Multi-level forest resource monitoring method and system

Abstract

The invention relates to the technical field of computers and the technical field of intelligent monitoring of forest resources, and particularly discloses a multi-level forest resource monitoring method and system. The method comprises the following steps: collecting computing power of edge nodes, storing and communicating state information, performing normalization processing, and constructing task resource mapping in combination with a forest resource monitoring task set; constructing a scheduling optimization model based on the mapping, executing multi-constraint scoring and iterative optimization, and generating a task-node matching structure; compressing the central graph neural network model according to a matching result and deploying the central graph neural network model to edge nodes; selecting a communication mode in combination with the communication state, and configuring channel resources; executing local reasoning and asynchronously returning data; and finally, extracting features to complete data aggregation, feeding back task weights and node scores, and updating a scheduling strategy in a closed loop. According to the method, the global monitoring scheduling efficiency and the system stability under the environment of uneven resource distribution and communication fluctuation are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical fields of computer technology and intelligent forest resource monitoring technology, and in particular to a multi-level forest resource monitoring method and system. Background Art

[0002] With the improvement of ecological protection requirements and the in-depth digital management of forest resources, the following problems have gradually emerged in traditional forest monitoring means: First, the monitoring system mostly relies on the central node for processing. Facing a forest environment with a wide geographical distribution and strong resource heterogeneity, it cannot effectively adapt to the reality of uneven computing power, storage, and communication capabilities among edge nodes, resulting in a relatively high overall task response delay; Second, the existing scheduling methods generally adopt static or rule-driven strategies, lacking the ability to perceive and respond to the dynamic state changes of edge nodes in real time, and it is difficult to achieve the global optimal allocation of monitoring tasks under resource-constrained conditions; Third, the problems of high bandwidth occupancy and poor stability of the data backhaul link further restrict the real-time feedback and closed-loop processing capabilities of monitoring results.

[0003] There is still a lack of a method and system in the prior art that can dynamically construct a task-resource mapping relationship under multi-dimensional resource constraints, and combine the partial derivative analysis of edge node states and the system functional control mechanism to achieve the optimal matching allocation of forest resource monitoring tasks between nodes. Summary of the Invention

[0004] The present invention provides a multi-level forest resource monitoring method and system to solve the problem of how to construct a multi-dimensional task-resource mapping based on the computing power state, storage capacity, and communication conditions of edge nodes, and achieve a dynamically stable optimal matching structure of forest resource monitoring tasks between nodes by introducing a multi-constraint scheduling scoring function, a node state partial derivative mechanism, and a system energy functional optimization model, thereby improving the global monitoring scheduling efficiency and feedback closed-loop ability in a heterogeneous resource environment.

[0005] To solve the above technical problems, the present invention provides a multi-level forest resource monitoring method, including:

[0006] Construct a task-resource mapping by normalizing the state information of each edge node and the set of forest resource monitoring tasks;

[0007] Construct a scheduling optimization model by obtaining the task-resource mapping and perform iterative matching to generate a matching allocation structure;

[0008] The expression of the scheduling optimization model is:

[0009]

[0010] Wherein, is the task and the edge node The comprehensive adaptation score; The task-node matching set formed by the current scheduling policy; The scheduling conflict cost function; The conflict penalty factor; The comprehensive matching score of the scheduling policy;

[0011] The matching allocation structure includes a model deployment priority score:

[0012]

[0013] Among them, The deployment priority score of the task-node pair; is the node The current scheduling offset risk indicator;

[0014] Compress the central GNN model according to the matching allocation structure and deploy it to the edge nodes, construct the model deployment structure, and complete the registration marking and version binding;

[0015] Select the communication mode and configure the communication channel according to the model deployment structure and communication status information, and complete the synchronization and storage of the communication configuration structure;

[0016] Execute the local inference task of the model deployment structure, complete the encoding and asynchronous feedback of the inference result, and generate the feedback data structure;

[0017] Extract features and aggregate data from the feedback data structure, generate the task weight policy and node resource score, and feedback to the task resource mapping and update the configuration policy.

[0018] Further, constructing the task resource mapping includes:

[0019] Normalize the computing power status information, storage status information, and communication status information of each edge node to generate edge node status information;

[0020] Structurally process the forest resource monitoring task set according to the task granularity, task priority, and computational complexity to generate task set information;

[0021] Construct a task resource mapping based on the edge node status information and task set information.

[0022] Further, constructing the scheduling optimization model includes:

[0023] Obtain the task resource mapping, and construct a scheduling optimization model based on the edge node status information and task set information;

[0024] Execute the task scheduling matching operation to generate the scheduling intermediate result;

[0025] Execute iterative optimization and constraint rearrangement based on the scheduling intermediate result to update the scheduling result.

[0026] Furthermore, generating a matching allocation structure includes:

[0027] Extract the optimal allocation relationship based on the scheduling result after iterative optimization, and construct a matching allocation structure;

[0028] Generate an edge node task execution list according to the matching allocation structure.

[0029] Furthermore, constructing a model deployment structure includes:

[0030] Obtain the matching allocation structure, perform compression processing on the central graph neural network model to generate a lightweight model;

[0031] Deploy the lightweight model to the edge nodes according to the matching allocation structure to construct a model deployment structure;

[0032] Register and mark the model deployment structure, and bind version information.

