0 to 30 meters, Y=0 to 20 meters, and Z=0 to 8 meters—into tiny cubes with sides of 10 centimeters. These tiny cubes are called voxels. The entire grain silo space is thus divided into 300 voxels in the X direction, 200 in the Y direction, and 80 in the Z direction, totaling 4.8 million regularly arranged voxels. Next, for the center point of each voxel, the server searches for all discrete data points in the output point cloud dataset within a preset radius (20 centimeters in this embodiment) around it. Then, the server uses inverse distance weighted interpolation to calculate the final temperature value of the voxel. The core idea of ​​this method is that the closer a known point is to the center of the voxel to be estimated, the greater its influence on the estimation result. Specifically, each known point searched for and participating in the interpolation is assigned a weight, which is its distance from the center of the voxel. The inverse of the square of the distance between the voxels is multiplied by the confidence weight assigned to the point in (6) and then combined. The temperature values ​​of all points are multiplied by their combined weights and summed, and then divided by the sum of all combined weights to obtain the temperature value of the voxel center. If no known points are found within a 20 cm radius around a voxel, the voxel is temporarily marked as null. The temperature is then filled smoothly using the voxel values ​​with calculated temperatures around the voxels through the three-dimensional spline interpolation algorithm, thus ensuring the integrity and continuity of the entire temperature field without any gaps. The final output is a complete three-dimensional temperature array of size 300×200×80. Each element in the array accurately represents the average temperature of each 10 cm side cube space region in the granary at the current moment. The inverse distance weight interpolation method and the three-dimensional spline interpolation algorithm are existing technologies in this field and are not the inventive solutions of this application. They will not be described in detail here.

[0085] (8) Encapsulate the three-dimensional temperature array into a structured data file as the standard data input for all pest feature analysis.

[0086] Based on the thermal radiation data of the grain pile, thermal feature analysis is performed to extract insect pest heat spot features and temperature difference gradient features, generating insect pest feature data, including:

[0087] S2.1: Based on the three-dimensional temperature field distribution data of the grain pile, a three-dimensional local maximum detection algorithm is used for traversal scanning. When the temperature value of any spatial point is detected to be higher than the average temperature of all points in its three-dimensional 26-neighborhood and the temperature difference exceeds the first preset threshold, the point is marked as a suspected hot spot seed point. Three-dimensional region growth is performed on the suspected hot spot seed points. Points with temperature differences within the second preset threshold are merged into the same connected region. Then, the geometric center coordinates, volume, surface area, and peak temperature of the connected region are calculated, and the geometric center coordinates, volume, surface area, and peak temperature are used as the characteristics of insect-infested hot spots.

[0088] Furthermore, in this embodiment, the entire temperature field is traversed, and the temperature difference between each point and its 26 neighbors is calculated. The 26 neighbors include adjacent points in the vertical, horizontal, front-back, and diagonal directions. Only when the temperature of a point is significantly higher than the average temperature of its surrounding microenvironment is it considered a seed. This effectively avoids false alarms caused by an overall increase in ambient temperature. During regional growth, the second preset threshold is set to 1.0℃, meaning that only points whose temperature difference from the seed point does not exceed 1℃ and are spatially connected will be included in the same hotspot. This ensures the accuracy of hotspot boundary extraction.

[0089] Furthermore, the specific steps of S2.1 include:

[0090] (1) Obtain three-dimensional temperature field distribution data of the grain pile;

[0091] (2) Before starting the traversal scan, two thresholds are preset for seed point determination and regional growth. The first is the first preset threshold, which is used to determine whether a point is hot enough to become a suspected hot spot seed point. This threshold represents the minimum difference between the temperature of the point and the temperature of the surrounding local environment. In this embodiment, based on the experimental determination of the metabolic heat production law of major grain storage pests such as corn weevil and grain borer at different densities, the first preset threshold is set to 2.5℃. The second is the second preset threshold, which is used to determine whether adjacent points belong to the same hot spot during regional growth. This threshold represents the maximum allowable range of temperature change inside the hot spot. In this embodiment, after statistical analysis of multiple typical pest hot spots, the second preset threshold is set to 1.0℃.

[0092] (3) Traverse the three-dimensional temperature field distribution data of the grain pile, calculate the difference between the average temperature of any point in the three-dimensional temperature field distribution data and its 26 three-dimensional neighbors, and generate a list of suspected hot spot seed points. Specifically, this includes: extracting the temperature value of each of these 26 neighboring points, calculating the arithmetic mean of these temperature values, obtaining the average temperature of the local environment around the point, and then subtracting this average temperature from the current point's temperature value to obtain a temperature difference value. If the temperature difference value is positive and greater than the first preset threshold, then the voxel point is identified as a suspected hot spot seed point and recorded. Record its three-dimensional coordinates; during the scanning process, assuming that a point with a temperature of 28.5℃ is detected at the position of the 120th grid in the X direction, the 48th grid in the Y direction, and the 21st grid in the Z direction, its three-dimensional 26-neighborhood average temperature is 23.8℃, and the temperature difference is 4.7℃, which is greater than the first preset threshold of 2.5℃, this point is marked as the first suspected hot spot seed point; continue scanning until the entire 300×200×80 array is traversed, find all points that meet the conditions, and finally output a list containing the three-dimensional coordinates of all suspected hot spot seed points, which is denoted as the seed point list;

[0093] (4) Select an unprocessed seed point from the seed point list and initialize a new hot spot connected region. Specifically, select the first unprocessed seed point from the generated seed point list as the starting point for the current region growth. Create a new data structure for the seed point to store the hot spot connected region to be formed. The data structure will record the coordinates and temperature values ​​of all voxel points belonging to the hot spot. In the initial state, the data structure only contains the current seed point itself. At the same time, mark the seed point as belonging and set its status to processed to prevent it from being reused in subsequent scanning or growth processes. Finally, output a newly created hot spot object containing only a single seed point. At the same time, the status of the seed point in the seed point list has been updated to processed.

[0094] (5) Starting from the current seed point, perform three-dimensional region growth, continuously absorbing adjacent points that meet the temperature difference conditions into the current hot spot connected region to generate a complete single hot spot;

[0095] Specifically, a dynamic list of growth front points is maintained. Initially, the list only contains the selected seed point (4). Then, a loop is entered, and a point is continuously taken from the list of growth front points. Each neighboring point in the three-dimensional 26-neighborhood of the point is checked. For each neighboring point, it is first determined whether it has been assigned to another hot spot or has been treated. If so, it is skipped. If not, the temperature value of the neighboring point is obtained and compared with the temperature value of the seed point of the current hot spot. The absolute value of the temperature difference between the two is calculated. If the absolute value of the temperature difference is less than or equal to the second preset threshold of 1.0℃, it is considered that the neighboring point belongs to the same pest activity area as the current hot spot. Therefore, it is added to the data structure of the current hot spot. At the same time, the neighboring point is marked as assigned and added to the list of growth front points so that it can continue to expand outward from it. Zhang; If the absolute value of the temperature difference is greater than the second preset threshold of 1.0℃, then the point is ignored. When all 26 neighborhoods of a point taken from the growth front point list have been checked and no new points have been added, the point is removed from the growth front point list. When the entire growth front point list becomes empty, it means that no new neighboring points that meet the temperature difference condition can be found. At this time, the region growth process terminates. In the example of Warehouse No. 5, growth starts with a seed point with a temperature of 28.5℃. The temperatures of the points around it are 27.9℃, 28.1℃, 27.8℃, etc., and the temperature difference with the seed point is less than 1.0℃. Therefore, they are continuously absorbed. The growth process continues to expand outward until it encounters a region with a temperature lower than 27.5℃ and then stops. A complete hot spot connected region containing the coordinates and temperature values ​​of several voxel points is output, which is recorded as a single hot spot object.

