Tea garden pest multi-source sensing monitoring and intelligent decision-making pesticide application system

Abstract

The application discloses a tea garden pest multi-source sensing monitoring and intelligent decision-making pesticide application system, and relates to the technical field of intelligent agriculture. The system comprises a mobile sensing array, a static monitoring base station and a time synchronization module. By collecting multispectral images and environmental parameters and completing space-time alignment, the system realizes multi-modal feature deep fusion relying on the cross-attention mechanism, combines space-time sequence repair technology and sheltering bottom-up strategy to deal with target sheltering problems, generates pesticide application action vectors based on a deep reinforcement learning algorithm, and integrates mechanical response characteristics and physical constraint filtering to ensure stable execution. Meanwhile, the system constructs a pesticide application amount residual feedback closed loop to realize online optimization of the model. The application effectively improves the accuracy of pest identification and pesticide application and the stability of equipment operation, solves the problems of poor environmental adaptability, insufficient sheltering response and disconnection between decision-making and execution in the prior art, and realizes intelligent and accurate prevention and control of tea garden pests.

Description

Technical Field

[0001] This invention relates to the field of intelligent agriculture technology, and in particular to a multi-source sensor monitoring and intelligent decision-making system for pesticide application in tea gardens. Background Technology

[0002] In the background technology, existing intelligent pesticide application technologies for tea gardens have gradually introduced multi-sensor perception and intelligent algorithm decision-making to improve the accuracy and efficiency of pesticide application. In the data acquisition stage, some solutions combine image sensors and environmental sensors to obtain tea garden information, but these are mostly independent applications or simple splicing of single-type data, failing to achieve deep collaboration of different modalities. In the target recognition and pesticide application control stage, although intelligent algorithms are used for pest and disease identification and pesticide application decisions, effective continuous control strategies are not designed for scenarios such as overlapping and shading of tea garden leaves. Often, when perception is interrupted, pesticide application is directly stopped or the original action is maintained, lacking a flexible fallback mechanism. Regarding the connection between decision-making and execution, existing technologies typically design algorithmic decision-making and mechanical execution separately, failing to fully consider the response characteristics of the plant protection machine's mechanical system, pursuing only optimal decision-making at the algorithmic level while neglecting the stability requirements during physical execution. Furthermore, most systems lack dynamic optimization capabilities after deployment, failing to establish a closed-loop data feedback system throughout the entire process, making it difficult to cope with the accuracy degradation caused by hardware wear and tear during long-term use.

[0003] Although existing technologies have made some progress in the intelligent application of pesticides to tea gardens, many technical problems still need to be solved: First, due to the lack of a deep integration mechanism between environmental parameters and image features, existing systems are easily affected by factors such as light and humidity in complex tea garden environments, resulting in insufficient environmental adaptability and reliability in pest and disease identification. Second, for scenarios such as overlapping and shading of leaves, there is a lack of effective time-series repair and fallback strategies, which can easily lead to gaps in the pesticide application control process, affecting the integrity and accuracy of pesticide application. Third, intelligent decision-making has not fully integrated the response characteristics of mechanical systems, and the execution process has not been designed with physical constraints, resulting in a disconnect between the scientific nature of decision-making and the stability of execution, which can easily lead to problems such as insufficient pesticide application accuracy or equipment impact damage. Fourth, due to the lack of a full-process data closed loop and online fine-tuning mechanism, the system cannot dynamically compensate for deviations caused by hardware wear and tear, making it difficult to guarantee the stability of long-term pesticide application. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a multi-source sensor monitoring and intelligent decision-making system for pest and disease control in tea gardens, so as to solve one or more problems in the prior art.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A multi-source sensor monitoring and intelligent decision-making system for pest and disease control in tea gardens includes the following steps:

[0007] Step 1: Acquire multimodal sensing data of the tea garden in real time. The multimodal sensing data includes multispectral image features with a frequency range of 25-35Hz and environmental sensing parameters with a frequency range of 5-15Hz.

[0008] Step 2: Perform feature fusion on the multimodal sensing data, map the environmental sensing parameters into environmental embedding vectors, and use a cross-attention mechanism to adjust the gain of multispectral image features. When the ambient light intensity is within the threshold range of 800-1200 lux, the confidence compensation logic is triggered.

[0009] Step 3: Perform spatiotemporal sequence restoration on the identified target, when the detection probability P of the identified target... det When the value is in the range of 0.3-0.5, the hidden state vector is used in conjunction with the spatial displacement ΔS for interpolation prediction.

[0010] Step 4: Generate drug administration decision action vector A based on deep reinforcement learning algorithm t During the generation process, a mechanical system response hysteresis parameter t with a preset range of 120-180ms is introduced. delay And according to the travel speed v of the plant protection machine current Dynamically adjust the decision-making and prediction time;

[0011] Step 5: Analyze the drug administration decision-making action vector A t Perform physical constraint filtering and output the result to the actuator.

[0012] Specifically, the feature fusion described in step two employs a multi-head cross-attention mechanism, with the number of heads h being 4, 6, 8, 10, or 12. The feature extraction and fusion logic is as follows:

[0013] Visual feature extraction vector is The environment embedding vector is Let the number of heads in the cross-attention mechanism be... First, the query matrix, key matrix, and value matrix are decomposed into h independent subsets through linear transformation: the query matrix of the i-th head. ,in Let be the query linear transformation matrix of the i-th head; and be the key matrix of the i-th head. ,in Let be the linear transformation matrix of the key for the i-th head; and be the value matrix of the i-th head. ,in Let be the linear transformation matrix for the value of the i-th head; where , ,satisfy i=1,2,...,h,d kLet be the feature dimension of the i-th head; Single-head attention calculation: , Multi-head attention concatenation and final transformation: The outputs of h heads are concatenated and then subjected to a linear transformation to obtain the final fused feature. The complete formula is as follows: ;in, To output a linear transformation matrix for multi-head attention, This represents a vector concatenation operation; the query matrix Q is derived from the context embedding vector E. e The mapping is generated by the key matrix K and value matrix V from the visual feature vector F. v Mapping generation.

[0014] Specifically, the interpolation prediction logic in step three is as follows: when the target is occluded for 400-600ms, the hidden state vector output by the Bi-LSTM model is used to maintain the continuity of the drug administration control logic; when the target is occluded for >600ms or <400ms, the interpolation prediction is stopped and a decision pause command is issued.

[0015] Specifically, the definition of the drug administration decision action vector in step four is as follows: Drug administration decision action vector formula: ; where v target p represents the target travel speed of the agricultural machinery. t Indicates the pressure of drug application, ω t This indicates the solenoid valve pulse frequency; when the travel speed v current When the speed is >1.5m / s, the generation and prediction time of the drug application decision vector increases by 40-60ms.

[0016] Specifically, the logic of the physical constraint filtering process in step five is as follows: Control vector smoothing constraint formula: By limiting the absolute value of the first derivative of the control vector to no more than 0.2, the physical impact of pressure and flow switching of the actuator is eliminated.