[0033] Furthermore, configuring the communication channel includes:

[0034] Obtain the model deployment structure and communication status information, and select a communication mode according to the communication quality level;

[0035] Construct a communication configuration structure and allocate communication channel resources;

[0036] Complete the edge node synchronization processing of the communication configuration structure, and store the configuration result in the communication configuration table.

[0037] Furthermore, executing a local inference task includes:

[0038] Obtain the model deployment structure deployed on the edge nodes, and call the lightweight model to execute the local inference task;

[0039] Obtain the inference result data, and perform data encoding according to the communication configuration structure;

[0040] Asynchronously transmit the inference result through the communication configuration structure to generate a backhaul data structure.

[0041] Furthermore, updating the configuration policy includes:

[0042] Obtain the backhaul data structure, perform global feature extraction and multi-node data aggregation;

[0043] Generate a task weight policy and a node resource score;

[0044] Feed back the task weight policy and the node resource score to the task resource mapping to complete the policy update.

[0045] Furthermore, the graph neural network model is compressed through knowledge distillation, and the communication configuration structure is constructed using a time-division multiple access mechanism.

[0046] Furthermore, a multi-level forest resource monitoring system is applied to any of the multi-level forest resource monitoring methods described above, comprising:

[0047] The task resource construction module is used to obtain edge node status information and forest resource monitoring task sets and construct task resource mapping;

[0048] The task scheduling module is used to perform scheduling optimization based on task resource mapping and generate matching allocation structure;

[0049] Model compression and deployment module, used to generate lightweight models according to the matching allocation structure and deploy them to edge nodes;

[0050] The communication configuration module is used to obtain the model deployment structure and communication status information, build the communication configuration structure and complete the channel configuration;

[0051] The local reasoning and feedback module is used to perform local reasoning tasks and complete the feedback of reasoning results, and generate the feedback data structure;

[0052] The data feedback module is used to generate task weight strategies and node resource scores based on the returned data structure, and update the task resource mapping.

[0053] The key innovative features of the present invention include:

[0054] (1) The system integrates the computing power, storage capacity, and communication status of edge nodes into the scheduling scoring structure to form a cross-dimensional fusion scheduling input feature system to improve the accuracy of task-node adaptation.

[0055] (2) Explicit modeling and iterative reconstruction of scheduling stability are achieved, solving the scheduling instability problem caused by rapid fluctuations in resource status.

[0056] (3) Integrate scoring, deployment, communication, and feedback into a unified structure, open up the entire chain of "scheduling-deployment-communication-feedback", and have a highly scalable and standardized implementation path.

[0057] The following are its main beneficial effects:

[0058] (1) This paper introduces a multi-dimensional task resource mapping mechanism that integrates computing power, storage and communication status. By constructing an adaptation scoring function to comprehensively score the task-node combination, high-precision adaptation is achieved in a resource heterogeneous environment. Compared with the traditional coarse-grained scheduling method based on task priority or geographical distance, it can effectively reduce task scheduling delays and improve the real-time response capability of forest resource monitoring tasks.

[0059] (2) The present invention constitutes a brand-new iterative optimization scheduling framework, in which the partial derivative index can identify high-fluctuation nodes in real time, while the functional objective can suppress the scheduling jitter problem caused by unstable resources, effectively avoiding problems such as node overload, data congestion, and rising failure rate. The system still operates stably in a dynamic environment, providing structural-level guarantee for edge intelligent monitoring.

[0060] (3) The deployment priority scoring function introduced in the matching allocation structure directly associates the scheduling result with subsequent model deployment and communication configuration, promoting the formation of a linkage closed-loop among model compression deployment, channel resource allocation, and feedback strategy, effectively improving the overall efficiency of collaborative monitoring among heterogeneous nodes, and ensuring that the system has end-to-end data consistency and scheduling execution consistency. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic flow chart of a multi-level forest resource monitoring method provided by an embodiment of the present application;

[0062] Figure 2 It is a structural block diagram of a multi-level forest resource monitoring system provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] Embodiment 1: Refer to Figure 1 , which is a schematic flow chart of a multi-level forest resource monitoring method provided by an embodiment of the present invention. The process can at least include steps S100-S600:

[0064] S100. Construct a task-resource mapping by normalizing the state information of each edge node and the forest resource monitoring task set.

[0065] S200. Construct a scheduling optimization model and perform iterative matching by obtaining the task-resource mapping, and generate a matching allocation structure.

[0066] S300. Compress the central GNN model according to the matching allocation structure and deploy it to the edge nodes, construct a model deployment structure, and complete registration marking and version binding.

[0067] S400. Select a communication mode and configure a communication channel according to the model deployment structure and communication status information, and complete the synchronization and storage of the communication configuration structure.

[0068] S500. Execute the local inference task of the model deployment structure, complete the encoding and asynchronous feedback of the inference result, and generate a feedback data structure.

[0069] S600. Perform feature extraction and data aggregation on the feedback data structure, generate a task weight strategy and a node resource score, and feedback them to the task-resource mapping and update the configuration strategy.

[0070] Step S100 includes at least steps S110 - S130:

[0071] S110. Obtain the computing power status information, storage status information, and communication status information of each edge node, perform unified normalization processing, and generate edge node status information.

[0072] Specifically, the edge node status information includes, but is not limited to: processor type, main frequency, memory capacity, remaining storage capacity, communication bandwidth, connection stability, current operating load level, and energy consumption evaluation index. The system first performs status collection operations on all in-network edge nodes through edge monitoring terminals or management centers, relying on the hardware configuration list uploaded during node registration and periodic operation status reports as preliminary raw data.