[0096] (6) Repeat (4)-(5) until all seed points in the seed point list have been processed, and generate a hot spot object list. Each element in the hot spot object list is an independent hot spot connected region. Each region contains a set of spatially connected voxel points with similar temperatures. In the entire three-dimensional temperature field of Warehouse 5, assuming that a total of three independent suspected hot spot seed points are detected, after region growth, three independent connected regions are finally formed. The output of this step is a list, denoted as the hot spot object list.

[0097] (7) For each hot spot connected region in the hot spot object list, calculate its geometric and physical characteristic parameters to generate insect pest hot spot characteristics.

[0098] Specifically, for each independent hotspot connected region in the hotspot object list, feature calculations are performed by traversing all voxel points contained within that independent hotspot connected region. First, the geometric center coordinates are calculated. The algorithm sums the X-coordinate values ​​of all voxel points within the region and divides by the total number of voxel points to obtain the X-coordinate of the geometric center. Similarly, the Y-coordinate values ​​are summed and divided by the total number to obtain the Y-coordinate of the geometric center. The Z-coordinate values ​​are summed and divided by the total number to obtain the Z-coordinate of the geometric center. Second, the volume is calculated. Since each voxel point represents a cube with a side length of ten centimeters, its volume is fixed at ten centimeters multiplied by ten centimeters, which is 0.00. The hotspot volume is 1 cubic meter, therefore, the total volume of voxels in the region is equal to the total number of voxels in the region multiplied by 0.001 cubic meters. Next, the surface area is calculated. The algorithm calculates the surface area by identifying the boundary voxels of the hotspot. If a voxel has at least one point in its 26-dimensional neighborhood that does not belong to the current hotspot, then the voxel is determined as a boundary voxel. The exposed faces of all boundary voxels are summed up. Each face is a square with a side length of 10 centimeters, i.e., an area of ​​0.01 square meters, so the total surface area of ​​the hotspot can be obtained. Finally, the peak temperature is determined. When traversing the temperature values ​​of all voxels in the region, the maximum value is recorded. This maximum value is the peak temperature of the hotspot. In this embodiment, the first hot spot contains 52 voxels. Its geometric center coordinates are calculated and converted to actual metric coordinates, i.e., X equals 12.03 meters, Y equals 4.81 meters, and Z equals 2.10 meters. Its volume is 0.052 cubic meters, its surface area is 1.24 square meters, and its peak temperature is 28.5℃. The final output is a complete set of insect pest hot spot characteristic parameters, including geometric center coordinates, volume, surface area, and peak temperature.

[0099] S2.2: Simultaneously, based on the three-dimensional temperature field distribution data of the grain pile, the Sobel operator is used to calculate the temperature gradient components of each spatial point in the horizontal, vertical and depth directions. The gradient vector magnitude of the point is synthesized according to the temperature gradient components, and the rate of change of the gradient vector magnitude and the consistency index of the gradient direction within a preset spatial scale are statistically analyzed to generate temperature difference gradient features. The Sobel operator is the prior art in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0100] Furthermore, the Sobel operator used in this embodiment contains three 3×3×3 three-dimensional convolution kernels to calculate the gradient in the horizontal direction. For example, its convolutional kernel's central layer, i.e., the second layer in the Z direction, has weight coefficients in the X direction of [-1, 0, 1] and their variations, used to detect changes in the X direction; the server combines the three-dimensional temperature field data volume with... Convolution with the kernel yields the gradient component matrix in the X direction; similarly, with... Kernel convolution yields the gradient component in the Y direction, and... Kernel convolution yields the gradient component in the Z direction. Then, at each spatial point, the three gradient components are squared, summed, and then squared to obtain the gradient vector magnitude. Finally, a gradient magnitude matrix is ​​generated. The larger the value on the gradient magnitude matrix, the more drastic the temperature change at that point, which is usually the edge of the hot spot. The gradient direction points to the direction of the fastest temperature increase. The gradient direction at all edge points should point to the center of the hot spot.

[0101] S2.3: The insect pest hot spot features and the temperature difference gradient features are concatenated to form a multi-dimensional feature vector describing the current insect pest activity status of the grain pile, i.e., insect pest feature data.

[0102] Furthermore, the feature concatenation of the insect pest hot spot features and the temperature difference gradient features involves normalizing the insect pest hot spot features and the temperature difference gradient features and then splicing them together end to end.

[0103] The pest characteristic data is input into a preset transfer learning model to analyze the pest reproduction pattern, and targeted control trigger commands are generated based on the analysis results, including:

[0104] S3.1: Load a pre-stored deep convolutional neural network model trained on a large dataset from the source domain as the base network; the base network includes at least an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer;

[0105] The source domain dataset comprises publicly available image or thermal imaging datasets of grain insects collected from grain depots in multiple different ecological zones. Specifically, the dataset contains 500,000 labeled images from eight major grain storage ecological zones, including Northeast, North, and Southwest China. These images cover visible light and infrared thermal imaging of twelve major grain storage pests, such as maize weevils, grain borers, sawtooth grain beetles, and Indian meal borers, at different developmental stages including adults, larvae, eggs, and pupae. The images were collected under various conditions, including different grain varieties, temperature and humidity environments, and lighting conditions, ensuring the breadth and diversity of the data sources.

[0106] The deep convolutional neural network model adopts the Densenet121 architecture. Its input layer receives normalized pest feature data. Its multiple densely connected blocks concatenate all feature maps of the previous layer along the channel dimension as input to the next layer. Its transition layer connects adjacent densely connected blocks to reduce the feature map size. Its pooling layer converts the feature map output by the last densely connected block into a feature vector. Its fully connected layer uses the Adam optimizer combined with the cross-entropy loss function for parameter fine-tuning. Its output layer uses the Softmax activation function to output the pest development stage classification result, where the pest development stage includes at least the egg stage, larval stage, pupal stage, and adult stage.

[0107] Furthermore, the Densenet121 architecture consists of 121 network layers. Its core feature is the use of a dense connection mechanism, where the input of each layer comes from the concatenation of the outputs of all previous layers along the channel dimension. This design can maximize the flow of features between layers, effectively alleviate the gradient vanishing problem, and achieve stronger feature representation capabilities with fewer parameters.

