A precision variable fertilization control system for farmland based on intelligent agents and edge computing

Abstract

This invention relates to a precision variable fertilization control system for farmland based on intelligent agents and edge computing, applied in the field of smart agriculture. The system includes a cloud-based model management unit, multiple edge intelligent agent nodes, a variable fertilization execution mechanism, and a human-machine collaborative interaction terminal. Multiple edge intelligent agent nodes collaborate based on distributed arbitration rules to resolve operational conflicts related to shared agricultural resources. This rule uses resource token management as its core mechanism. When spatiotemporal conflicts arise between nodes regarding resource demands, they exchange negotiation messages containing token status, resource demands, and benefit values. Each node independently runs the arbitration algorithm, reaches a consensus, and generates a collaborative operation sequence table as a scheduling credential, thereby driving the execution mechanism. This invention fills the technological gap in autonomous collaborative rules at the physical execution layer of distributed agricultural equipment, achieving decentralized, highly robust, and highly secure precision variable fertilization operations.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and more specifically, to a precision variable fertilization control system for farmland based on intelligent agents and edge computing. Background Technology

[0002] As smart agriculture develops towards multi-device, distributed, and highly autonomous directions, the control core of precision variable fertilization systems is being decentralized from centralized decision-making in the cloud to the edge. When multiple edge intelligent agents (such as autonomous mobile variable fertilization machines) with local environmental perception and real-time decision-making capabilities work collaboratively in the field, they will inevitably face conflicts at the physical level regarding the real-time and spatiotemporal requirements for limited shared resources (such as shared fertilizer mother liquor supply pumps, main irrigation canals, and overlapping operation paths).

[0003] Currently, there are two main types of solutions for collaboration in agricultural systems, but both have fundamental limitations:

[0004] The first type is the traditional solution based on centralized cloud scheduling. This method uploads all decision requests for unified scheduling by a central server. This type of solution has a significant risk of single point of failure; the system's collaborative capabilities are completely lost when there is network latency or interruption, making it difficult to meet the real-time, concurrent conflict resolution needs of distributed devices.

[0005] The second category comprises more advanced collaborative decision-making models based on multi-agent reinforcement learning and graph neural networks (e.g., the scheme described in Chinese invention patent CN118586432B). These schemes optimize model parameters by constructing an interaction graph between agents during the training phase to output theoretically superior collaborative strategies. However, their technical essence remains at the level of model training and strategy optimization in virtual space. When multiple physical execution units making decisions based on such models experience resource contention during actual operation in real farmland, existing technologies do not provide any specific, executable communication protocols or real-time arbitration rules to resolve conflicts instantly. They provide a "smart decision-making brain," but lack the "collaborative rules" that allow "multiple autonomous hands and feet" to cooperate harmoniously in the physical world.

[0006] Therefore, designing a collaborative control mechanism for distributed farmland operation agents that can autonomously, efficiently, and fairly negotiate and resolve physical resource conflicts during system operation has become a core bottleneck in achieving truly reliable and efficient distributed precision operations, and is also a significant gap in the existing technology system. Summary of the Invention

[0007] To fill the gap in existing technologies regarding distributed physical execution coordination rules, the core technical problem this invention aims to solve is: in the scenario of precision variable fertilization in farmland based on agents and edge computing, when multiple distributed edge agent nodes have real-time and spatiotemporal conflicting operational needs for the same shared agricultural resources due to their local decisions, how to provide a decentralized collaborative control mechanism so that conflicting nodes can autonomously negotiate and generate a consistent executable schedule based on clear rules without relying on centralized scheduling in the cloud, thereby ensuring the overall operational efficiency, robustness, and security of the system.

[0008] To address the aforementioned technical problems, this invention provides a precision variable fertilization control system for farmland based on intelligent agents and edge computing. The system includes a cloud-based model management unit, multiple edge intelligent agent nodes deployed in the field, a variable fertilization execution mechanism, and a human-machine collaborative interaction terminal. The multiple edge intelligent agent nodes are configured to collaborate based on the following distributed arbitration rules to resolve operational conflicts related to shared agricultural resources:

[0009] Each edge agent node holds a resource token with a unique identifier and dynamic priority. The resource token is used to represent the exclusive right to use shared resources within a unit of time period.