[0073] Among them, the collection of computing power status information includes dynamic sampling of the available processing resources of the node, and combining historical task response time and peak scheduling load to construct the computing power index value in the current cycle; the storage status information includes the current remaining storage capacity of the node, the size of the data buffer area, and the write rate; the communication status information includes the average signal-to-noise ratio, packet loss rate, average bandwidth occupancy rate, and the number of recent communication failures with the communication base station.

[0074] After completing the collection of raw status data, the system performs unified normalization processing operations. The normalization method uses interval mapping and time window sliding average processing to map various types of indicators to the closed interval [0, 1] with a unified standard to eliminate the impact of inconsistent data dimensions on the scheduling score. The normalization parameter template is set initially by the system or dynamically updated through historical model feedback.

[0075] After normalization, a structured edge node status information structure is generated. This structure contains the above-mentioned multiple-dimensional status attributes and is bound with a unique node identifier. The edge node status information structure will be used as one of the input elements for the subsequent task resource mapping construction operation and participate in the task adaptation scoring process.

[0076] S120. Obtain the forest resource monitoring task set, perform structured processing according to task granularity, task priority, and computational complexity, and generate task set information.

[0077] Specifically, the forest resource monitoring task set comes from multiple data scheduling sources, including regular inspection tasks issued by the central forestry management platform, high-priority response tasks triggered by sudden abnormal alarms, and temporary tasks generated based on remote sensing satellite or unmanned aerial vehicle telemetry analysis.

[0078] Each monitoring task includes a task identifier, a monitoring target location, a task content type (such as image acquisition, audio monitoring, temperature and humidity monitoring, etc.), an execution period, data transmission requirements, and a business emergency level. The system processes the above information by task splitting, extracts the smallest executable task unit, and adds a granularity level label to the task according to predefined rules.

[0079] The task granularity label is evaluated based on the task space range and sampling density; the task priority label is assigned by referring to the task source type and response time limit; the computational complexity evaluation is based on the inference time of past similar task models, the volume of collected data, the processing path structure, etc. for evaluation and modeling.

[0080] Subsequently, the system performs a structured process on the above task set to generate a standardized task set information structure, which includes fields such as task type, target location encoding, priority score, computational complexity score, etc. This task set information structure will be coupled with the edge node status information structure for subsequent processing and participate in the construction process of task resource mapping.

[0081] S130. Construct a task resource mapping based on the edge node status information and the task set information.

[0082] In this step, after obtaining the edge node status information and the task set information, a construction operation of the task resource matching structure is performed. Specifically, it includes: multi-dimensional feature matching, schedulability judgment, and task scoring function calculation.

[0083] First, the system performs a spatial matching on the geographical distribution of edge nodes according to the task location parameters recorded in the task set information, and eliminates invalid nodes whose distance from the target task location exceeds the deployment threshold range.

[0084] Secondly, for each remaining node-task combination pair, a computing power adaptability matching is performed. The system calculates the computing power matching score based on the normalized computing power index of the node and the task computational complexity label; at the same time, it scores the storage adaptability based on the node storage capacity and the task data volume; and combines the communication status index and the task data transmission requirements to evaluate the communication adaptability score.

[0085] The above three scoring dimensions constitute the evaluation core of task resource matching. The system calculates the overall matching score of each node-task pair based on a weighting function and marks the matching feasibility flag in the scoring structure. This scoring structure will be stored in a table structure and used as the output object of the task resource mapping.

[0086] Finally, the constructed task resource mapping includes fields such as task number, node number, comprehensive score value, schedulability flag, etc., forming standardized input data, which is the key input for the construction of the scheduling optimization model and task matching calculation.

[0087] Through the implementation of this step, the system can achieve multi-dimensional structured modeling between the task set and the edge node capabilities, forming a complete task-resource mapping relationship. The mapping result significantly improves the perception accuracy of the subsequent scheduling model for resource heterogeneity, providing a high-quality scoring basis. At the same time, the unified quantification and schedulability expression of three types of core resources, namely computing power, storage, and communication, are realized, providing key support for the efficient allocation and execution of multi-level forest resource monitoring tasks, effectively reducing the overall scheduling delay, and enhancing the resource scheduling robustness of the system in a large-scale node environment.

[0088] Step S200 includes at least steps S210 - S230:

[0089] S210. Obtain the task-resource mapping, construct a reinforcement scheduling optimization model, and perform a task scheduling matching operation.

[0090] The system first obtains the task-resource mapping constructed by S130, which includes task numbers, node numbers, computing adaptation scores, storage adaptation scores, communication adaptation scores, and schedulability flags. Based on the mapping result, the system constructs a task-node bipartite graph structure and expresses the multi-dimensional resource adaptation relationship in the form of edge weights to construct a reinforcement scheduling optimization model.

[0091] Formula ①: Comprehensive adaptation score function

[0092]

[0093] Where:

[0094] : Task and node 's comprehensive adaptation score;

[0095] : Task 's adaptation score in the computing power dimension;

[0096] : Task 's adaptation score in the storage dimension;

[0097] : Task 's adaptation score in the communication dimension;

[0098] : Adaptation weight coefficient, satisfying , predefined by the system or adjusted by historical feedback.

[0099] Based on this, the system generates a comprehensive scoring matrix, which constitutes the edge weight set of the scheduling graph structure and serves as the input basis for the subsequent scheduling model matching. The system takes this scoring matrix as the input and constructs a reinforcement scheduling matching objective function.