[0108] Furthermore, after determining the basic network architecture, a loading operation needs to be performed within the deep learning framework. In this embodiment, the system uses the PyTorch deep learning framework as the basic platform for model training and inference. The specific execution method of the loading operation is to call the pre-trained model interface provided in the torchvision model library, specify the model name as Densenet121, and set the pre-training parameters to true. This operation will automatically load the Densenet121 model, which has been pre-trained on the ImageNet large-scale image dataset, from the cloud or local cache. This model contains the complete network structure definition and the pre-trained weight parameters of all layers. After loading, the model is instantiated as a workable Python object containing a complete structure including an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer. The input layer is pre-defined as a 240×240 pixel three-channel color image, corresponding to the standardized input requirements of a large source domain dataset. The main body of the network consists of four densely connected blocks and three transition layers connecting them. Each densely connected block contains multiple convolutional layers: the first block contains 6 convolutional layers, the second contains 12, the third contains 24, and the fourth contains 16. Within each densely connected block, the feature map size remains constant, while the number of channels gradually increases with depth. The first layer is responsible for connecting adjacent densely connected blocks. Each transition layer contains a batch normalization layer, a convolutional layer, and an average pooling layer. Its function is to reduce the spatial size of the feature map by half and adjust the number of channels. After the last densely connected block, there is a global average pooling layer, which converts the feature map into a one-dimensional feature vector. Finally, there are two fully connected layers. The first fully connected layer maps the feature vector to 1024 dimensions, and the second fully connected layer maps the 1024-dimensional features to 1000 dimensions, corresponding to the 1000 classification categories of the ImageNet dataset. The output layer uses the Softmax activation function to convert the output of the fully connected layers into the predicted probability of each category. Finally, a complete Densenet121 model instance containing pre-trained weights is output, which is denoted as the original base network.

[0109] In this embodiment, the input layer of the deep convolutional neural network model receives normalized multidimensional pest feature data. This normalized multidimensional pest feature data is organized into a format similar to a three-dimensional image, with multiple densely connected blocks at its core. In each densely connected block, the input of each layer comes from the outputs of all preceding layers, and these outputs are concatenated along the channel dimension. For example, the third layer not only receives the output of the second layer but also receives the original output of the first layer as part of its input. This design greatly enhances the flow of features between layers, alleviates the gradient vanishing problem, and achieves better performance with fewer parameters. The transition layers connect the densely connected blocks through convolution and pooling. Operations are performed to reduce the size of the feature map; finally, a global average pooling layer compresses the 3D feature map output by the last densely connected block into a 1D feature vector; in the fine-tuning stage, by fixing the parameters of the first few densely connected blocks, only the last few densely connected blocks and the fully connected layer are trained. The optimizer is Adam, and the specific parameters of the Adam optimizer are set to a learning rate of 0.0001 and a loss function of cross-entropy loss. The final output layer uses the Softmax activation function, the batch size is set to 32, and the number of iterations is set to 50. The output of the fully connected layer is converted into a probability distribution, for example, 5% for the egg stage, 92% for the larval stage, 2% for the pupal stage, and 1% for the adult stage in this embodiment.

[0110] S3.2: Fix the network parameters of the first convolutional layer in the basic network, and fine-tune the network parameters of the second convolutional layer and the fully connected layer in the basic network using the historical pest feature data labeled in the target domain to form a transfer learning model adapted to the current grain depot scenario;

[0111] The fine-tuning process requires updating model parameters using labeled data in the target domain. In this embodiment, the historical pest feature data labeled in the target domain are pest feature data collected by an infrared thermal imager array and manually verified and labeled during the current grain depot's historical storage cycle. Specifically, a total of 5,000 labeled samples from the past three storage cycles are extracted from the grain depot's database. Each sample corresponds to a set of pest feature data vectors, which are 92 in length and include pest heat spot features and temperature gradient features. Each sample is also accompanied by a manually labeled pest development stage tag, with the tag value being an integer from zero to three, corresponding to the egg stage, larval stage, pupal stage, and adult stage, respectively. Before inputting the data into the model, preprocessing is required. First, the 5,000 labeled samples are randomly divided into training, validation, and test sets in a 7:2:1 ratio. The training set contains 3,500 samples, the validation set contains 1,000 samples, and the test set contains 500 samples. Then, each 92-dimensional feature vector is converted into the format desired by the model's input layer. Since the DenseNet121 input layer is originally designed to process two-dimensional image data, this embodiment requires reshaping the one-dimensional feature vector into a two-dimensional feature map. Specifically, the 92-dimensional feature vector is filled into an 11×11 two-dimensional matrix, with the first 10 rows each filled with 9 values, the last row filled with 2 values, and the remaining positions filled with zeros, forming an 11×11 single-channel feature map. Finally, the feature map is normalized to between zero and one by dividing each pixel value by the maximum value of all pixels in the sample. The final output is a target domain dataset that has been divided into training, validation, and test sets and has undergone format conversion and normalization.

[0112] S3.3: Input the real-time extracted pest feature data into the transfer learning model that has completed parameter fine-tuning. After layer-by-layer feature transformation and nonlinear mapping within the model, output the classification result of the current developmental stage of the pest population, the pest population density level, and the predicted value of the pest outbreak probability within a preset time period.

[0113] Furthermore, the specific steps of S3.3 include:

[0114] (1) Obtain pest feature data and perform preprocessing. Specifically, first, fill the 92-dimensional vector, i.e. pest feature data, into an 11×11 two-dimensional matrix to form a complete 11×11 single-channel feature map. After filling, normalize the single-channel feature map. After processing, the original 92-dimensional feature vector is converted into a single-channel two-dimensional feature map of size 11×11, which is denoted as the input feature map.

[0115] (2) The preprocessed input feature map is fed into the deployed transfer learning model. The input layer of the transfer learning model receives the input feature map and passes it to the first dense connection block. Since the model has been set to evaluation mode, all batch normalization layers use the global mean and variance obtained during the training phase. The parameters of all layers remain fixed. The entire propagation process does not involve gradient calculation and only performs forward numerical operations.

[0116] (3) General basic features are extracted after freezing the first convolutional layer in the first densely connected block. Specifically, the input feature map first enters the first densely connected block, which contains 6 convolutional layers. The parameters of these convolutional layers are completely frozen during the fine-tuning stage, so their kernel weights remain the values ​​obtained from pre-training on the source domain large dataset. The first convolutional layer contains 64 3×3 kernels, and the weight value of each kernel is fixed after pre-training. The input feature map performs two-dimensional convolution operations with these kernels. The stride is set to 1 and the padding is set to 1 to ensure that the spatial size of the output feature map remains unchanged at 11×11. After processing by the first convolutional layer, a feature map with 64 channels is generated. Subsequently, these 64-channel feature maps are processed by the next 5 convolutional layers in the first densely connected block. Each layer takes the concatenation of the outputs of all previous layers in the channel dimension as input. After all 6 convolutional layers of the first densely connected block, the number of channels of the output feature map increases to 256, while the spatial size remains 11×11. This is called the shallow feature map.

[0117] (4) The shallow feature map then enters the first transition layer. The first transition layer first performs batch normalization to standardize the shallow feature map. Then, it compresses the number of channels from 256 to 128 through a convolutional layer. Finally, it performs an average pooling operation with a stride of 2 to reduce the spatial size of the feature map from 11×11 to 5×5. Finally, it outputs a feature map with 128 channels and a spatial size of 5×5, which is called the first intermediate layer feature map.