[0010] When fertilization control commands generated by at least two nodes conflict in time and space with the demand for the same shared resource, a distributed negotiation process is triggered, and the conflicting nodes exchange negotiation messages containing the current state of their resource tokens, resource demand vectors, and expected agronomic benefit values.

[0011] Each conflicting node independently evaluates and sorts all received negotiation messages based on an arbitration algorithm aimed at maximizing overall agronomic benefits, and reaches a consensus based on the evaluation results to generate a globally consistent collaborative operation timeline. The collaborative operation timeline includes at least the agent node identifiers allocated shared resources, the start and end timestamps of resource usage, and the maximum permitted resource flow within that time period. This timeline will serve as the sole scheduling credential for variable fertilization implementers to access shared resources.

[0012] Edge agent nodes generate final control commands based on the collaborative operation schedule and local decisions to drive the variable fertilization actuator.

[0013] Furthermore, the dynamic priority of resource tokens is periodically calibrated based on the performance evaluation results of the node's historical tasks; the performance evaluation results include at least the completion rate of the task, the fertilization uniformity index, the operational efficiency per unit area, and the contribution of compromise or cooperation to the overall system benefits in historical collaborative tasks.

[0014] Furthermore, in the distributed negotiation process, the negotiation message adopts a lightweight encoding format containing fixed-length fields. The fields include at least the message type, sending node ID, timestamp, resource token structure, resource demand vector, and expected agronomic benefit value. The resource token structure includes token ID, basic priority, and current dynamic priority. The resource demand vector includes at least the type identifier of the required shared resource, the absolute time window requested, and the resource demand curve generated based on the fertilizer prescription map within that window.

[0015] Furthermore, the header of the negotiation message has an encryption flag. When the flag indicates sensitive negotiation, the resource token structure and resource requirement vector field in the message body are transmitted in encryption.

[0016] Furthermore, the human-computer collaborative interaction terminal is configured as follows:

[0017] Provides a conflict decision-making interface that displays the status of conflict nodes, resource requirements, and a comparison of expected benefits based on different arbitration results.

[0018] It receives strategy preference parameters input from agronomic experts, which are used to adjust the weights of yield, quality, and energy consumption in the arbitration algorithm.

[0019] It has an emergency braking channel that can generate the highest priority global braking command and send it directly to the safety controller of the variable fertilization actuator. The global braking command covers any ongoing collaborative operation sequence.

[0020] Furthermore, the edge agent node also includes an instruction pre-execution simulation module;

[0021] Before the final control command is issued, this module quickly simulates the execution effect of the command based on a digital twin model of the farmland area;

[0022] If the simulation results show that there is a risk of over-fertilization, trajectory conflict, or violation of preset agronomic rules, the following operations will be automatically triggered: mark the current instruction and simulation risk as a high-risk event and submit it to the human-machine collaborative interaction terminal for confirmation; while waiting for confirmation, initiate a new local distributed negotiation aimed at avoiding this risk.

[0023] Furthermore, the edge agent node also includes an incremental learning engine and a logically isolated data isolation area;

[0024] The data isolation area is used to store experience data packages consisting of historical operation data, which are tagged with time and space. The experience data packages contain at least environmental data, execution instructions, and actual feedback data.

[0025] During system idle periods, the incremental learning engine asynchronously retrieves experience data packages to fine-tune the local decision-making model, and then encrypts the resulting model parameter differences before uploading them to the cloud-based model management unit.

[0026] Furthermore, the variable fertilization actuator is an autonomous mobile multi-compartment variable fertilizer dispenser, which is configured as follows:

[0027] Receive the final control command and parse it into parallel control signals for the motors of each batching bin, the mixing device, and the traveling mechanism;

[0028] Through the built-in high-precision fertilizer discharge sensor and positioning module, the actual fertilizer discharge volume and the real-time location information of the machine are combined to form a feedback data stream, which is transmitted back to the corresponding edge intelligent agent node in real time for decision-making closed-loop calibration.