[0100] Formula ②: Reinforcement scheduling matching objective function

[0101] To optimize the overall resource utilization efficiency of the system under multiple constraints, the system constructs a reinforcement scheduling optimization model, specifically the following scheduling objective function:

[0102]

[0103] Where:

[0104] : The task-node matching set formed by the current scheduling strategy;

[0105] : Scheduling conflict cost function, measuring the degree of scheduling violations such as task reallocation and node overload;

[0106] : Conflict penalty factor, reflecting the priority of the current scheduling stability;

[0107] : Comprehensive matching score of the scheduling strategy.

[0108] After executing the above reinforcement objective function, the system completes the initial matching of the first round of task scheduling and forms an initial scheduling result set for the next step.

[0109] S220. Iteratively optimize and re-arrange the constraints for the intermediate results of the task scheduling matching operation to obtain the optimal allocation structure.

[0110] To solve the situations of sudden changes in the computing power of edge nodes and dynamic communication failures, this step introduces a stability partial derivative mechanism and a system functional reconstruction method to iteratively adjust the initial matching structure generated in S210.

[0111] After the system obtains the initial scheduling results output by S210, it further identifies the unstable nodes in the scheduling structure and performs constraint re-arrangement optimization.

[0112] Specifically, for each matched node , according to its computing power, storage, and communication status in the normalized state vector , perform a partial derivative operation on the comprehensive scoring function to obtain the scheduling deviation index , as shown in Formula ③:

[0113] Formula ③: Calculation of node scheduling partial derivative

[0114]

[0115] Wherein:

[0116] : Node Current scheduling offset risk index;

[0117] : Node State vector, including three dimensions of computing power, storage, and communication;

[0118] : Sensitivity partial derivative function of the adaptation score to the node state;

[0119] : Vector norm operation, indicating the absolute influence degree of the partial derivative.

[0120] This partial derivative index is used to identify scheduling sensitive nodes. When exceeds the stability threshold it is regarded as a high-fluctuation node and rearrangement replacement needs to be performed. Further, to achieve global optimal scheduling stability control, the system constructs a node energy functional function, as defined in Equation ④:

[0121] Equation ④: Definition of the system scheduling energy functional

[0122]

[0123] Wherein:

[0124] : Scheduling strategy System energy functional value;

[0125] : Single-node scheduling pressure function;

[0126] : Total number of nodes;

[0127] : Node offset index calculated by Equation ③;

[0128] : Matching score from Equation ①.

[0129] Further, the system takes this functional as the minimization objective, while satisfying that the comprehensive score is not lower than the preset threshold ( ). The system obtains the optimal matching structure by replacing the task matching corresponding to high nodes through multiple rounds of iteration until the functional value converges and satisfies the score constraint.

[0130] S230. Generate a matching allocation structure based on the optimal allocation structure.

[0131] After completing the functional optimization iteration of the scheduling matching structure, the system converts the final optimal matching structure into a matching allocation structure body. The structure body records the task number, node number, final adaptation score, node current state vector, deployment priority score, etc. of each task-node pair, and is used to guide subsequent model deployment and communication configuration.

[0132] Formula ⑤: Model deployment priority score

[0133]

[0134] Where:

[0135] : Deployment priority score of the task-node pair;

[0136] : Task and node comprehensive adaptation score;

[0137] : Node current scheduling offset risk index.

[0138] The matching allocation structure body contains task number, node number, final score, deployment priority score and status vector information. This structure body will be used as the input decision basis for the subsequent compression deployment of the S300 module, and is synchronously mapped with the communication channel priority to ensure the closed-loop stable operation of the system resource dynamic allocation mechanism.

[0139] Through the design and implementation of the above S200 module, the present invention constructs a reinforcement scheduling optimization mechanism integrating graph matching, partial derivative detection and system functional control. Compared with the traditional Hungarian algorithm, this mechanism realizes the following technical performances while ensuring the resource adaptation accuracy:

[0140] Optimize the task-node scheduling time complexity and improve the scheduling real-time performance;

[0141] The multi-constraint scoring structure realizes the joint scheduling evaluation of computing power, storage and communication resources;

[0142] The output of the matching allocation structure can be directly served for subsequent model deployment and communication configuration strategies, providing a mathematical and structural basis for realizing the "task scheduling - model deployment - data feedback" closed-loop system.

[0143] Step S300 includes at least steps S310 - S330:

[0144] S310. Obtain the matching allocation structure, perform compression processing on the central GNN model, and generate a lightweight model.

[0145] First, call the matching allocation structure generated in S230. The structure contains task numbers, node numbers, comprehensive score values, deployment priority scores, and node state vectors. The system sorts the task-node pairs according to the deployment priority scores and determines a compression processing priority queue.

[0146] Specifically, adopt the knowledge distillation strategy to lightweight the central graph neural network model. The compression processing includes the following core processes: First, obtain the central GNN model structure and its pre-trained parameters on the original dataset; then construct an auxiliary teacher-student network structure, where the teacher model is a complete structure and the student model is the target compressed version; subsequently, use the training subset associated with the task number in the matching allocation structure as the distillation input data, execute the training process of the joint loss function of cross-entropy and KL divergence, adjust the weight parameters of the student model, and optimize its generalization ability for edge task scenarios.