[0118] (5) The first middle layer feature map then enters the second dense connection block, which contains 12 convolutional layers. During the fine-tuning stage, the parameters of the first 6 convolutional layers are frozen, and the parameters of the last 6 convolutional layers are updated using historical pest data of the target domain. The first middle layer feature map is processed by these 12 convolutional layers in sequence. Each layer takes the concatenation of the outputs of all the previous layers in the channel dimension as input. After all 12 convolutional layers of the second dense connection block, the number of channels of the output feature map increases to 512, and the spatial size remains unchanged at 5×5. This is called the first deep feature map.

[0119] (6) The first deep feature map then enters the second transition layer. The transition layer first performs batch normalization, then compresses the number of channels from 512 to 256 through a convolutional layer, and finally performs an average pooling operation with a stride of 2 to reduce the spatial size of the feature map from 5×5 to 2×2, which is called the second middle layer feature map.

[0120] (7) The second middle layer feature map enters the third dense connection block. The third dense connection block contains 24 convolutional layers. The parameters of these convolutional layers were all updated using historical pest data of the target domain during the fine-tuning stage. The second middle layer feature map is processed by 24 convolutional layers in sequence. Each layer takes the concatenation of the outputs of all previous layers in the channel dimension as input. After all 24 convolutional layers of the third dense connection block, the number of channels of the output feature map increases to 1024, and the spatial size remains unchanged at 2×2. This is called the second deep layer feature map.

[0121] (8) The second deep feature map then enters the third transition layer. The transition layer first performs batch normalization, then compresses the number of channels from 1024 to 512 through a convolutional layer, and finally performs an average pooling operation with a stride of 2 to reduce the spatial size of the feature map from 2×2 to 1×1, which is called the third middle layer feature map.

[0122] (9) The third middle layer feature map then enters the fourth dense connection block, which contains 16 convolutional layers. The third middle layer feature map is processed by 16 convolutional layers in sequence. Each layer takes the concatenation of the outputs of all previous layers in the channel dimension as input. After all 16 convolutional layers of the fourth dense connection block, the number of channels of the output feature map increases to 1024, and the spatial size remains unchanged at 1×1. Since the spatial size has been reduced to 1, this feature map is actually equivalent to a one-dimensional feature vector with a length of 1024, which is denoted as the third deep layer feature map.

[0123] (10) The third deep feature map is fed into the global average pooling layer. Since the spatial size of the third deep feature map is already 1×1 and each channel contains only one value, the global average pooling layer does not actually need to perform any averaging operation. It directly outputs these 1024 values ​​to form a one-dimensional feature vector with a length of 1024, which is called the pooling feature vector.

[0124] (11) The pooled feature vector is simultaneously fed into three parallel fully connected branches constructed during the fine-tuning stage. The first branch is used for pest development stage classification and contains two fully connected layers. The first fully connected layer maps the 1024-dimensional pooled feature vector to 256 dimensions and applies the ReLU activation function. The second fully connected layer maps the 256-dimensional feature vector to four dimensions, obtaining four raw score values. Then, the Softmax activation function is used to convert the four raw scores into four probability values ​​between zero and one, with a sum of one, corresponding to the predicted probabilities of the egg stage, larval stage, pupal stage, and adult stage, respectively. The second branch is used for pest population density level regression and contains two fully connected layers. The first fully connected layer maps the 1024-dimensional pooled feature vector to 128 dimensions and applies the ReLU activation function. The second fully connected layer maps the 128-dimensional feature vector to one dimension and outputs the result. A continuous numerical value, trained during the fine-tuning phase to directly correspond to the insect population density level; the third branch is used for predicting the probability of insect outbreak within a preset time period, also containing two fully connected layers. The first fully connected layer maps the 1024-dimensional pooled feature vector to 128 dimensions and applies the ReLU activation function. The second fully connected layer maps the 128 dimensions to one dimension, outputting a value between zero and one, which directly represents the predicted probability of insect outbreak within the preset time period. These three branches are executed in parallel, sharing the same pooled feature vector as input, and outputting three parallel inference results: a four-dimensional developmental stage classification probability vector, a continuous numerical insect population density regression value, and a predicted insect outbreak probability value between zero and one. In this embodiment, the preset time period is set to 7 days.

[0125] S3.4: Based on the developmental stage classification results, insect population density level, and predicted insect outbreak probability, a preset prevention and control rule library is matched. When the triggering condition is determined to be met, a targeted prevention and control triggering instruction containing the target area coordinates, prevention and control operation parameters, and execution time period is generated.

[0126] The matching preset prevention and control rule base, when the triggering condition is determined to be met, generates a targeted prevention and control triggering instruction containing the target area coordinates, prevention and control operation parameters, and execution time period, including:

[0127] S3.4.1: Query the prevention and control rule base to obtain the minimum insect population density trigger threshold corresponding to the classification result of the current insect development stage;

[0128] In this embodiment, the table structure of the prevention and control rule base includes a primary key identifier field, a pest development stage field, a minimum pest density trigger threshold field, a suggested operation mode field, a suggested pesticide type field, a suggested pesticide concentration field, a suggested pesticide dosage field, and a rule effective time range field. Among them, the pest development stage field stores text label values, which are completely consistent with the development stage text label format output by the model, including four values: egg stage, larval stage, pupal stage, and adult stage. The minimum pest density trigger threshold field stores an integer value, representing the minimum pest density level at which prevention and control measures need to be initiated under the corresponding development stage.

[0129] Furthermore, a database query statement is constructed using the developmental stage label as the key value, specifying the target field for the query. Since the pest developmental stage field is designed to have a unique constraint, meaning that each developmental stage corresponds to only one record in the rule base, the query statement is expected to return only one matching result. The query statement also needs to explicitly specify the target field to be returned. In this embodiment, the target field to be returned is the minimum pest population density trigger threshold field, meaning that only the value of this field needs to be extracted from the matching record, and it is not necessary to return all fields of the entire record. The output of this step is a structured database query instruction, denoted as the query to be executed.

[0130] Furthermore, a query operation needs to be performed to retrieve matching records from the prevention and control rule base and extract the target field value. In this embodiment, based on grain storage experience and historical data statistics, for larval pests, due to their large feeding volume, severe damage, and rapid growth period, the set minimum insect population density trigger threshold is relatively low. The value stored in this field is secondary. The query engine extracts this value and returns it as the query result. At the same time, to ensure the accuracy and reliability of the query results, the query engine will also return a status flag indicating that the query was executed successfully and the result is valid. The final output is a query result set containing the minimum insect population density trigger threshold value, denoted as the trigger threshold query result. This result set explicitly contains the secondary value.

[0131] S3.4.2: Convert the insect population density level into an insect population density value, and determine whether the insect population density value is greater than or equal to the minimum insect population density trigger threshold;

[0132] Furthermore, the process of converting the insect population density level into an insect population density value includes:

[0133] (1) Obtaining the insect population density level. Specifically, the insect population density level is obtained by rounding the continuous values ​​output by the regression branch by the transfer learning model output module. The value is an integer value, and the range is predefined as level 1 to level 5. Level 1 corresponds to 1 to 5 insects per kilogram of grain, level 2 corresponds to 5 to 15 insects per kilogram of grain, level 3 corresponds to 15 to 30 insects per kilogram of grain, level 4 corresponds to 30 to 50 insects per kilogram of grain, and level 5 corresponds to 50 insects per kilogram of grain. In this embodiment, the insect population density level output by the transfer learning model is level 3, which is recorded as the current level value.