[0029] Furthermore, the cloud-based model management unit is configured as follows:

[0030] Receive encrypted model parameter differences and associated job performance evaluation vectors uploaded by each edge agent node;

[0031] Based on the federated learning framework, the weights of each node in the global model aggregation are dynamically assigned according to the performance evaluation vectors uploaded by each node, and the global decision model is updated.

[0032] The updated global model is made lightweight, and the parameters of the new version model are distributed differently based on the historical performance and hardware capabilities of each node.

[0033] Furthermore, the specific strategy for distributing new version model parameters in a differentiated manner is as follows:

[0034] For nodes whose operational efficiency assessments consistently exceed preset thresholds and whose data quality is stable, a cutting-edge model version with a higher contribution weight is issued; for nodes whose efficiency assessments are average or whose data is abnormal, a basic, robust model version is issued and a system diagnostic process for that node is triggered simultaneously.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. This invention proposes a complete set of decentralized arbitration rules for resolving resource conflicts at the physical execution layer among multiple farmland operation agents, filling the technological gap from "intelligent decision-making" to "harmonious execution" and making highly autonomous distributed precision operations possible.

[0037] 2. This invention, based on a peer-to-peer negotiation mechanism using resource tokens, completely eliminates the reliance on centralized cloud scheduling. Even in the event of network outages or cloud service failures, the edge node cluster can still autonomously maintain collaborative operations, resulting in a significant improvement in system availability and robustness.

[0038] 3. By deeply integrating instruction pre-execution simulation with risk-triggered renegotiation mechanism, this invention achieves a paradigm shift from "execution-inspection-remediation" to "prediction-negotiation-prevention," greatly reducing the safety risks of on-site operations. This is something that cannot be achieved by simply optimizing the decision-making model.

[0039] 4. The dynamic priority calibration mechanism based on historical contribution enables the arbitration rules of this invention not only to resolve instantaneous conflicts, but also to guide agents to form a collaborative consensus, spontaneously optimize the overall efficiency of the system in long-term operation, and have the ability to self-evolve. Attached Figure Description

[0040] Figure 1 This is a structural block diagram of the precision variable fertilization control system for farmland based on intelligent agents and edge computing in an embodiment of the present invention;

[0041] Figure 2 This is a flowchart of the distributed arbitration rules and conflict resolution process in an embodiment of the present invention;

[0042] Figure 3 This is a flowchart illustrating the risk closed-loop processing of the instruction pre-execution simulation module and the arbitration mechanism in an embodiment of the present invention. Detailed Implementation

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and do not constitute a limitation thereof. Example:

[0044] 1. System Overall Architecture and Distributed Arbitration Process

[0045] Figure 1 This paper presents a precision variable fertilization control system for farmland based on intelligent agents and edge computing, including a cloud-based model management unit, multiple edge intelligent agent nodes, a variable fertilization execution mechanism, and a human-machine collaborative interaction terminal.

[0046] Edge intelligence nodes are deployed in the fields, and their core task is to coordinate based on distributed arbitration rules to resolve operational conflicts over shared agricultural resources. The core of this distributed arbitration rule is a resource token management mechanism, which is described below in conjunction with... Figure 2 Taking the dispute between node A and node B over a shared fertilizer mother liquor pump P as an example, the distributed arbitration process is described in detail:

[0047] Conflict Detection and Triggering: Node A and Node B simultaneously generate fertilization control commands requiring the use of pump P based on their respective local decision models, and the time windows for requesting use overlap. Each node identifies this spatiotemporal conflict by listening to the communication channel or receiving resource status broadcasts.