[0147] During the compression process, dynamically adjust the model layers, feature dimensions, and activation structures according to the state vector of the deployment node (including computing power, storage, and communication dimensions), and strictly control the total amount of compressed model parameters to be lower than the preset storage threshold. After compression, multiple lightweight model instances are generated, and each instance establishes a unique mapping and binding relationship with the target deployment node. The finally generated lightweight model structure and parameters will be used as the direct input for step S320.

[0148] S320. Deploy the lightweight model to the edge nodes according to the matching allocation structure to construct a model deployment structure.

[0149] Complete the model distribution and deployment operations according to the lightweight model set generated in S310 and the task-node binding information in the matching allocation structure. Specifically, the system writes the corresponding lightweight model file to the local model directory of the node through a secure transmission channel according to the edge node identifier corresponding to each task, and initializes the model call interface.

[0150] The deployment operation depends on the device driver adaptation layer of the edge node. The system pre-parses the node computing architecture (such as CPU architecture, GPU type, tensor processing unit, etc.) before distribution, and automatically completes the selection and conversion of the model conversion format (such as ONNX, TensorRT, TFLite). During the deployment process, the system will perform integrity verification and sandbox loading tests on the transmitted model file to ensure that the deployed model can be called without errors and respond with low latency on the target edge node.

[0151] Subsequently, the system constructs a model deployment structure, which is a data encapsulation body designed for a distributed environment. The content includes a model identification code, a node unique identifier, a deployment timestamp, a compression ratio, a call interface path, and a running status flag. This structure is persistently stored locally at the edge node and uploaded to the central scheduling platform for subsequent registration marking processing and unified management of call scheduling.

[0152] S330. Perform registration marking and version binding on the model deployment structure.

[0153] After completing the model deployment at the edge node, perform the registration marking and version binding operations for the deployed model. Specifically, the system first calls the model identification code and node identifier included in the deployment structure to establish a unique registration index for each model-node binding entity and generate a registration token.

[0154] The registration process uses a model scheduling registration interface based on the central scheduling system. The interface receives structured information including a model hash digest, a deployment path, a version number, and a resource binding status, and writes it into the central version control database. This database supports version tracking, historical model rollback, and inference performance recording, providing support for subsequent model maintenance.

[0155] The version binding operation is achieved by introducing a unified model version identification specification. Each lightweight model generates an initial version number after distillation compression is completed, and automatically triggers a version upgrade mechanism after each subsequent round of online fine-tuning or policy update. The version upgrade event will be synchronously written into the deployment structure and the central scheduling platform, and the corresponding version will be automatically loaded by the model loader during inference to ensure the consistency and stability of the inference results.

[0156] Finally, after completing registration and version binding, the model deployment structure becomes a callable deployment entity and participates in the execution process of the local inference task in S500. This structure will also be used as one of the basis for transmission priorities in the subsequent S420 communication configuration, and achieve multi-dimensional joint mapping with communication resource scheduling.

[0157] Through the implementation of the S300 module, the system has achieved the following technical performance and structural support functions:

[0158] Under the drive of the matching allocation structure, perform customized knowledge distillation compression on the central GNN model, taking into account both model accuracy and resource load control;

[0159] Realize the dynamic deployment of lightweight models among heterogeneous edge nodes, supporting automatic format conversion and fast loading under different node architectures;

[0160] Construct a structured model deployment structure and complete the registration marking and version binding operations, providing structural guarantee for unified scheduling and version control of distributed inference;

[0161] The output result of the model deployment structure will be used as one of the bases for the communication configuration priority in the S400 module. At the same time, it provides a structural interface for local inference in the S500 and intermediate structural support for the system performance closed-loop scheduling mechanism.

[0162] Step S400 at least includes steps S410 - S430:

[0163] S410. Obtain the model deployment structure and the communication status information, and select a communication mode according to the communication quality level.

[0164] The system first extracts fields such as deployment node identifiers, deployment timestamps, model compression ratios, interface call paths, and running status flags from the model deployment structure output by the S330 module, and binds them to the normalized communication status information in the S110 module. The communication status information includes, but is not limited to, attribute dimensions such as communication link bandwidth, average signal-to-noise ratio between nodes, historical communication failure times, maximum packet loss rate, and current channel occupancy rate.

[0165] Specifically, the system calculates the communication quality level based on the communication status evaluation results of the deployment nodes. The communication quality level is divided by the central scheduling system according to the standardized index classification strategy, including three categories: high level, medium level, and low level. For deployment nodes with a high communication quality level, the system preferentially selects a high-speed and high-frequency transmission mode (such as Wi-Fi 6 or 5G NR mode); for deployment nodes with a medium communication quality level, the system adopts a medium-speed and controllable transmission mode (such as traditional 4G LTE or TDMA time-division multiplexing mode); while for deployment nodes with a low communication quality level, it switches to a low-frequency communication mode with priority on reliability (such as LoRa or ZigBee and other low-power long-distance solutions) to ensure the integrity of the transmission task and a low packet loss rate.

[0166] The selection basis of the above communication mode not only considers the current communication performance of the node, but also combines the model compression ratio and call path requirements recorded in the model deployment structure to ensure that the communication mode meets the minimum requirement threshold of the corresponding model operation task in terms of throughput capacity and response latency.

[0167] After the system selects the communication mode, it binds the communication mode identifier to the corresponding node identifier to form a preliminary communication configuration unit, providing the basic input for constructing the communication configuration structure in the subsequent S420.

[0168] S420. Based on the communication mode and the model deployment structure, construct a communication configuration structure and configure communication channel resources.