[0134] (2) Locate and access the insect population density level reference table pre-stored in the edge computing server. The insect population density level reference table is a structured data mapping table pre-built and stored in the local memory of the edge computing server. The construction of this structured data mapping table is based on the insect pest density grading standard in the field of grain storage. It establishes a one-to-one correspondence between the abstract level concept and the specific insect population density numerical range. It is organized using a key-value pair data structure, where the key is the level value and the value is a description of the corresponding insect population density numerical range. In this embodiment, the specific content of the insect population density level reference table is as follows: Level 1 corresponds to a lower limit of 1 insect and an upper limit of 5 insects; Level 2 corresponds to a lower limit of 5 insects and an upper limit of 15 insects; Level 3 corresponds to a lower limit of 15 insects and an upper limit of 30 insects; Level 4 corresponds to a lower limit of 30 insects and an upper limit of 50 insects; and Level 5 corresponds to a lower limit of 50 insects and an upper limit of 80 insects based on historical data statistics. In addition, the insect population density level comparison table predefines representative values ​​for each level. Typically, the median of the value range for that level is selected as the conversion target, i.e., 3 insects for level 1, 10 insects for level 2, 22.5 insects for level 3, 40 insects for level 4, and 65 insects for level 5. The output confirms that the insect population density level comparison table has been loaded into memory and can be accessed quickly, and is recorded as a usable comparison table.

[0135] (3) Construct a lookup command to retrieve the conversion result that matches the value of the insect population density level from the insect population density level comparison table. The core logic of the lookup command is to use the current level value of level three as the key to find the corresponding value in the comparison table. In this embodiment, what needs to be obtained is the representative value corresponding to level three, that is, the median of the level, 22.5 insects.

[0136] (4) Perform a table lookup operation, retrieve matching records from the insect population density level comparison table and extract representative values. In this embodiment, the output is a table lookup result set containing the representative value 22.5, which is recorded as the insect population density value query result.

[0137] (5) After obtaining the query result of the insect population density value, perform validity verification. First, check the status flag returned by the query to confirm that the table lookup was executed successfully and no memory access error occurred. Then check whether the returned value is a valid floating-point number. Confirm that 22.5 is a reasonable value within the preset range. In this embodiment, the value range corresponding to the third level is 15 to 30 insects. 22.5 falls within this range and meets the definition of the median. Therefore, the verification is passed. Finally, confirm that the returned value format is a floating-point number type and can be directly used for numerical comparison calculation.

[0138] (6) After obtaining the current insect population density value of 22.5, the system compares it with the value corresponding to the lowest insect population density trigger threshold obtained from the prevention and control rule base. Before this, the trigger threshold level 2 needs to be converted into the corresponding representative value. Repeat (2)-(5) to look up the table with level 2 as the key value and obtain the representative value corresponding to level 2 as 10 insects. Then, the value comparison operation is performed to determine whether the current insect population density value of 22.5 is greater than or equal to the value corresponding to the trigger threshold of 10 insects. The result of the comparison operation is true, that is, 22.5>10. At the same time, the system records the entire process of this conversion and comparison, including the current level value, the value obtained from the table lookup, the trigger threshold level, the value corresponding to the trigger threshold, and the comparison result, forming an audit log.

[0139] S3.4.3: If the judgment result is yes, then further obtain the control operation parameter template that is jointly associated with the current pest development stage classification result and the pest outbreak probability prediction value within the preset time period; the control operation parameter template includes at least the pesticide type, pesticide concentration, pesticide dosage and operation mode;

[0140] S3.4.4: Determine the coordinates of the target area based on the geometric center coordinates of the pest hot spot features, fill in the prevention and control operation parameters according to the prevention and control operation parameter template, determine the execution period based on the pest outbreak probability prediction value, and assemble and generate a targeted prevention and control trigger command.

[0141] Furthermore, the geometric center coordinates in the pest hotspot features represent the core location of the pest activity area. However, in actual control operations, it is necessary to target an area with a certain spatial range rather than a single point. Therefore, based on these geometric center coordinates and combined with the size information of the pest hotspot, a target spatial range that can cover the entire pest activity area needs to be calculated. This includes: firstly, obtaining the volume and shape information of the hotspot from the pest hotspot features. In this embodiment, the volume of the hotspot is 0.052 cubic meters, and its shape is approximately an ellipsoid. Based on the volume and shape, the hotspot is estimated to cover an area of ​​approximately 100 square kilometers. The extension radii in the X, Y, and Z directions are approximately 0.3 meters, 0.2 meters, and 0.2 meters, respectively. Based on this, the target area is defined as an ellipsoidal space with the geometric center coordinates as its center and the radius as the extension radius in each direction. To facilitate path planning and drug application control for the targeted control equipment, this ellipsoidal space needs to be converted into a regular cuboid bounding box. The conversion rule is to extend 0.3 meters to the left and right in the X direction, 0.2 meters forward and backward in the Y direction, and 0.2 meters upward and downward in the Z direction, forming a cuboid region. Calculations show that the X coordinate range of this cuboid region is 11.73 meters to 12.33 meters, the Y coordinate range is 4.61 meters to 5.01 meters, and the Z coordinate range is 1.90 meters to 2.30 meters. Meanwhile, in order to ensure that the nozzle can accurately reach the working position, the geometric center coordinates need to be converted into a path planning target point that the equipment can recognize, that is, the three-dimensional coordinate point that the end of the robotic arm needs to reach. This point directly uses the original geometric center coordinates of 12.03 meters, 4.81 meters, and 2.10 meters. The final output is a complete target area description, including a cuboid bounding box for pesticide coverage and a central target point for equipment positioning, which is denoted as the final target area coordinates.

[0142] Furthermore, filling in the prevention and control operation parameters according to the prevention and control operation parameter template involves filling and formatting the parameters according to each field in the prevention and control operation parameter template to generate specific operation control parameters. For example, the value of the treatment agent concentration field given in the prevention and control operation parameter template is 2%, which needs to be converted into the control parameter of the equipment proportional valve. The opening degree of the proportional valve controls the mixing ratio of the mother liquor and the dilution water to achieve the target concentration.

[0143] Furthermore, the predicted pest outbreak probability reflects the urgency of pest development, which is used to determine the optimal timing for control operations. The system first obtains the current system time, assuming it is 2:30 AM on July 15, 2024. Then, according to preset rules, the predicted pest outbreak probability of 0.85 is mapped to different execution time strategies. The control rule base defines three levels: when the outbreak probability is below 0.3, the execution time is set to the next regular operation window, i.e., 8 PM to 10 PM that day; when the outbreak probability is between 0.3 and 0.7, the execution time is set to execute as soon as possible, i.e., the first idle operation period after the current time, such as 4 AM; when the outbreak probability is above 0.7, the execution time is set to execute immediately. Since 0.85 > 0.7 in this example, the immediate execution strategy is triggered. Immediate execution means that once the instruction is generated, it should be immediately sent to the equipment and the operation process should begin, without waiting. To clearly express this strategy in the instruction, the execution time field is filled with a special identifier, along with the current timestamp as the instruction generation time. In addition, considering that nighttime operations may require consideration of factors such as personnel on duty and equipment status, the system will also check whether the current time is within the preset workable time window. In this embodiment, the workable time window set by the grain depot is 24 hours a day, so the immediate execution strategy is effective.