[0017] Specifically, it also includes step six: real-time monitoring of the actual dosage Q. real With model prediction Q pred The residual ε is calculated such that when |ε|>5% and the duration reaches 3 sampling periods, the terminal bias parameters of the policy network are fine-tuned using online gradient descent, where the learning rate η for fine-tuning is 1×10⁻⁶. -5 .

[0018] Specifically, it also includes a mobile sensor array deployed on the plant protection machine, the mobile sensor array including a 4-channel multispectral camera; a static monitoring base station deployed in the blind area of ​​the tea garden, the static monitoring base station including a homogeneous 4-channel multispectral camera, temperature and humidity sensor and soil parameter sensor; and a time synchronization module used to perform spatiotemporal alignment between the mobile sensor array and the static monitoring base station.

[0019] Furthermore, the deployment rules for the mobile sensor array are as follows: in the case of young tea gardens, the multispectral camera is installed at a height of 50-80cm above the tea tree canopy; in the case of mature tea gardens, the multispectral camera is installed at a height of 80-120cm above the tea tree canopy; and the lens of the multispectral camera is tilted downwards at 30° relative to the vertical direction.

[0020] Furthermore, the technical parameters of the time synchronization module are as follows: it adopts the RTK-GPS time synchronization protocol, and the timestamp synchronization accuracy between the mobile sensor array and the static monitoring base station is ±1ms.

[0021] Furthermore, the deployment density of the static monitoring base stations is 3333.33m. 2 Deploy a group, and perform data sampling on the static monitoring base station once every 15 minutes.

[0022] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0023] (i) By combining a multimodal stereo perception network with a cross-attention fusion mechanism, the limitations of existing technologies that rely on single perception or simple data stitching are overcome, achieving deep synergy between environmental parameters and multispectral image features. This combination can dynamically adjust the weights of different types of data in complex tea garden environments, effectively offsetting the interference of environmental factors such as light and humidity on pest and disease identification, and significantly improving the environmental adaptability and reliability of the identification results.

[0024] (ii) By combining spatiotemporal sequence restoration technology with an occlusion fallback mechanism, the problem of perception discontinuity that easily occurs in existing technologies under leaf overlap occlusion scenarios is solved. This combination uses a time-series prediction model to maintain the continuity of the pesticide application logic during short-term occlusion, while using a fallback strategy of static base station data verification and decision pause to avoid the risk of mis-spraying caused by long-term occlusion, thus ensuring the integrity and accuracy of the pesticide application control process.

[0025] (III) By combining deep reinforcement learning decision-making with physical constraint filtering, the response characteristics of the mechanical system are integrated into the intelligent decision-making process, which is different from the design of existing technologies where algorithmic decision-making and physical execution are disconnected. This combination not only improves the accuracy of drug delivery through logic such as mechanical hysteresis compensation and speed adaptive prediction, but also eliminates the physical impact caused by frequent equipment switching through control vector derivative constraints, thus achieving a unity of scientific decision-making and execution stability.

[0026] (iv) By combining online deviation fine-tuning technology with a closed-loop data system throughout the entire process, the system's adaptive optimization capability is constructed, making up for the shortcomings of existing technologies in adapting to hardware wear. This combination utilizes the residual feedback of pesticide application to adjust model parameters in real time, combined with incremental optimization of the model supported by long-term data retention, enabling the system to dynamically compensate for accuracy deviations caused by hardware wear such as nozzle wear, and maintain stable pesticide application results over the long term. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the logical flow of monitoring and intelligent decision-making for drug administration in this invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and exemplary descriptions. It should be understood that the structures, proportions, sizes, etc., illustrated in the accompanying drawings are merely for illustrative purposes to aid those skilled in the art and are not intended to limit the implementation of this invention. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to the size, without affecting the effects and objectives achieved by this invention, should still fall within the scope of the technical content disclosed in this invention.

[0029] Application Overview

[0030] This invention discloses a multi-source sensing monitoring and intelligent decision-making system for pesticide application in tea gardens. Its core features, distinguishing it from existing technologies, are as follows: First, it innovatively employs a "multimodal three-dimensional sensing network + cross-attention fusion mechanism," rather than simply splicing multi-source data. Instead, it dynamically adjusts data weights through deep collaboration between environmental parameters and multispectral image features to resist complex environmental interference. Second, for tea garden leaf overlap and shading scenarios, it designs a "spatiotemporal sequence repair technology + shading fallback strategy." This maintains pesticide application continuity during short-term shading through temporal prediction, and combines static base station verification and decision-pause to avoid long-term shading. The invention addresses several key aspects: First, it mitigates the risk of accidental spraying and ensures the integrity of the control process. Second, it organically combines deep reinforcement learning decision-making with physical constraint filtering, integrating the mechanical response characteristics of the plant protection machine into the decision-making process. Mechanical hysteresis compensation and adaptive speed prediction improve application accuracy, while control vector derivative constraints eliminate physical impacts on the equipment, achieving synergistic optimization of decision-making and execution. Third, it constructs an adaptive optimization system of "online deviation fine-tuning + full-process data closed loop," using residual feedback of application rate to adjust model parameters in real time. Combined with incremental optimization supported by long-term data, it dynamically compensates for accuracy deviations caused by hardware wear, maintaining stable application results over the long term. Through the innovative combination of these technical features, this invention effectively solves the shortcomings of existing technologies in environmental adaptability, shading response, decision-making and execution coordination, and long-term stability, significantly improving the intelligence and precision of pest and disease application in tea gardens.

[0031] Comprehensive explanation Detailed Implementation

[0032] This invention discloses a multi-source sensor monitoring and intelligent decision-making system for pest and disease control in tea gardens. By organically combining multimodal perception, deep fusion, intelligent decision-making, and closed-loop optimization, it achieves accurate monitoring and efficient application of pesticides for pests and diseases in tea gardens, and solves problems such as interference identification, occlusion response, decision-making and execution coordination, and long-term stability in complex environments. The specific implementation of the system is described in detail below.

[0033] I. System Overall Architecture and Deployment

[0034] The core architecture of this system is built around the entire process of "perception-processing-decision-execution-optimization," encompassing a multimodal three-dimensional perception network, a communication and timing module, a data processing module, an intelligent decision-making module, an execution control module, and a maintenance and optimization module. These modules work collaboratively to ensure the comprehensiveness of data acquisition, the accuracy of processing, the scientific nature of decision-making, and the stability of execution, enabling those skilled in the art to fully implement the system's functions based on the details described below.

[0035] (I) Deployment of Multimodal Stereo Sensing Network

[0036] The three-dimensional sensing network consists of a mobile sensor array and a static monitoring base station, which work together to achieve data collection across the entire tea garden without blind spots.