[0169] This sub-step constructs a communication configuration structure body and performs a refined configuration operation of communication channel resources based on the communication mode result determined in S410 and the model deployment structure parameters.

[0170] Specifically, the system groups according to the communication mode and separately establishes communication configuration templates. During the construction of the configuration structure, the system needs to combine the model call frequency, transmission data volume, and estimated inference feedback cycle recorded in the model deployment structure to perform configuration calculations on communication time slots, channel allocation priorities, retransmission mechanism parameters, data packet encapsulation formats, and target coding methods. If the communication mode is a TDMA structure, a time slot allocation table is generated according to the model period and the predicted feedback window; if the communication mode is a LoRa structure, dynamic adjustment calculations of the spreading factor, coding rate, and transmission power are performed.

[0171] Meanwhile, the historical channel congestion rate and access frequency data in the communication status information obtained by the system from S110 will be used as input for communication channel conflict avoidance parameters, and the system will perform interference avoidance optimization on the channel allocation strategy of each edge node in the same communication area accordingly. If the system detects that there are multi-task centralized deployment nodes in the current area, the communication channel frequency configuration will preferentially adopt a frequency hopping strategy, and the backoff timer parameters will be configured to reduce the probability of competition conflicts.

[0172] After the construction of the communication configuration structure is completed, the system generates a communication configuration structure body. The communication configuration structure body includes fields such as deployment node identification, bound communication mode identification, channel number, channel frequency band, start and end times of time slots, data packet encapsulation format, retransmission strategy parameters, and channel priority tags. This structure body is also bound to the model identification code in the model deployment structure to realize a three-dimensional integrated configuration linkage structure of task-model-communication.

[0173] S430. Complete the edge node synchronization process of the communication configuration structure and store the configuration result in the communication configuration table.

[0174] After the construction of the communication configuration structure body is completed, the system immediately starts the communication configuration synchronization mechanism and performs the distribution of configuration instructions and the status registration operation of edge nodes.

[0175] Specifically, the system issues each communication configuration structure body to the corresponding deployment node in the form of a structured command. After receiving the configuration instruction, the node updates the parameters of the local communication module and reloads the communication interface according to the configuration content. The system deploys a confirmation mechanism module on the node side, and after the configuration update is completed, a configuration confirmation signal will be generated and sent back to the central system as an indication of the configuration taking effect.

[0176] After the edge node synchronization process is completed, the system writes all communication configuration structures into the communication configuration table. The communication configuration table is a structured communication resource configuration list maintained by the scheduling system, containing a complete record of the model-communication parameter combination relationship corresponding to each task. The communication configuration table supports real-time query and dynamic update, and is used as the basis for the selection of transmission paths and protocols during the asynchronous feedback channel matching process of the subsequent S500 module.

[0177] Furthermore, communication behavior statistical metrics between edge nodes are additionally recorded in the communication configuration table, including performance parameters such as actual bandwidth occupancy, retransmission ratio, average latency, and error frame ratio, for reference by the S600 module's task resource scoring and feedback update mechanism.

[0178] Complete the parameter binding between the communication configuration structure and the model deployment structure to form an end-to-end transmission control closed-loop, and establish a structural dependence basis for the data feedback of the S500 module and the data feedback of the S600 module.

[0179] Step 500 includes at least steps S510 - S530:

[0180] S510. Obtain the model deployment structures deployed on each edge node and execute the corresponding local inference tasks.

[0181] Specifically, the system first retrieves the model deployment structures registered and bound in step S330. The model deployment structures include edge node identifiers, lightweight model identification codes, deployment paths, compression ratios, model call interface paths, and current running state flags.

[0182] Before performing local inference, the system initializes the edge node inference module according to the deployment path and loads the model file that matches the node architecture. The model loading process depends on the running environment of the edge node, including but not limited to the operating system type, available computing units (such as CPU, GPU, TPU), and their driver versions. After successful loading, the system verifies the callability of the model interface to ensure that the model can normally accept input data and output inference results.

[0183] The input data for the inference task comes from the information collected by the forest resource monitoring sensors mounted on the edge nodes. Specifically, for nodes deployed with video monitoring functions, the input data includes the image frame sequence of the target monitoring area; for audio acquisition nodes, the input is noise sampling or bird sound segments; and for environmental monitoring tasks, the input is multi-dimensional structured environmental data such as temperature and humidity, light intensity, or PM2.5.

[0184] The model input dimensions and formats are pre-configured in the model deployment structure. The system accordingly performs formatting processing and tensor conversion operations on the original collected data, and executes standardization and normalization steps to generate data tensors that conform to the model input interface. Subsequently, the model inference interface is called to perform a forward propagation operation on the above input data to obtain a preliminary analysis result of the forest resource status.

[0185] The inference result is returned in the form of a structure, and the content includes but is not limited to: inference tags (such as fire risk level, illegal logging probability, pest and disease index), corresponding confidence scores, output timestamps, and inference time consumption evaluations, etc. This structure is cached in the local memory space of the node and used as the input data for step S520.

[0186] After the local inference task is completed, the system updates the running status flag field in the model deployment structure and records the inference task number, node identifier, model version, and execution time to support subsequent version comparison and performance statistics.

[0187] S520. Obtain the inference result data of the local inference task and perform data encoding according to the communication configuration structure.

[0188] After the local inference task is completed, the system obtains the inference result structure cached locally on the edge node and calls the communication configuration structure output by S430 for encoding adaptation processing. The communication configuration structure contains fields such as the communication mode selected by the node, channel number, allocated time slot, data encapsulation format, retransmission strategy parameters, and channel priority label, etc.