[0144] The response to receiving the targeted control trigger command, controlling the targeted control equipment to perform precise control operations on the grain pile area corresponding to the insect infestation hotspots, includes:

[0145] S4.1: Parse the targeted prevention and control trigger command to obtain the target area coordinates and prevention and control operation parameters;

[0146] S4.2: Based on the coordinates of the target area, plan the movement path of the targeted control equipment in the grain warehouse, and calculate the attitude angle required for the end effector of the equipment to reach the target position;

[0147] Furthermore, the specific steps of S4.2 include:

[0148] (1) Analyze the targeted prevention and control trigger command to obtain the coordinates of the target area. At the same time, obtain the current position of the mobile chassis, the current angle of each joint of the robotic arm, and the three-dimensional position information of the fixed obstacles in the grain warehouse from the sensor and status monitoring system of the targeted prevention and control equipment in real time. The three-dimensional position information of the fixed obstacles includes the position and size of fixed facilities such as warehouse columns, ventilation ducts, and temperature measuring cable suspension points.

[0149] (2) Based on the current position of the mobile chassis, the coordinates of the target area are transformed from the overall coordinate system of the grain warehouse to the coordinate system of the robotic arm base, and the working posture required by the end effector at the target point is determined, so as to obtain the end target position and target posture in the coordinate system of the robotic arm base. The two together constitute the target posture of the robotic arm.

[0150] Furthermore, the complete three-dimensional coordinates of the robotic arm base in the overall coordinate system of the grain silo need to be calculated through the orbital geometric model. The orbital geometric model is existing technology in this field and is not an inventive solution of this application, so it will not be described in detail here.

[0151] Furthermore, the coordinates of the target point relative to the robotic arm base are the overall coordinates of the grain silo minus the coordinates of the base.

[0152] In this embodiment, the desired orientation for point-like precision injection is that the nozzle is perpendicular to the grain pile surface and pointing downwards, that is, the Z-axis of the end effector points directly downwards.

[0153] (3) Based on the target position in the target pose of the robotic arm, and combined with the positive kinematic workspace model of the robotic arm, calculate the target stopping position of the mobile chassis as it moves along the track; the target stopping position must ensure that when the chassis stops at this time, the target position is within the workspace that the robotic arm can reach from its base.

[0154] (4) Plan the track movement path of the mobile chassis from the current position to the target parking position. When planning the path, load the three-dimensional position information of the fixed obstacle at the same time and perform collision detection. If it is detected that the mobile chassis or the robotic arm it carries will interfere with the obstacle in the default posture, specify the corresponding obstacle avoidance posture sequence for the robotic arm in the path planning and generate a chassis movement path containing segmented speed commands and robotic arm obstacle avoidance posture requirements.

[0155] (5) When the mobile chassis arrives at and locks at the target parking position, the updated target pose of the robotic arm is used as input, and the inverse kinematics model of the robotic arm is used to solve the problem and calculate a set of target angles of each joint of the robotic arm that enable the end effector to reach the target pose. The inverse kinematics model of the robotic arm is the prior art in this field and is not an inventive solution of this application. It will not be described in detail here.

[0156] (6) Starting from the current angle of each joint of the robotic arm and ending at the target angle of each joint of the robotic arm, trajectory interpolation is performed in the joint space to generate a time-continuous joint angle motion trajectory. During the generation of the joint angle motion trajectory, real-time collision detection and joint limit verification are performed on the discrete intermediate points on the trajectory to ensure that the entire trajectory is collision-free and that the angle of each joint is always within the range allowed by the design of the robotic arm.

[0157] (7) The chassis movement path and the joint angle movement trajectory are spatiotemporally aligned and integrated to generate a complete time-sequential equipment control instruction sequence; the equipment control instruction sequence includes the chassis movement speed instruction at each time period, the joint angle instruction of the robotic arm at each time point, and the operation preparation instruction of the end effector after reaching the target pose.

[0158] S4.3: At the start of the execution period, a movement control command is sent to the target control device to drive it to move along the planned path and adjust its attitude to the target position;

[0159] S4.4: Once the targeted control equipment is in place, a drug application control command is generated based on the control operation parameters to control the coordinated operation of the drug pipeline, proportional valve, and atomizing nozzle of the targeted control equipment to perform precise drug application to the target area. The control operation parameters include the operation mode, drug identification, drug concentration, and application rate.

[0160] Furthermore, the operation mode selection logic: When generating instructions, the system intelligently selects the operation mode. In this embodiment, since the hot spot volume is small, such as 0.5 cubic meters, and exists in isolation, it meets the conditions for the point-like precision injection mode. The preset logic in the prevention and control rule base is:

[0161] If the area of ​​the hot spot in the insect pest hot spot characteristics is less than a preset area threshold, such as 2 square meters, and the number of hot spots is less than a preset number threshold, such as 3, then the point-like precision injection mode is selected. In the point-like precision injection mode, the robot will control a single atomizing nozzle to open and adjust the nozzle to a low flow rate, such as 10-15 mL / s, and accurately aim the atomizing nozzle at the geometric center of each hot spot to perform point-to-point and quantitative injection of pesticide. The pesticide directly acts on the core area of ​​the pest, resulting in high control efficiency and minimal pesticide usage. It is suitable for the control of isolated and small-scale insect pest hot spots.

[0162] If a hot spot is detected to be distributed in a strip along any direction of the grain pile, such as along the cold and hot interface of the silo wall, the system will select the strip continuous spraying mode. In the strip continuous spraying mode, the robot will control the two atomizing nozzles to turn on and adjust the nozzles to a medium flow rate, such as 20-25 mL / s. According to the extension direction and length of the strip hot spot, the robot will plan the linear movement path of the robotic arm and drive the robotic arm to move at a constant speed along the path to achieve continuous and uniform spraying of the strip hot spot area. This is suitable for the control of medium-sized insect pest hot spots that are linearly distributed.

[0163] If the area of ​​the hot spots in the pest hot spot characteristics is greater than the preset area threshold or the number of hot spots is greater than the preset number threshold and the distribution is scattered, the area full coverage spraying mode is triggered. In the area full coverage spraying mode, the robot will control all three atomizing nozzles to be turned on and adjust the nozzles to a high flow rate, such as 30-35 mL / s. According to the distribution range of the pest hot spots, the robot plans the grid-like movement path of the robotic arm and drives the robotic arm to move at a constant speed along the path to achieve full coverage and uniform spraying of the entire pest distribution area. It is suitable for the control of large-scale pest hot spots with large areas and scattered distribution.