[0037] The mobile sensor array is deployed on the agricultural drone, with the installation height adjusted according to the tea tree's growth stage: 50-80cm from the canopy for young tea trees; 80-120cm from the canopy for mature tea trees, with the lens tilted downwards at 30° relative to the vertical direction. The array integrates a 4-channel multispectral camera (including RGB, NIR, and Red-edge bands), with a video sampling frequency of 25-35Hz, and simultaneously acquires environmental sensor parameters (temperature, humidity, light intensity, etc.) at a frequency of 5-15Hz.

[0038] Static monitoring base stations are deployed in tea garden blind spots (such as steep slopes and gullies) where plant protection machines cannot access, with a deployment density of 3333.33m. 2 One group (i.e., 5 acres / group) integrates a 4-channel multispectral camera, temperature and humidity sensor and soil parameter sensor that are derived from the same source as the mobile sensor array. The sampling frequency is 15 minutes / time, which is used for blind spot data compensation and verification in occlusion scenarios.

[0039] (ii) Communication transmission and time synchronization

[0040] The communication transmission adopts a layered design: multispectral images and environmental parameters collected by the mobile sensor array and the static monitoring base station are transmitted with high bandwidth through 4G or 5G industrial modules; the drug application decision command is issued through the MQTT protocol to ensure the reliability of command transmission.

[0041] The time synchronization relies on the RTK-GPS time synchronization protocol. The time stamp synchronization accuracy between the mobile sensor array and the static monitoring base station is controlled within ±1ms, realizing the spatiotemporal alignment of multi-source data and providing a unified time reference for subsequent data fusion and decision-making.

[0042] II. Multimodal Data Processing and Feature Fusion

[0043] The collected multimodal data needs to undergo feature extraction and deep fusion to be transformed into high-dimensional feature vectors that can be used for decision-making. The core is to achieve synergistic enhancement of environmental parameters and visual features to resist interference from complex environments.

[0044] (I) Feature Extraction

[0045] Visual Feature Extraction: A Backbone network is used to process images acquired by a multispectral camera, extracting image feature vectors at a resolution of 1280×720. These vectors have a dimension of 480 and are mathematically expressed as follows: Its core function is to capture the visual morphological characteristics of pests and diseases (such as color, texture, outline, etc.).

[0046] Environmental feature embedding: Environmental sensing parameters such as temperature, humidity, and light intensity are nonlinearly mapped through three fully connected layers to generate an environmental embedding vector with the same dimension as the visual feature vector. Mathematically, this is expressed as... This vector is used to characterize the real-time environmental state of the tea garden, providing a basis for subsequent feature weight adjustment.

[0047] (ii) Cross-attention fusion

[0048] A multi-head cross-attention mechanism is employed to achieve deep fusion of visual and environmental features, with the number of heads h being 4, 6, 8, 10, or 12, rather than simple data concatenation. Its core logic is based on the environmental embedding vector E. e As the query matrix Q, with the visual feature vector F v As the key matrix K and value matrix V, the fused features after gain are calculated using the following formula, which is called the "Cross-Attention Feature Fusion Formula": Let the number of heads in the cross-attention mechanism be... First, the query matrix, key matrix, and value matrix are decomposed into h independent subsets through linear transformation: the query matrix of the i-th head. ,in Let be the query linear transformation matrix of the i-th head; and be the key matrix of the i-th head. ,in Let be the linear transformation matrix of the key for the i-th head; and be the value matrix of the i-th head. ,in Let be the linear transformation matrix for the value of the i-th head; where , ,satisfy i=1,2,...,h,d k Let be the feature dimension of the i-th head; Single-head attention calculation: Multi-head attention splicing and final transformation: ;in, To output a linear transformation matrix for multi-head attention, This represents a vector concatenation operation; the query matrix Q is derived from the context embedding vector E. e The mapping is generated by the key matrix K and value matrix V from the visual feature vector F. v Mapping generation. During the fusion process, the system dynamically adjusts feature weights based on real-time environmental conditions: when the ambient light intensity is within the 800-1200 lux threshold range, the environmental embedding vector E is automatically increased. e Attention weights are applied, and environmental prior logic is used to suppress visual noise interference, thereby improving feature reliability under complex lighting conditions. Through a cross-attention feature fusion formula, synergistic optimization of environmental parameters and visual features is achieved, solving the problem of single features being susceptible to environmental interference.

[0049] To ensure alignment of matrix operations, the total feature dimension D must be an integer multiple of the number of heads h. In the preferred embodiment of this scheme, D=480 and h is an even number, ensuring that the dimensions of each subspace are positive integers.

[0050] III. Spatiotemporal sequence repair and occlusion fallback strategy (core invention point)

[0051] To address the issue of target shading caused by overlapping tea leaves, the system introduces a Bi-LSTM time series prediction model, combined with a fallback mechanism to ensure the continuity and accuracy of pesticide application control logic.

[0052] Feature compensation logic: When the target detection probability of the current frame is in the range of 0.3-0.5, the system calls the hidden state vector h output by the Bi-LSTM model. t-1 By combining the spatial displacement $\DeltaS$ provided by the inertial measurement unit (IMU) for linear interpolation prediction, the location and status of pests and diseases can still be accurately estimated when visual perception is temporarily interrupted.

[0053] Occlusion classification and fallback mechanism: When the occlusion time threshold is set to 400-600ms, the continuity of the pesticide application control logic is maintained by relying on the above timing prediction to avoid equipment shock caused by frequent start and stop of solenoid valves; when the occlusion time is >600ms or <400ms, the system automatically triggers multispectral data from nearby static monitoring base stations for verification. If the target still cannot be locked, a decision pause command is issued. After the visual perception system re-establishes stable tracking, the pesticide application process is resumed, effectively avoiding the risk of blind spraying caused by long-term occlusion.

[0054] IV. Deep Reinforcement Learning Intelligent Decision Making (Core Invention Point)

[0055] The system builds an intelligent decision-making module based on the PPO algorithm, realizing end-to-end mapping from fused features to control actions, while incorporating physical system characteristics and constraints to ensure the accuracy of decision-making and the stability of execution.

[0056] (I) Decision Model Construction and Physical Delay Compensation

[0057] End-to-end mapping: The input to the decision model is high-dimensional features fused by cross-attention, and the output is a continuous control action vector. The mathematical expression of this vector is the drug administration decision action vector formula. , where v target p represents the target travel speed of the agricultural machinery. t Indicates the pressure of drug application, ω t This indicates the pulse frequency of the solenoid valve.

[0058] Physical delay compensation: Introducing the mechanical system response hysteresis parameter t into the reward function of the decision-making model. delay(Typical range is 120-180ms) The physical response characteristics of the plant protection machine's spraying mechanism are clearly considered to avoid spraying deviations caused by the disconnect between decision-making and execution.

[0059] (II) Adaptive velocity prediction and physical constraint filtering

[0060] Speed ​​adaptive prediction: When the plant protection machine travels at speed v current When the speed is >1.5m / s, the decision logic automatically increases the prediction time of the application command by 40-60ms to compensate for the positional deviation caused by the air drift of the agent under high-speed movement, and ensures that the agent is accurately sprayed to the target area.