[0189] First, the system determines the encoding format requirements of the current edge node according to the communication mode identifier. If the communication mode is high-speed and high-frequency type (such as Wi-Fi 6), the encoding format can select the binary packaging method with high compression ratio and low transmission delay; if the communication mode is low-frequency and low-power type (such as LoRa), a redundant control encoding scheme needs to be adopted to increase the error correction field to reduce the bit error rate during long-distance transmission.

[0190] During the encoding process, the system maps each field in the inference result structure to the data packet field table of the communication protocol, performs hash or index conversion compression processing on the text data therein, and performs fixed-point quantization processing on the numerical data to compress the data packet size. At the same time, set the transmission priority flag for the data packet according to the channel priority label allocated in the communication configuration structure. If the node communication load is higher than the set threshold, the system will hang the data packet into the delayed transmission queue and write the start time point of the configured return window.

[0191] Furthermore, if the communication configuration structure enables the sub-packet retransmission mechanism, the system will split the encoded data structure into multiple sub-packets, and each sub-packet is attached with a sequence number identifier and a check redundancy bit for re-splicing and error correction recovery at the receiving end.

[0192] After the data encoding is completed, a standardized inference data packet structure is generated, and the fields include: node identifier, task number, inference tag, confidence score, data packet number, encapsulation format type, priority flag, whether to enable the retransmission mechanism, etc. This structure will be used as the input object for the asynchronous return processing of S530.

[0193] S530. Asynchronously back-transmit the inference result data through the communication configuration structure and generate a back-transmission data structure.

[0194] After the encoding of the data packet structure is completed, the system calls the communication parameters recorded in the communication configuration table constructed by S430 to start the asynchronous back-transmission module to perform data transmission operations. The asynchronous back-transmission module is a multi-threaded asynchronous task scheduling unit, with functions such as automatic retransmission, delay queue management, channel conflict detection, and transmission completion confirmation.

[0195] First, the system binds the local communication interface of the edge node according to the channel number and communication frequency band recorded in the communication configuration structure, initializes the communication channel, and applies for bandwidth resources. If the system detects that the channel resources in the communication area where the current node is located are saturated, the task will be automatically added to the delay queue and the back-off timer will be started.

[0196] The back-transmission process adopts an asynchronous non-blocking transmission mechanism. The system writes each inference result data packet into the transmission buffer in the order of sub-packets and monitors the transmission completion flag bit. If an exception occurs during the transmission (such as bandwidth interruption, packet loss, retransmission failure), the system executes the retransmission logic according to the retransmission strategy parameters in the communication configuration structure. When the maximum number of retries exceeds the threshold, the transmission event will be marked as failed and written into the error log for subsequent call by the S600 feedback mechanism.

[0197] After all data packets are sent, the system assembles the inference result data into a structured back-transmission data structure. The fields include: task number, edge node identifier, model version, inference label, confidence score, communication mode identifier, back-transmission success status flag, transmission time consumption statistics, number of retransmissions, channel number, total number of data packets, and received confirmation timestamp, etc.

[0198] After the back-transmission data structure is locally stored, it is immediately uploaded to the back-transmission buffer queue of the central scheduling system, and the communication status statistical parameters of the corresponding node in the communication configuration table are updated. After the central system obtains the back-transmission data structure, it can perform global feature extraction and data aggregation processing in S600.

[0199] Under the linkage control of the model deployment structure and the communication configuration structure, the local inference operation of the edge node is completed, supporting the consistency of inference calls in heterogeneous model environments; constructing a standardized back-transmission data structure to achieve the data consistency of the whole process from model deployment to communication control and then to data aggregation, providing a reliable input basis for the scoring and policy update of the S600 module.

[0200] Step S600 at least includes steps S610 - S630:

[0201] S610. Obtain the backhaul data structure and perform global feature extraction and multi-node data aggregation processing.

[0202] Specifically, the system first batch extracts the backhaul data structures from multiple edge nodes from the backhaul data cache module of the central scheduling platform. The backhaul data structure is asynchronously generated and uploaded in step S530, and the fields include task number, edge node identifier, model version, inference label, confidence score, communication mode identifier, backhaul success status flag, transmission time consumption statistics, retransmission times, channel number, total number of data packets, and receive confirmation timestamp, etc.

[0203] After the acquisition is completed, the system performs global feature extraction operations. This process is based on the inference label and confidence score fields recorded in the backhaul data structure to extract risk factors, target status, and abnormal distribution characteristics related to forest resource monitoring. For example, for the fire risk task, the system will aggregate the occurrence frequencies of fire situation labels and the corresponding confidence distributions on different nodes to extract the change pattern of the forest heat index; for the pest and disease monitoring task, the consistency index of abnormal plant characteristics identified by multiple nodes in the same target area is calculated.

[0204] During the feature extraction process, the system synchronously reads the deployed model version and inference time consumption fields to judge the adaptation degree of the current model performance on each node and identify the response ability bottleneck of the model under specific resource configurations. At the same time, communication-related fields (such as retransmission times, channel number, backhaul success status) are also involved in the analysis to evaluate the communication path stability and the transmission reliability of each node.