[0164] Furthermore, the robots work collaboratively: upon receiving the pesticide application control command, the track-mounted mobile pesticide application robot does not simply move in a straight line. Based on the location of equipment inside the warehouse, such as axial flow fans and temperature measuring cables, it dynamically plans the optimal obstacle avoidance path. After reaching the target area, the descent speed of the robotic arm is adjusted in real time according to the pressure feedback from the nozzle to prevent damage to the nozzle or compaction of the grain due to excessive descent. After the pesticide application begins, the flow rate and atomization angle of the nozzle are controlled in a closed loop according to the application rate and operation mode to ensure uniform coverage of the pesticide solution without dripping or dead zones.

[0165] Through the detailed description of the above embodiments, this invention fully demonstrates how to integrate technologies such as infrared thermal imaging, three-dimensional temperature field reconstruction, machine learning, transfer learning, and automated control to construct a complete, intelligent, and precise grain depot pest control system, encompassing the entire chain from perception to analysis to execution. This system can automatically inspect 24 / 7 without human intervention, accurately locating and eliminating pests at the initial stage of an outbreak, even in the larval stage invisible to the naked eye. This changes the traditional extensive management model of grain depots, which relies on manual experience, regular inspections, and full-warehouse pesticide application, thereby improving the safety, economy, and environmental friendliness of grain storage.

[0166] Example 2:

[0167] Please see Figure 3 Another embodiment of the present invention provides: an intelligent pest control system for grain depots based on infrared thermal feature sensing, comprising:

[0168] The data acquisition module 10 simultaneously collects the original thermal radiation intensity values ​​of the surface and different depth layers of the grain pile through the first infrared thermal imager group arranged in a grid on the roof of the grain silo and the second infrared thermal imager group on the detection rod inside the grain pile. After temperature conversion, spatial coordinate registration and data fusion, it generates three-dimensional temperature field distribution data that can completely characterize the temperature spatial distribution of the entire grain pile.

[0169] The insect pest thermal feature analysis module 20, based on the three-dimensional temperature field distribution data of the grain pile, extracts the core hot spot features such as the geometric center coordinates and volume of the insect pest hot spots through the three-dimensional local maximum detection algorithm and the region growing algorithm to obtain the insect pest hot spot features. At the same time, it uses the three-directional Sobel operator to perform three-dimensional convolution operation, calculates and statistically obtains the temperature difference gradient features, and finally concatenates the insect pest hot spot features and temperature difference gradient features into multi-dimensional insect pest feature data to realize the quantitative characterization of the insect pest activity state in the grain pile.

[0170] The transfer learning model analysis module 30 first uses a deep convolutional neural network model based on Densenet121 as the base network, and combines the source domain large dataset and the historical pest feature data labeled in the target domain grain depot to complete the model fine-tuning, and build a transfer learning model adapted to the current grain depot scenario. Then, real-time pest feature data is input into the transfer learning model, and through feature transformation and nonlinear mapping, the classification results of pest development stage, pest density level and pest outbreak probability prediction value within the preset time period are accurately output.

[0171] The targeted prevention and control instruction generation module 40, based on the analysis results of the transfer learning model, matches the preset prevention and control rule library, first determines whether the insect population density value has reached the minimum insect population density trigger threshold. If it has, it retrieves the associated prevention and control operation parameter template, determines the target area coordinates by combining the geometric center coordinates in the insect hot spot features, selects the execution time period based on the predicted insect outbreak probability value, and finally assembles and generates a standardized targeted prevention and control trigger instruction containing the target area coordinates, prevention and control operation parameters, and execution time period.

[0172] The targeted control equipment execution module 50 receives and parses the targeted control trigger command, extracts the control operation parameters, first plans the movement path of the track-type mobile spraying robot and calculates the target posture angle of the robotic arm, drives the equipment to move and adjust its posture to the target position during the specified execution period; then, according to the operation mode, pesticide parameters, etc., it controls the pesticide pipeline, proportional valve and atomizing nozzle of the equipment to work together to achieve directional, quantitative and fixed-mode precise application of pesticide liquid, and completes the targeted control operation on the grain pile area corresponding to the insect infestation hot spots;

[0173] The prevention and control rule base and parameter management module 60 is used to store and manage various rules, thresholds and parameter templates required for the entire process of prevention and control. These include the minimum insect population density trigger threshold corresponding to different insect development stages, the operation mode selection threshold corresponding to insect hot spot characteristics, and prevention and control operation parameter templates such as pesticide type and pesticide concentration associated with insect development stage and outbreak probability. It also supports flexible adjustment and updating of relevant parameters according to the actual scenario of the grain depot and the type of insect pest, providing standardized and adaptable rules and parameters for the targeted prevention and control instruction generation module 40.

[0174] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for intelligent pest control in grain depots based on infrared thermal feature sensing, characterized in that, include: Acquire thermal radiation data of the grain pile collected by an array of infrared thermal imagers inside the warehouse; Based on the heat radiation data of the grain pile, thermal feature analysis is performed to extract the characteristics of insect pest hot spots and temperature difference gradient features, and insect pest feature data is generated. The pest characteristic data is input into a preset transfer learning model to analyze the pest reproduction pattern, and a targeted control trigger command is generated based on the analysis results. In response to receiving the targeted control trigger command, the targeted control equipment is controlled to perform control operations on the grain pile area corresponding to the insect infestation hot spots. The acquisition of thermal radiation data of the grain pile collected by the infrared thermal imager array inside the warehouse includes: A first infrared thermal imager group is set up along a preset grid below the grain silo roof to collect thermal radiation from the surface of the grain pile. At the same time, at least one detection rod is set up vertically at a preset depth inside the grain pile. A second infrared thermal imager group is set up at equal intervals along the direction of the rod on each detection rod to collect thermal radiation from different depth layers inside the grain pile. According to the preset synchronous scanning cycle, the first infrared thermal imager group and the second infrared thermal imager group are controlled to synchronously collect the raw thermal radiation intensity values. The raw thermal radiation intensity values ​​collected by each infrared thermal imager are converted into temperature values. Based on the three-dimensional spatial coordinates and pixel field of view parameters of each infrared thermal imager in the grain warehouse, spatial coordinate registration and data fusion are performed on all temperature values ​​to generate three-dimensional temperature field distribution data of the grain pile that characterizes the temperature spatial distribution of the entire grain pile, which serves as the grain pile thermal radiation data. Based on the thermal radiation data of the grain pile, thermal feature analysis is performed to extract insect pest heat spot features and temperature difference gradient features, generating insect pest feature data, including: Based on the three-dimensional temperature field distribution data of the grain pile, a three-dimensional local maximum detection algorithm is used for traversal scanning. When the temperature value of any spatial point is detected to be higher than the average temperature of all points in its three-dimensional 26-neighborhood and the temperature difference exceeds the first preset threshold, the corresponding point is marked as a suspected hot spot seed point. Three-dimensional region growth is performed on the suspected hot spot seed points. Points with temperature differences within the second preset threshold are merged into the same connected region. Then, the geometric center coordinates, volume, surface area, and peak temperature of the connected region are calculated, and the geometric center coordinates, volume, surface area, and peak temperature are used as the characteristics of insect-infested hot spots. Simultaneously, based on the three-dimensional temperature field distribution data of the grain pile, the Sobel operator is used to calculate the temperature gradient components of each spatial point in the horizontal, vertical and depth directions. The gradient vector magnitude of the corresponding point is synthesized according to the temperature gradient components, and the rate of change of the gradient vector magnitude and the consistency index of the gradient direction within the preset spatial scale are statistically analyzed to generate temperature difference gradient features. The insect pest hot spot features and the temperature difference gradient features are concatenated to form a multi-dimensional feature vector describing the current insect pest activity status of the grain pile, i.e., insect pest feature data.