[0048] Exchange of Negotiation Messages: After a conflict is triggered, Node A and Node B enter a distributed negotiation process. Both parties construct and exchange negotiation messages. Taking Node A as an example, its message includes: the resource token it holds. The current state (e.g., priority value), detailed resource requirement vector (request to use pump P in) to The demand flow curve for the specified time period is as follows: ), and expected agronomic benefit values (e.g., the estimated benefits of increased production).

[0049] Independent evaluation and consensus: Node A and Node B each run the same arbitration algorithm aimed at maximizing overall agronomic benefits. Each node independently calculates the total system benefit under all possible scheduling schemes and selects the scheme with the highest total benefit as the local preferred scheme.

[0050] Generating scheduling vouchers: Since the algorithms are consistent and the inputs are the same, nodes A and B will independently arrive at the same optimal solution. Both parties confirm this solution as a collaborative work sequence table. This sequence table becomes a binding scheduling voucher, clearly defining the three elements of resources, time, and flow.

[0051] Command execution: Nodes A and B adjust their final control commands according to the timing table reached at this moment, and drive the variable fertilizer actuator to access pump P in sequence to complete the operation.

[0052] 2. Dynamic priority calibration of resource tokens

[0053] Dynamic prioritization of resource tokens is crucial for the fairness and evolution of the arbitration mechanism. After each job, the system prioritizes the nodes. Generate performance evaluation report This includes not only task completion rate fertilization uniformity Work efficiency In addition to basic metrics, there is also a collaborative contribution metric. This indicator quantifies the "big picture" behavior of nodes in historical collaboration, such as by statistically analyzing their most recent... The overall benefit gained to the system due to compromises made during this negotiation. An example of the calculation formula is as follows:

[0054] ;

[0055] in, It is the overall system benefit gain resulting from compromise in the k-th negotiation. For indicator functions, when node The value is 1 if a compromise is made in the k-th negotiation, and 0 otherwise.

[0056] Dynamic priority of tokens Every Calibrate once per work cycle. The calibration formula incorporates historical priority, basic performance, and contribution. For example:

[0057] ;

[0058] in, This is the historical weight decay coefficient. and These are the weighting coefficients for basic performance and contribution, respectively. This is a comprehensive scoring function for basic indicators. Contribution. The introduction of this feature enables actively cooperating nodes to receive higher priority in future arbitrations, thus achieving incentive compatibility.

[0059] 3. Specific implementation of the negotiation communication protocol

[0060] The efficiency and security of negotiation rely on standardized communication protocols. Negotiation messages use a lightweight, fixed-length encoding format (e.g., 128 bytes), with fields including: message type, sending node ID, timestamp, encryption flag, resource token structure, resource demand vector, expected agronomic benefit value, and CRC checksum. The resource demand vector details the resource request; for example, for irrigation resources, it encodes the resource type, start and end timestamps, and flow curve data (10 time-flow sampling points) generated from a fertilizer prescription map. The encryption flag triggers selective encryption: for example, when negotiation involves resources in the core commercial production area, the flag is set, and the system only encrypts the token structure and demand vector fields using AES-128, minimizing communication overhead while protecting business confidentiality.

[0061] 4. Human-machine collaboration and safety closed-loop control

[0062] The system constructs a reliable closed loop through deep human-machine collaboration and proactive safety simulation. The human-machine interface provides a conflict decision-making interface, displaying the status of conflict nodes, requirements, and a comparison of benefits under different arbitration schemes. Strategy preference parameters input by agronomic experts (such as a 6:3:1 weighting ratio of yield, quality, and energy consumption) are converted in real-time into weighting coefficients in the objective function of the arbitration algorithm. This guides the consensus outcome. Before the instruction is issued, the instruction pre-execution simulation module performs a rapid simulation based on a digital twin model of farmland. For example... Figure 3As shown, if the simulation detects risks such as excessive fertilization or trajectory conflicts, the system not only generates an alarm requesting manual confirmation, but more importantly, it immediately initiates a localized distributed negotiation aimed at risk avoidance, forming a proactive safety closed loop of "simulation-negotiation". Furthermore, the terminal's emergency braking channel can generate a high-priority global braking command, directly overriding any currently executing timing schedule.