[0061] Physical constraint filtering: To eliminate the physical shock during the pressure and flow switching process of the actuator, a physical constraint filter is applied before the actuator to limit the absolute value of the first derivative of the control vector to no more than 0.2. The mathematical expression of this constraint is the control vector smoothing constraint formula. This allows for a smooth switching between pressure and flow, protecting equipment and improving the uniformity of pesticide application.

[0062] V. Closed-loop optimization and execution control (core invention point)

[0063] The system constructs a closed-loop data system throughout the entire process, dynamically compensates for hardware wear through online deviation fine-tuning, and maintains the stability of application accuracy over the long term.

[0064] Drug administration execution: Control action vectors output by the decision module The signal is transmitted to the actuator, which drives components such as solenoid valves and pressure pumps to work together to achieve precise control of application pressure, flow rate, and spraying range.

[0065] Online deviation fine-tuning: Real-time monitoring of actual application rate Q real With model prediction Q pred The residual ε, when |ε|>5% for 3 consecutive sampling periods, triggers the online gradient descent algorithm, with a 1×10 -5 The learning rate is fine-tuned to adjust the terminal bias parameters of the network, automatically compensating for application rate deviations caused by nozzle wear.

[0066] Data closed-loop storage: Local industrial-grade servers (i7 level) retain 180 days of perception data, decision logs and drug administration records. This data serves as the training set for quarterly incremental fine-tuning of the model (iteration 50 epochs), continuously optimizing the adaptability and accuracy of the decision model.

[0067] VI. System Maintenance Specifications

[0068] To ensure long-term stable operation of the system, the following maintenance guidelines must be followed:

[0069] Optical calibration: Monthly radiometric calibration of the multimodal camera based on a 99% reflectivity standard white board is performed to ensure consistency and accuracy of image acquisition.

[0070] Hardware maintenance: Clean the sensor lens monthly to prevent outdoor dew and dust from affecting sensing accuracy; calibrate the solenoid valve response speed quarterly and update the mechanical system response hysteresis parameter t. delay This ensures the accuracy of physical delay compensation.

[0071] Model optimization: The decision model is incrementally fine-tuned every quarter using 180 days of locally stored data, iterating for 50 epochs to ensure that the model continuously adapts to the occurrence patterns of pests and diseases in tea gardens and environmental changes, maintaining the stability of long-term pesticide application effects.

[0072] This implementation method, through the organic combination and collaborative work of the above modules, comprehensively covers the entire process details of the system from deployment, data collection, processing, decision-making, execution to maintenance. All technical means and parameters are clear and explicit, and those skilled in the art can implement the entire technical solution without ambiguity based on the above description, effectively solving many shortcomings of the existing technology.

[0073] To further verify the practical application effect of the core technology combination of this system and clarify the impact of key technical variables on the overall system performance, the following comparative experiment was designed. The experiment was conducted strictly according to objectively existing independent testing standards. By controlling the core variables and keeping other operating parameters consistent, the system performance under different variable combinations was tested to demonstrate that the constraints on the relevant technical variables have clear practical significance.

[0074] I. Testing Methods

[0075] (a) Accuracy test for disease and pest identification

[0076] 1. Sample Preparation: Typical pest and disease samples were collected from the tea garden (covering 8 common pests and diseases including tea anthracnose, tea white spot disease, tea geometrid moth, and tea green leafhopper, as well as healthy tea leaves). 200 samples were collected for each type, for a total of 1800 samples. All samples were photographed using a high-definition camera under natural light conditions, with a uniform resolution of 1280×720. The shooting distance was consistent with the lens distance used in actual system operation (50-120cm). Samples were precisely labeled manually using professional annotation tools. The annotations included the coordinate frame of the pest / disease area (accurate to pixels) and the pest / disease category. The labeling accuracy was cross-validated by three professionals to ensure no labeling errors.

[0077] 2. Test environment setup: The actual lighting scene of the tea garden was simulated in a constant temperature and humidity laboratory (temperature 25±2℃, humidity 60±5%RH). The light intensity was adjusted to the range of 800-1200 lux using an adjustable light source device (covering the typical operating environment of the system). The angle of the light source was consistent with the downward tilt angle of the system lens (30°).

[0078] 3. Testing process: The labeled samples are input into the system of each experimental group in random order. The system completes the image acquisition, feature extraction, fusion and recognition process according to the preset parameters, and records the recognition category and predicted region coordinates of each sample.

[0079] 4. Accuracy calculation: Identification accuracy = (Number of correctly identified samples / Total number of test samples) × 100%. Among them, "correct identification" must meet two conditions at the same time: First, the disease and pest categories are consistent; second, the intersection-union ratio (IOU) between the predicted area and the manually labeled area is ≥ 0.5 (IOU = overlapping area between the predicted area and the labeled area / union area between the predicted area and the labeled area).

[0080] (ii) Testing the accuracy of pesticide application

[0081] 1. Simulated Tea Garden Construction: Construct a 1:1 simulated tea garden of 10m×10m and plant tea seedlings at the same density as the actual tea garden (0.5m spacing between plants and 1.5m spacing between rows). Pre-determine 20 target areas in the canopy of the tea seedlings (each area is a circle with a diameter of 10cm, marked with red reflective markers that are approximately the same size as the actual pests and diseases). The target areas are randomly distributed in different locations in the canopy (including the upper layer, middle layer, and overlapping leaf areas).

[0082] 2. Application parameter settings: The baseline values ​​for the plant protection machine's travel speed in all experimental groups were set to 1.5 m / s, the baseline value for application pressure was set to 0.3 MPa, the baseline value for the solenoid valve pulse frequency was set to 10 Hz, and the baseline value for spraying rate was set to 50 mL / m. 2 Only the core variables were adjusted according to the experimental group settings.

[0083] 3. Testing Procedure: The plant protection machine completes the spraying operation in the simulated tea garden according to the preset path (S-shaped path, path spacing 1.5m). After the operation is completed, a 4K high-definition camera is used to take pictures of each target area from an angle perpendicular to the canopy. The shooting distance is fixed at 1.5m to ensure that the image clearly shows the spraying traces.

[0084] 4. Accuracy Calculation: The spray coverage area within the target region is extracted using image segmentation technology, and the area of ​​the covered area is measured relative to the area of ​​the target region (78.5 cm²). 2 The accuracy of pesticide application is calculated as (covered area / target area) × 100%. If the sprayed area exceeds the target area, the excess area is not included in the effective coverage area.

[0085] (III) Equipment Operation Stability Test

[0086] 1. Test environment setup: An outdoor simulated tea garden work area (500m²) was constructed. 2 The system includes flat areas, gentle slope areas with a gradient of 15°, and simulated blind spots for machinery. The ambient temperature is controlled between 15-35℃, and the humidity is 40-85%RH. The power supply uses the plant protection machine's matching lithium battery (voltage 24V, capacity 100Ah) to ensure stable power supply.