2. The intelligent pest control method for grain depots based on infrared thermal feature sensing as described in claim 1, characterized in that, The pest characteristic data is input into a preset transfer learning model to analyze the pest reproduction pattern, and targeted control trigger commands are generated based on the analysis results, including: A pre-stored deep convolutional neural network model trained on a large dataset from the source domain is loaded as the base network; the base network includes at least an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer; The network parameters of the first convolutional layer in the basic network are fixed, and the network parameters of the second convolutional layer and the fully connected layer in the basic network are fine-tuned using the historical pest feature data labeled in the target domain to form a transfer learning model adapted to the current grain depot scenario. The real-time extracted pest feature data is input into the transfer learning model with fine-tuned parameters. After layer-by-layer feature transformation and nonlinear mapping within the model, the classification result of the current developmental stage of the pest population, the pest population density level, and the predicted value of the pest outbreak probability within a preset time period are output. Based on the developmental stage classification results, insect population density levels, and predicted insect outbreak probability, a preset prevention and control rule library is matched. When the triggering conditions are determined to be met, a targeted prevention and control triggering instruction containing the target area coordinates, prevention and control operation parameters, and execution time period is generated.

3. The intelligent pest control method for grain depots based on infrared thermal feature sensing as described in claim 2, characterized in that, The deep convolutional neural network model adopts the Densenet121 architecture. Its input layer receives normalized pest feature data. Its multiple densely connected blocks use the concatenation of all feature maps from the previous layer along the channel dimension as the input to the next layer. Its transition layer connects adjacent densely connected blocks. Its pooling layer converts the feature map output by the last densely connected block into a feature vector. Its fully connected layer uses the Adam optimizer combined with the cross-entropy loss function for parameter fine-tuning. Its output layer uses the Softmax activation function to output the pest development stage classification results, where the pest development stage includes at least the egg stage, larval stage, pupal stage, and adult stage.

4. The intelligent pest control method for grain depots based on infrared thermal feature sensing as described in claim 3, characterized in that, The source domain big data set is a publicly available grain insect image dataset or thermal imaging dataset collected from grain depots in multiple different ecological zones; the historical pest feature data labeled in the target domain is data collected and labeled by infrared thermal imager arrays during the historical storage period of the current grain depot, showing the actual occurrence and reproduction stages of pests. The general feature representation capabilities learned in the source domain are transferred to the specified scenario of the current grain depot through transfer learning.

5. The intelligent pest control method for grain depots based on infrared thermal feature sensing as described in claim 4, characterized in that, The matching preset prevention and control rule base, when determining that the triggering condition is met, generates a targeted prevention and control triggering instruction containing the target area coordinates, prevention and control operation parameters, and execution time period, including: Query the control rule base to obtain the minimum insect population density trigger threshold corresponding to the classification result of the current insect development stage; The insect population density level is converted into an insect population density value, and it is determined whether the insect population density value is greater than or equal to the minimum insect population density trigger threshold. If the judgment result is yes, then obtain the control operation parameter template that is jointly associated with the current pest development stage classification result and the pest outbreak probability prediction value within the preset time period; the control operation parameter template includes at least the pesticide type, pesticide concentration, pesticide dosage and operation mode. The coordinates of the target area are determined based on the geometric center coordinates of the pest hot spot features. The control operation parameters are filled in according to the control operation parameter template. The execution period is determined according to the pest outbreak probability prediction value. The targeted control trigger command is then assembled and generated.

6. The intelligent pest control method for grain depots based on infrared thermal feature sensing as described in claim 1, characterized in that, The step of responding to receiving the targeted control trigger command and controlling the targeted control equipment to perform control operations on the grain pile area corresponding to the insect infestation hot spots includes: The targeted prevention and control trigger command is parsed to obtain the target area coordinates and prevention and control operation parameters; Based on the coordinates of the target area, plan the movement path of the targeted control equipment in the grain warehouse, and calculate the attitude angle of the end effector of the equipment when it reaches the target position; At the start of the execution period, a movement control command is sent to the targeted control device to drive it to move along the planned path and adjust its posture to the target position; Once the targeted control equipment is in place, a pesticide application control command is generated based on the control operation parameters. This command controls the pesticide pipeline, proportional valve, and atomizing nozzle of the targeted control equipment to work together to apply pesticides to the target area.

7. The intelligent pest control method for grain depots based on infrared thermal feature sensing as described in claim 6, characterized in that, The targeted control device is a track-mounted mobile spraying robot, whose end effector is a multi-degree-of-freedom robotic arm, with at least one independently controlled atomizing nozzle mounted at the end of the robotic arm. The specific method for generating spraying control instructions is as follows: select the corresponding pesticide pipeline according to the pesticide identification in the control operation parameters; control the opening degree of the ratio valve between the mother liquor and dilution water according to the pesticide concentration in the control operation parameters; and control the opening and closing sequence and flow rate of the atomizing nozzle according to the spraying rate and operation mode in the control operation parameters.

8. A grain depot pest intelligent control system based on infrared thermal feature sensing, used to implement the grain depot pest intelligent control method based on infrared thermal feature sensing as described in any one of claims 1-7, characterized in that, include: The data acquisition module synchronously collects the original thermal radiation intensity values ​​of the surface layer and different depth layers of the grain pile. After temperature conversion, spatial coordinate registration and data fusion, it generates three-dimensional temperature field distribution data of the grain pile. The insect pest thermal feature analysis module extracts insect pest hot spot features and temperature difference gradient features based on the three-dimensional temperature field distribution data of the grain pile, and concatenates the insect pest hot spot features and temperature difference gradient features into insect pest feature data. The transfer learning model analysis module is used to build a transfer learning model adapted to the current grain depot scenario. Real-time pest feature data is input into the transfer learning model, and after feature transformation and nonlinear mapping, the model outputs the classification results of pest development stage, pest population density level, and the predicted value of pest outbreak probability within a preset time period. The targeted prevention and control instruction generation module, based on the analysis results of the transfer learning model, matches the preset prevention and control rule library and assembles and generates targeted prevention and control trigger instructions containing target area coordinates, prevention and control operation parameters, and execution time period; The targeted prevention and control equipment execution module receives and parses the targeted prevention and control trigger command, extracts the prevention and control operation parameters, drives the equipment to move and adjust its posture to the target position during the specified execution period, and controls the equipment's agent pipeline, proportional valve, and atomizing nozzle to work together. The prevention and control rule base and parameter management module is used to store and manage various rules, thresholds and parameter templates for the entire prevention and control process.

Images