[0063] 5. Edge-side data management and incremental learning

[0064] Edge nodes have logically isolated data isolation areas (or data caches) for storing experience data packages tagged with <timestamp, GPS coordinates, growth cycle>. These data packages must contain at least environmental data (such as soil moisture). The system executes commands (such as motor speed) and receives actual feedback data (such as actual fertilizer discharge). During system idle periods, the incremental learning engine asynchronously retrieves data packets with specific labels to fine-tune the local decision model, generating model parameter differences. . Encrypted and associated performance evaluation vector Upload them together to the cloud.

[0065] 6. Precise control of variable-rate fertilization execution mechanisms

[0066] The variable fertilizer application actuator is an autonomous mobile multi-compartment variable fertilizer dispenser. The final control commands it receives are parsed into parallel control signals for the motors, mixing devices, and traveling mechanisms of each feed compartment. During operation, a real-time feedback data stream of <time, location, actual fertilizer application [N, P, K]> is generated using built-in high-precision fertilizer discharge sensors and a positioning module and transmitted back to the edge nodes for closed-loop calibration in decision-making.

[0067] 7. Cloud-based model co-evolution and differentiated distribution

[0068] The cloud-based model management unit is based on a federated learning framework and uses performance evaluation vectors uploaded by each node. Dynamically allocate aggregate weights (For example, let ( This is used to weight and aggregate the encrypted model parameter differences. The global model is updated. The updated global model undergoes knowledge distillation to generate a cutting-edge version and a robust base version. The cloud then updates the model based on the historical performance of the nodes (e.g., ...). and Based on factors such as whether the performance consistently exceeds the threshold and hardware capabilities (e.g., memory, computing power), differentiated deployment is implemented: advanced versions are deployed to high-performance nodes with strong hardware; basic, robust versions are deployed to nodes with average performance, abnormal data, or limited hardware, and diagnostic processes are triggered (e.g., remotely retrieving data logs to investigate sensor or model failures), thereby achieving adaptive and robust model evolution and deployment.

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A precision variable fertilization control system for farmland based on intelligent agents and edge computing, comprising a cloud-based model management unit, multiple edge intelligent agent nodes deployed in the field, a variable fertilization execution mechanism, and a human-machine collaborative interaction terminal, characterized in that, Multiple edge agent nodes are configured to collaborate based on the following distributed arbitration rules to resolve operational conflicts involving shared physical resources such as fertilizer mother liquor supply pumps, main irrigation canals, and other field operations: Each edge agent node holds a resource token with a unique identifier and dynamic priority. The resource token is used to represent the exclusive right to use shared resources within a unit of time period. The dynamic priority of the resource token is periodically calibrated based on the performance evaluation results of the node's historical operations. The performance evaluation results include at least the completion rate of the task, the fertilization uniformity index, the work efficiency per unit area, and the assessment of the contribution of compromise or cooperation to the overall system benefits in historical collaborative tasks. When fertilization control commands generated by at least two nodes conflict in time and space with the same shared resource requirement, a distributed negotiation process is triggered. The conflicting nodes exchange negotiation messages containing the current state of their resource tokens, resource requirement vectors, and expected agronomic benefit values. The negotiation messages adopt a lightweight encoding format with fixed-length fields, which include at least message type, sending node ID, timestamp, resource token structure, resource requirement vector, and expected agronomic benefit value. The resource token structure includes token ID, basic priority, and current dynamic priority. The resource requirement vector includes at least the type identifier of the required shared resource, the absolute time window requested, and the resource requirement curve generated based on the fertilization prescription map within that window. Each conflicting node independently evaluates and sorts all received negotiation messages based on an arbitration algorithm aimed at maximizing overall agronomic benefits, and reaches a consensus based on the evaluation results to generate a globally consistent collaborative operation sequence table. The collaborative operation time sequence table includes at least the agent node identifier allocated the shared resources, the start and end timestamps of resource usage, and the maximum permitted resource flow within that time period; this time sequence table will serve as the sole scheduling credential for the variable fertilization execution agency to access the shared resources, and nodes without this time sequence table shall not be allowed to access the corresponding shared resources; The edge agent node generates control instructions to be executed based on the collaborative operation schedule and local decisions; The edge agent node also includes an instruction pre-execution simulation module; before the control instruction to be executed is issued, this module quickly simulates the execution effect of the instruction based on a digital twin model of the farmland area; if the simulation results show that there is a risk of over-fertilization, trajectory conflict or violation of preset agronomic rules, the following operations are automatically triggered: the current control instruction to be executed and the simulated risk are marked as high-risk events and submitted to the human-machine collaborative interaction terminal for confirmation; while waiting for confirmation, a new local distributed negotiation is initiated with the goal of avoiding this risk; The control instructions to be executed, which were simulated and found to be risk-free, were determined to be the final control instructions. The edge agent node drives the variable fertilization execution mechanism according to the final control command.