[0087] 2. Test Procedure: Each experimental group's system ran continuously for 8 hours (480 minutes), executing the "data acquisition - feature fusion - decision-based drug administration - status feedback" process in a loop according to the actual operation mode. The system operation status data was recorded every minute, including the response time of the identification module, the delay in issuing decision instructions, the feedback signal of the solenoid valve action, and the communication connection status.

[0088] 3. Fault Judgment and Statistics: Faults are classified as follows: response time of the identification module exceeding 500ms, decision command delay exceeding 100ms, no feedback from the solenoid valve, communication interruption lasting more than 10s, or application pressure / flow rate deviating from the baseline value by ±10%. The identification module response time threshold (≤500ms) is set based on the multispectral image acquisition frequency (25-35Hz) and the computational efficiency of cross-attention feature fusion to ensure real-time adaptation between identification and decision-making. The number of faults and the duration of each fault are statistically analyzed during the test. The cumulative fault-free runtime = total test duration - cumulative fault duration.

[0089] 4. Stability calculation: Equipment operation stability = (cumulative fault-free operation time / total test time) × 100%.

[0090] II. Experimental Variables and Experimental Group Design

[0091] (a) Experimental variables

[0092] Three key variables were selected from the core technologies of the system: the number of cross-attention mechanism heads (variable 1), the occlusion time threshold for spatiotemporal sequence repair (variable 2), and the increase in speed adaptive prediction time (variable 3).

[0093] (II) Experimental group setup

[0094] A total of 10 experimental groups were designed. Except for the three variables mentioned above, all other operating parameters (including sensing sampling frequency, communication protocol, maintenance specifications, and application pressure baseline value) were kept consistent across all experimental groups. The specific groupings are as follows:

[0095] 1. Standard Group (Groups 1-5): All three variables are within the specified range, and the complete technical solution of this system is adopted;

[0096] 2. Control group (groups 6-9): The same technical framework as the control group was used, but at least one variable was outside the specified range;

[0097] 3. Blank control group (Group 10): Using existing technology (single multispectral perception + traditional PID decision-making mode, without cross-attention fusion, spatiotemporal sequence repair and speed adaptive prediction function).

[0098] (III) Weighted scoring mechanism

[0099] 1. Performance index weight setting: Taking into account the impact of the three performance indices on the overall system effect, the weight allocation is set as follows: pest and disease identification accuracy (40%), pesticide application accuracy (40%), and equipment operation stability (20%).

[0100] 2. Calculation method for individual indicators:

[0101] Pest and disease identification accuracy (P1) = (Number of correctly identified samples / Total number of test samples) × 100%;

[0102] Application accuracy (P2) = (Coverage area within the target area / Area of ​​the target area) × 100%;

[0103] Equipment operational stability (P3) = (cumulative fault-free running time / total test time) × 100%.

[0104] 3. System overall performance scoring formula:

[0105]

[0106] In this calculation, P1, P2, and P3 are all substituted into the formula as percentages (e.g., if P1 = 92.45%, then substitute 92.45), and the overall score is rounded to two decimal places.

[0107] III. Recording Experimental Results

[0108] Table 1. Relevant Data Information of the Experimental Group

[0109] Experimental group number Cross-attention mechanism head count Occlusion time threshold (ms) Increase in predicted time (ms) Pest and disease identification accuracy rate (%) Application accuracy (%) Equipment operational stability (%) Overall score (points) 1 4 400 40 88.23 86.57 92.35 88.39 2 6 450 45 90.17 89.32 94.18 90.63 3 8 500 50 92.45 91.68 96.24 92.90 4 10 550 55 91.36 90.85 95.76 92.04 5 12 600 60 89.78 88.93 93.87 90.26 6 3 500 50 83.56 81.29 88.42 83.62 7 14 500 50 84.19 82.37 87.95 84.21 8 8 350 50 85.32 83.74 89.16 85.46 9 8 650 50 86.05 84.51 88.73 85.97 10 - - - 72.38 70.65 80.24 73.26

[0110] IV. Summary of Experimental Results

[0111] The experimental results clearly show that the overall scores of the conventional groups (groups 1-5) were all above 90 points, significantly better than the control groups (groups 6-9, overall scores 83.62-85.97 points) and the blank control group (group 10, overall score 73.26 points). In all four control groups, at least one of the three indicators—pest and disease identification accuracy, pesticide application precision, and equipment operational stability—was significantly better than the blank control group. The highest overall score was in group 3 (92.90 points), placing it in the middle of the conventional groups, indicating that the overall performance of the system is not linearly related to a single variable, but rather is the result of the synergistic effect of multiple technical variables. These results fully demonstrate that the constraints on the relevant technical variables in this system have clear practical significance, and that the overall technical solution has significant advantages over existing technologies.

[0112] Based on the above experimental results, and combined with the working principle of the system's core algorithm and the inherent logic of variable setting, the following analysis will examine the performance trends and underlying causes reflected in the experimental data from the perspectives of algorithm collaboration mechanism and variable adaptability, revealing the correlation between variable setting and the overall system performance.

[0113] The reason why the conventional group exhibits superior overall performance is that its three key variables are all within the optimal adaptation range of the core algorithm, achieving synergistic effects among multiple algorithm modules. From an algorithmic logic perspective, the core value of the cross-attention mechanism lies in the precise weight allocation of multimodal features. The number of heads directly affects the granularity and computational efficiency of attention allocation: when the number of heads is within a reasonable range, a balance can be achieved between feature dimension and computational complexity, which can fully capture the correlation information between environmental parameters and visual features, achieving deep fusion of the two, without causing response delay or overfitting due to computational redundancy; however, when the number of heads exceeds this range, too few heads will lead to overly coarse attention allocation granularity, making it impossible to accurately identify effective feature associations in complex environments, while too many heads will introduce redundant computation and noise information, destroying the stability of feature fusion, and thus affecting the accuracy of recognition and decision-making.

[0114] The setting of the occlusion time threshold is highly related to the working mechanism of the Bi-LSTM time series prediction model. Essentially, this threshold is an adaptation parameter between the time series repair window and the occlusion scenario. Within a reasonable range, the threshold can accurately distinguish between short-term and long-term occlusion: for short-term occlusion, the Bi-LSTM model can accurately predict the target location and state based on historical hidden state vectors and spatial displacement, maintaining the continuity of drug administration control; for long-term occlusion, the fallback strategy of static base station data verification and decision-making pause can effectively avoid the risk of blind drug administration. If the threshold exceeds a reasonable range, a threshold that is too short will misclassify some short-term occlusions as long-term occlusions, leading to frequent triggering of pause commands and disrupting the continuity of the drug administration process; a threshold that is too long will misclassify long-term occlusions as short-term occlusions, and the error in time series prediction will accumulate over time, resulting in a significant decrease in drug administration accuracy and increasing the probability of ineffective drug administration.