2. The precision variable fertilization control system for farmland based on intelligent agents and edge computing according to claim 1, characterized in that, The header of the negotiation message has an encryption flag. When the flag indicates sensitive negotiation, the resource token structure and resource requirement vector field in the message body are transmitted in encryption.

3. The precision variable fertilization control system for farmland based on intelligent agents and edge computing according to claim 1, characterized in that, The human-computer collaborative interaction terminal is configured as follows: Provides a conflict decision-making interface that displays the status of conflict nodes, resource requirements, and a comparison of expected benefits based on different arbitration results. The system receives strategy preference parameters input by agronomic experts, which are used to adjust the benefit calculation weights of yield, quality, and energy consumption in the arbitration algorithm. It has an emergency braking channel that can generate the highest priority global braking command and send it directly to the safety controller of the variable fertilization actuator. The global braking command covers any ongoing collaborative operation sequence.

4. The precision variable fertilization control system for farmland based on intelligent agents and edge computing according to claim 1, characterized in that, The edge agent node also includes an incremental learning engine and a data isolation area logically isolated from it; The data isolation area is used to store experience data packages consisting of historical operation data and tagged with time and space. The experience data packages include at least environmental data, execution instructions and actual feedback data. During system idle periods, the incremental learning engine asynchronously retrieves the experience data package to fine-tune the local decision model, and then encrypts the resulting model parameter differences before uploading them to the cloud-based model management unit.

5. The precision variable fertilization control system for farmland based on intelligent agents and edge computing according to claim 1, characterized in that, The variable fertilization actuator is an autonomous mobile multi-compartment variable fertilizer dispenser, which is configured as follows: Receive the final control command and parse the final control command into parallel control signals for the motors of each batching bin, the mixing device and the traveling mechanism; Through the built-in high-precision fertilizer discharge sensor and positioning module, the actual fertilizer discharge volume and the real-time location information of the machine are combined to form a feedback data stream, which is transmitted back to the corresponding edge intelligent agent node in real time for decision-making closed-loop calibration.

6. The precision variable fertilization control system for farmland based on intelligent agents and edge computing according to claim 1, characterized in that, The cloud-based model management unit is configured as follows: Receive encrypted model parameter differences and associated job performance evaluation vectors uploaded by each edge agent node; Based on the federated learning framework, the weights of each node in the global model aggregation are dynamically assigned according to the performance evaluation vectors uploaded by each node, and the global decision model is updated. The updated global model is made lightweight, and the parameters of the new version model are distributed differently based on the historical performance and hardware capabilities of each node.

7. The precision variable fertilization control system for farmland based on intelligent agents and edge computing according to claim 6, characterized in that, The specific strategy for distributing new version model parameters in a differentiated manner is as follows: For nodes whose operational performance evaluation consistently exceeds the preset threshold and whose data quality is stable, a cutting-edge model version with a higher contribution weight is issued; for nodes whose performance evaluation is average or whose data is abnormal, a basic and robust model version is issued and a system diagnostic process for that node is triggered simultaneously.

Images