[0115] The setting of the speed-adaptive prediction time increment is essentially a coordinated adaptation between the decision-making model and the mechanical system's response characteristics. This variable complements the mechanical system's response hysteresis parameter, accurately compensating for the airborne drift effect of pesticides at high speeds within a reasonable range: when the agricultural drone's speed reaches a specific threshold, the increase in prediction time precisely matches the pesticide's flight time from the nozzle to the target area with the mechanical system's response delay, ensuring accurate pesticide coverage of the target area. The variable settings in the conventional group perfectly align with this coordinated logic, while the control group, due to the absence of this adaptation setting or variables exceeding the adaptation range, experiences a time lag between the decision command and the actual execution, thus affecting the accuracy of pesticide application.

[0116] The middle group in the standard group exhibited the best performance, the underlying reason being the optimal synergy of the three variables in this group across multiple core algorithms: the feature fusion effect of the cross-attention mechanism, the temporal sequence restoration accuracy of Bi-LSTM, and the predictive compensation efficiency of the decision model mutually supported each other. This avoided the performance limitations of any single variable and the negative impacts of variable redundancy or mismatch, resulting in an optimal working state at the algorithm level. In contrast, the control group, due to a single variable exceeding a reasonable range, disrupted this algorithmic balance. The performance limitations of one module would propagate to the entire system process, leading to a decline in overall performance. The blank control group, lacking the support of core algorithms such as cross-attention feature fusion, spatiotemporal sequence restoration, and speed adaptive prediction, could only achieve simple data collection and decision execution, unable to cope with the challenges posed by complex environmental interference, occlusion scenarios, and mechanical response characteristics, naturally failing to reach the performance levels of the standard and control groups.

[0117] In summary, the performance trend reflected by the experimental data is essentially a direct manifestation of the compatibility between variable settings and the collaborative mechanism of the core algorithm. Reasonable variable settings can maximize the collaborative advantages of multiple core algorithms, while variable deviation will disrupt the algorithm balance and lead to performance degradation. This further confirms the scientific nature and necessity of limiting the scope of variables.

[0118] Exemplary Description

[0119] To ensure the practicality and rigor of the embodiments of the present invention, the following embodiments directly use the variable parameters and corresponding performance data of the aforementioned ten experimental groups. The implementable process of the present invention is fully presented through specific value parameters. The variables and performance data of each embodiment correspond one-to-one with the experimental groups to ensure the feasibility of the technical solution.

[0120] Example 1

[0121] This embodiment describes a specific implementation of a multi-source sensor monitoring and intelligent decision-making system for pest and disease control in tea gardens. It utilizes a young tea garden as an operational scenario. The core variables and corresponding performance data are as follows: 4 heads for the cross-attention mechanism, a spatiotemporal sequence repair occlusion time threshold of 400ms, and a speed adaptive prediction time increase of 40ms. The corresponding pest and disease identification accuracy is 88.23%, application precision is 86.57%, and equipment operational stability is 92.35%. The specific implementation steps are as follows:

[0122] 1. System Deployment and Parameter Settings

[0123] (1) Deployment of multimodal three-dimensional sensing network: The mobile sensor array is deployed on the plant protection machine, installed at a height of 50cm above the canopy of young tea trees, with the lens tilted down 30° relative to the vertical direction, integrating a 4-channel multispectral camera (including RGB, NIR, and Red-edge bands); static monitoring base stations are set at every 3333.33m 2 A single deployment integrates a homogeneous 4-channel multispectral camera, temperature and humidity sensor, and soil parameter sensor.

[0124] (2) Communication and time synchronization: 4G industrial modules are used to transmit data, and the drug application decision instructions are issued through the MQTT protocol (QoS=1); the RTK-GPS time synchronization protocol is used to ensure that the time stamp synchronization accuracy between the mobile sensor array and the static monitoring base station is ±1ms.

[0125] (3) Core variable settings: The number of heads for cross-attention mechanism is set to 4; the occlusion time threshold for spatiotemporal sequence repair is set to 400ms; the increase in speed adaptive prediction time is set to 40ms.

[0126] (4) Other constant settings: mobile sensor array video sampling frequency 25Hz, environmental parameter acquisition frequency 5Hz; static monitoring base station sampling frequency 15min / time; plant protection machine travel speed reference value 1.5m / s, pesticide application pressure reference value 0.3MPa, solenoid valve pulse frequency reference value 10Hz, spraying volume reference value 50mL / m 2 Online bias fine-tuning learning rate 1×10 -5The residual threshold is 5%, and the continuous sampling period is 3 times; the system runs continuously for 8 hours.

[0127] 2. Multimodal data acquisition

[0128] After the system is started, the mobile sensor array moves along an S-shaped path (path spacing 1.5m) with the plant protection machine, and collects multispectral images (resolution 1280×720) and environmental parameters (temperature, humidity, light) of the young tea garden in real time; the static monitoring base station simultaneously collects multispectral images and environmental data of the blind area. All collected data is transmitted to the data processing module through the 4G module and stored after being synchronized by RTK-GPS.

[0129] 3. Feature extraction and cross-attention fusion

[0130] (1) Feature extraction: The multispectral image is processed using a Backbone network to extract a 480-dimensional visual feature vector (Fv∈R). 480 The environment parameters are mapped to a 480-dimensional environment embedding vector (Ee∈R) through three fully connected layers. 480 ).

[0131] (2) Cross-attention fusion: The environment embedding vector is used as the query matrix Q, and the visual feature vector is used as the key matrix K and the value matrix V. The fusion feature is calculated by the cross-attention feature fusion formula. The ambient light intensity is 800 lux. The system automatically increases the attention weight of the environment embedding vector and reduces light interference.

[0132] 4. Spatiotemporal sequence repair and occlusion fallback

[0133] The system determines the target detection probability in real time. When the detection probability is in the range of 0.3-0.5, it calls the hidden state vector h of the Bi-LSTM model. t-1 Linear interpolation prediction is performed by combining the spatial displacement ΔS provided by the IMU; when the identified target is occluded, if the occlusion time is between 400-600ms, the continuity of the drug application control logic is maintained by relying on time-series prediction; if the occlusion time is >600ms or <400ms, data verification of the nearby static monitoring base station is triggered, and if there is no valid reference, a decision pause command is issued.

[0134] 5. Deep reinforcement learning intelligent decision-making

[0135] (1) End-to-end mapping: Based on the PPO algorithm, the high-dimensional features after cross-attention fusion are mapped into drug administration decision action vectors. (v) target =1.5m / s, p t =0.3MPa、ω t =10Hz).

[0136] (2) Physical delay compensation and speed adaptive prediction: The mechanical system response hysteresis parameter t is introduced into the reward function. delay =120ms; when the agricultural drone's travel speed v current When the speed is >1.5m / s, the prediction time for the application command is increased by 40ms to compensate for the drift deviation of the agent.

[0137] (3) Physical constraint filtering: By using the control vector smoothing constraint formula, the absolute value of the first derivative of the control action vector is limited to ≤0.2, so as to achieve a smooth switching between the application pressure and the flow rate.

[0138] 6. Optimization of pesticide application execution and closed-loop system

[0139] (1) Drug application execution: The filtered control action vector is transmitted to the actuator, which drives the solenoid valve and pressure pump to work together to complete the precise drug application.

[0140] (2) Online deviation fine-tuning: Real-time monitoring of actual application rate Q real With model prediction Q pred The residual ε, when |ε|>5% for 3 consecutive sampling periods, triggers the online gradient descent algorithm, with a 1×10 -5 The learning rate fine-tuning strategy is used to adjust the bias parameters at the network endpoints.

[0141] (3) Data storage and optimization: The local server retains the perception data, decision logs and drug application records of this operation for subsequent quarterly model fine-tuning.

[0142] 7. System performance verification

[0143] The test was completed according to the preset test method. The final system performance was as follows: pest and disease identification accuracy rate 88.23%, pesticide application accuracy 86.57%, and equipment operation stability 92.35%, which met the preset operation requirements.

[0144] Example 2

[0145] The differences between this embodiment and Embodiment 1 are: 6 heads for the cross-attention mechanism, 450ms threshold for occlusion time in spatiotemporal sequence repair, and 45ms increase in speed adaptive prediction time; the corresponding system performance is: 90.17% accuracy in pest and disease identification, 89.32% accuracy in pesticide application, and 94.18% stability in equipment operation.

[0146] Example 3

[0147] The differences between this embodiment and Embodiment 1 are: 8 heads for the cross-attention mechanism, 500ms threshold for occlusion time in spatiotemporal sequence repair, and 50ms increase in speed adaptive prediction time; the corresponding system performance is: 92.45% accuracy in pest and disease identification, 91.68% accuracy in pesticide application, and 96.24% equipment operation stability.

[0148] Example 4

[0149] The differences between this embodiment and Embodiment 1 are: 10 heads for the cross-attention mechanism, 550ms threshold for occlusion time in spatiotemporal sequence repair, and 55ms increase in speed adaptive prediction time; the corresponding system performance is: 91.36% accuracy in pest and disease identification, 90.85% accuracy in pesticide application, and 95.76% stability in equipment operation.

[0150] Example 5

[0151] The differences between this embodiment and Embodiment 1 are: 12 heads for the cross-attention mechanism, 600ms threshold for occlusion time in spatiotemporal sequence repair, and 60ms increase in speed adaptive prediction time; the corresponding system performance is: 89.78% accuracy in pest and disease identification, 88.93% accuracy in pesticide application, and 93.87% equipment operation stability.

[0152] Example 6

[0153] The differences between this embodiment and Embodiment 1 are: the number of heads in the cross-attention mechanism is 3 (exceeding a reasonable range), the occlusion time threshold for spatiotemporal sequence repair is 500ms, and the speed adaptive prediction time is increased by 40-60ms; the corresponding system performance is: pest and disease identification accuracy 83.56%, pesticide application accuracy 81.29%, and equipment operation stability 88.42%.

[0154] Example 7

[0155] The differences between this embodiment and Embodiment 1 are: the number of heads in the cross-attention mechanism is 14 (exceeding a reasonable range), the occlusion time threshold for spatiotemporal sequence repair is 500ms, and the speed adaptive prediction time is increased by 40-60ms; the corresponding system performance is: pest and disease identification accuracy 84.19%, pesticide application accuracy 82.37%, and equipment operation stability 87.95%.

[0156] Example 8

[0157] The differences between this embodiment and Embodiment 1 are: 8 heads for the cross-attention mechanism, 350ms threshold for occlusion time in spatiotemporal sequence repair (exceeding a reasonable range), and an increase of 40-60ms in the speed adaptive prediction time; the corresponding system performance is: 85.32% accuracy in pest and disease identification, 83.74% accuracy in pesticide application, and 89.16% stability in equipment operation.

[0158] Example 9

[0159] The differences between this embodiment and Embodiment 1 are: 8 heads for the cross-attention mechanism, 650ms threshold for occlusion time in spatiotemporal sequence repair (exceeding a reasonable range), and an increase of 40-60ms in the speed adaptive prediction time; the corresponding system performance is: 86.05% accuracy in pest and disease identification, 84.51% accuracy in pesticide application, and 88.73% stability in equipment operation.

[0160] Example 10 (Blank control group, using existing technology)

[0161] This embodiment employs existing tea garden pest and disease application technology, specifically "single multispectral sensing + traditional PID decision-making mode," without cross-attention feature fusion, spatiotemporal sequence repair, and speed adaptive prediction functions. The specific implementation steps are as follows:

[0162] 1. System Deployment and Parameter Settings

[0163] The plant protection machine is equipped with a single 4-channel multispectral camera, installed 50cm above the canopy of young tea trees, with the lens tilted down at 30° and a video sampling frequency of 25Hz. It uses a traditional PID controller as the decision-making core, and communication is via a 4G module. There is no static monitoring base station or RTK-GPS high-precision timing module. The machine's travel speed is 1.5m / s, the application pressure is 0.3MPa, the solenoid valve pulse frequency is 10Hz, and the spraying rate is 50mL / m². 2 The PID controller has a proportional coefficient Kp = 5.0, an integral coefficient Ki = 0.1, and a derivative coefficient Kd = 0.5.

[0164] 2. Data Collection and Recognition

[0165] Multispectral cameras acquire tea garden images in real time, and directly perform grayscale and threshold segmentation on the images to identify pest and disease areas without collecting environmental parameters or fusing features; during the identification process, only image pixel features are recorded, without environmental interference compensation logic.

[0166] 3. Decision-making and execution of drug administration

[0167] Traditional PID controllers output pesticide application control commands according to a preset ratio based on the identified area of ​​pests and diseases, without considering mechanical system response lag and pesticide drift, and without speed adaptive prediction logic; control commands are directly transmitted to the actuators, without physical constraint filtering modules, and solenoid valves are directly started and stopped according to commands.

[0168] 4. System performance verification

[0169] The test was conducted using the same testing method as in the aforementioned embodiments. The final system performance was as follows: pest and disease identification accuracy 72.38%, pesticide application accuracy 70.65%, and equipment operation stability 80.24%.

[0170] Specific work process

[0171] Please refer to Figure 1 After the system is started, the mobile sensor array moves along the preset path with the plant protection machine, and synchronously collects multispectral images and environmental parameters of the tea garden. The static monitoring base station collects multispectral images, temperature and humidity and soil parameters of the blind area in parallel. All collected data are synchronized by RTK-GPS and then transmitted to the data processing module through 4G or 5G modules.

[0172] The Backbone network processes multispectral images to extract visual feature vectors. Environmental parameters are nonlinearly mapped through fully connected layers to generate environmental embedding vectors. Using the environmental embedding vectors as query matrices and the visual feature vectors as key and value matrices, a cross-attention mechanism is used to calculate fused features and dynamically adjust the attention weights of the environmental embedding vectors to compensate for environmental interference.

[0173] The system determines the target detection probability in real time. When the detection probability is within a specific range, it calls the hidden state vector of the Bi-LSTM model and performs linear interpolation prediction in combination with the spatial displacement provided by the IMU. It also counts the target occlusion time in real time. When the occlusion time does not exceed the threshold, it maintains the continuity of the drug application control logic based on time-series prediction. When the occlusion time exceeds the threshold, it triggers data verification from the nearby static monitoring base station. If there is no effective reference, it issues a decision to pause the process and resumes drug application after the visual perception stabilizes the tracking.

[0174] Based on the PPO algorithm, the high-dimensional features after cross-attention fusion are mapped into a continuous application decision action vector. The mechanical system response hysteresis parameter is introduced into the reward function to achieve physical delay compensation. The speed of the plant protection machine is monitored in real time, and the prediction time of the application command is increased when the speed exceeds the threshold. Physical constraint filtering is performed by limiting the first derivative of the control action vector to achieve a smooth switching between application pressure and flow rate.

[0175] The filtered control action vector is sent to the actuator via the MQTT protocol to drive the solenoid valve and pressure pump to work together to complete the drug application; the residual between the actual drug application amount and the model prediction value is monitored in real time. When the absolute value of the residual and the duration reach the set conditions, the online gradient descent algorithm is triggered to fine-tune the terminal bias parameters of the strategy network; the local server retains the perception data, decision logs and drug application records for subsequent incremental model optimization.

[0176] The technical features described above can be combined in any way. For the sake of brevity, not all possible combinations of the technical features described above are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A multi-source sensor monitoring and intelligent decision-making system for pesticide application in tea gardens, characterized in that: Includes the following steps: Step 1: Acquire multimodal sensing data of the tea garden in real time. The multimodal sensing data includes multispectral image features with a frequency range of 25-35Hz and environmental sensing parameters with a frequency range of 5-15Hz. Step 2: Perform feature fusion on the multimodal sensing data, map the environmental sensing parameters into environmental embedding vectors, and use a cross-attention mechanism to adjust the gain of multispectral image features. When the ambient light intensity is within the threshold range of 800-1200 lux, the confidence compensation logic is triggered. Step 3: Perform spatiotemporal sequence restoration on the identified target, when the detection probability P of the identified target... det When the value is in the range of 0.3-0.5, the hidden state vector is used in conjunction with the spatial displacement ΔS for interpolation prediction. Step 4: Generate drug administration decision action vector A based on deep reinforcement learning algorithm t During the generation process, a mechanical system response hysteresis parameter t with a preset range of 120-180ms is introduced. delay And according to the travel speed v of the plant protection machine current Dynamically adjust the decision-making and prediction time; Step 5: Analyze the drug administration decision-making action vector A t Perform physical constraint filtering and output the result to the actuator.

2. The system as described in claim 1, characterized in that: The feature fusion described in step two employs a multi-head cross-attention mechanism, with the number of heads h being 4, 6, 8, 10, or 12. The feature extraction and fusion logic is as follows: Visual feature extraction vector is The environment embedding vector is Let the number of heads in the cross-attention mechanism be... First, the query matrix, key matrix, and value matrix are decomposed into h independent subsets through linear transformation: the query matrix of the i-th head. ,in Let be the query linear transformation matrix of the i-th head; and be the key matrix of the i-th head. ,in Let be the linear transformation matrix of the key for the i-th head; and be the value matrix of the i-th head. ,in Let be the linear transformation matrix for the value of the i-th head; where , ,satisfy i=1,2,...,h,d k Let i be the feature dimension of the i-th head; Single-head attention calculation: , Multi-head attention concatenation and final transformation: The outputs of h heads are concatenated and then subjected to a linear transformation to obtain the final fused feature. The complete formula is as follows: ;in, To output a linear transformation matrix for multi-head attention, This represents a vector concatenation operation; the query matrix Q is derived from the context embedding vector E. e The mapping is generated by the key matrix K and value matrix V from the visual feature vector F. v Mapping generation.

3. The system as described in claim 1, characterized in that: The specific logic of interpolation prediction in step three is as follows: when the time of occlusion of the identified target is 400-600ms, the hidden state vector output by the Bi-LSTM model is used to maintain the continuity of the drug administration control logic; when the time of occlusion of the identified target is >600ms or <400ms, the interpolation prediction is stopped and a decision pause command is issued.

4. The system as described in claim 1, characterized in that: The definition of the drug administration decision action vector in step four is as follows: Drug administration decision action vector formula: ; Among them, v target p represents the target travel speed of the agricultural machinery. t Indicates the pressure of drug application, ω t This indicates the solenoid valve pulse frequency; when the travel speed v current When the speed is >1.5m / s, the generation and prediction time of the drug application decision vector increases by 40-60ms.

5. The system as described in claim 1, characterized in that: The logic of the physical constraint filtering process described in step five is as follows: Control vector smoothing constraint formula: By limiting the absolute value of the first derivative of the control vector to no more than 0.2, the physical impact of pressure and flow switching of the actuator is eliminated.

6. The system as described in claim 1, characterized in that: It also includes step six: real-time monitoring of the actual dosage Q. real With model prediction Q pred The residual ε is calculated such that when |ε|>5% and the duration reaches 3 sampling periods, the terminal bias parameters of the policy network are fine-tuned using online gradient descent, where the learning rate η for fine-tuning is 1×10⁻⁶. -5 .

7. The system as described in claim 1, characterized in that: It also includes a mobile sensor array deployed on the plant protection machine, the mobile sensor array including a 4-channel multispectral camera; and a static monitoring base station deployed in the blind area of ​​the tea garden, the static monitoring base station including a homogeneous 4-channel multispectral camera, temperature and humidity sensor and soil parameter sensor; The timing synchronization module is used to perform spatiotemporal alignment between the mobile sensor array and the static monitoring base station.

8. The system as described in claim 7, characterized in that: The deployment rules for the mobile sensor array are as follows: in young tea gardens, the multispectral camera is installed at a height of 50-80cm above the tea tree canopy; in mature tea gardens, the multispectral camera is installed at a height of 80-120cm above the tea tree canopy; and the lens of the multispectral camera is tilted downwards at 30° relative to the vertical direction.

9. The system as described in claim 7, characterized in that: The technical parameters of the time synchronization module are as follows: it adopts the RTK-GPS time synchronization protocol, and the timestamp synchronization accuracy between the mobile sensor array and the static monitoring base station is ±1ms.

10. The system as described in claim 7, characterized in that: The deployment density of the static monitoring base stations is 3333.33m. 2 Deploy a group, and perform data sampling on the static monitoring base station once every 15 minutes.

Images