[0205] After completing the global feature extraction, the system classifies the backhaul data of all edge nodes by task number and performs multi-node data aggregation operations. The aggregation process includes but is not limited to: cross-node feature alignment, redundant data elimination, confidence weighted fusion, and abnormal result elimination. The system automatically identifies data abnormal nodes or inference failure nodes according to the consistency difference index before and after aggregation and marks them to enter the subsequent policy regulation module for priority adjustment.

[0206] After the processing of this step, the system generates an aggregated data structure, and the fields include task number, aggregated inference label, multi-node weighted confidence, average inference time consumption, average retransmission rate, communication success rate, model version distribution, and list of abnormal node identifiers, etc., providing basic data support for the scoring and policy generation of the S620 module.

[0207] S620. Update the task weight policy and node resource score based on the multi-node data aggregation processing result.

[0208] After obtaining the aggregated data structure output by S610, the system enters the task regulation and node performance scoring phase. First, the system performs a task importance assessment based on the task dimension. The assessment basis includes: the aggregated inference confidence level, the consistency degree of multi-node inference results, the communication quality level, and the model version consistency index, etc.

[0209] The system generates a weight label for each task, and the weight label reflects the priority processing degree of the task in the next scheduling cycle. The task weight policy includes fields such as task number, current cycle weight score, difference from the previous cycle, task type identifier, and recommended scheduling frequency suggestion, etc. If the task result has characteristics such as low multi-node consistency, confidence level decline, or reduced communication stability, the system automatically increases the weight score of the task to ensure that subsequent resources are preferentially matched; otherwise, the weight is appropriately reduced to relieve the resource allocation pressure.

[0210] At the same time, the system performs a resource scoring operation on the edge nodes. The scoring logic basis includes multi-dimensional indicators such as model execution time, inference accuracy, return success rate, number of retransmissions, and communication channel occupancy rate, etc. During the scoring process, the system references the edge node status information generated in S110 and the return results recorded in S530 to construct a node operation performance model.

[0211] Each node generates a node resource scoring structure, and the fields include node identifier, execution stability score, communication reliability score, model adaptability score, average inference time, abnormal behavior count, and score update timestamp, etc. For nodes with excellent execution performance and stable communication quality, the system marks them as "priority nodes" and assigns a higher priority weight in the next round of S210 task scheduling; while for nodes with frequent failures, return delays exceeding the threshold, or large fluctuations in inference confidence, they are marked as "weight-reduced nodes" and are preferentially excluded or only assigned low-complexity tasks in the next round of task resource mapping.

[0212] Through this step, the system forms a dynamic scoring result in the task-node dimension and outputs it to the S630 module in a structured manner to drive the parameter update process of the task resource mapping.

[0213] S630 feeds back the task weight policy and the node resource scoring to the task resource mapping to update the task and resource configuration policies.

[0214] After the task weight policy and the node resource scoring structure are generated, the system immediately starts the mapping feedback mechanism to complete the parameter write-back and configuration update process. This sub-step is based on the task resource mapping constructed in S130, and combines the scoring results output in S620 to reconstruct the original task-node adaptation score and scheduling feasibility status.

[0215] Specifically, the system first maps the weight scoring field in the task weight policy to the corresponding task entry and updates its priority parameter in the resource mapping table. At the same time, it maps the execution stability and communication scoring fields in the node resource scoring structure to the corresponding node information and replaces its original status indicator value. This process ensures that the task-resource mapping can be dynamically adapted according to the execution feedback in each scheduling cycle, so as to cope with the node performance fluctuations and resource mutations in the actual operating environment.

[0216] Furthermore, the system re-executes a scoring function calculation operation to re-score the updated task-node combination, generating a new task-resource mapping scoring matrix as the input basis for the task scheduling model in the next scheduling cycle S210. At the same time, to improve the adaptation quality after mapping reconstruction, the system introduces a scoring smoothing mechanism to compare the scoring change range and weight adjustment trend in the previous round, avoiding drastic scoring fluctuations caused by instantaneous anomalies and affecting the system scheduling stability.

[0217] Finally, the system stores the updated task-resource mapping in the central task configuration database and writes it into the mapping version control module to ensure that the task-node mapping version is consistent with the model version and the communication configuration structure version. This mapping version number will be used as the retrieval basis for subsequent S210 task scheduling operations to ensure that the system task scheduling chain has a complete configuration backtracking ability.

[0218] Through the implementation of this step, the system realizes the closed-loop control from task execution feedback to configuration structure adjustment, providing support for the global multi-node task scheduling mechanism with high robustness and dynamic regulation capabilities.

[0219] The present invention realizes the following technical performance optimization and system function closed-loop in the multi-level forest resource monitoring task:

[0220] Establish a standardized extraction mechanism for the backhaul data structure to support cross-node data consistency evaluation and multi-dimensional index fusion processing, improving the system perception accuracy;

[0221] Construct a dual feedback system of task weight policy and node resource scoring to support dynamic adjustment of task priorities and hierarchical classification of node capabilities, effectively improving the resource scheduling efficiency;

[0222] Realize the version-level closed-loop update of the task-resource mapping structure to ensure that the task-resource relationship structure has the ability to dynamically adapt and evolve with the operation results, enhancing the overall stability of the system.

[0223] Embodiment 2: Figure 2 Show a structural block diagram of a multi-level forest resource monitoring system according to an embodiment of the present invention. As Figure 2 shown, the structure may include:

[0224] The task resource construction module 10 is used to obtain the edge node status information and the forest resource monitoring task set, and construct a task resource mapping.

[0225] Obtain the computing power, storage, and communication status information of the edge node;

[0226] Normalize the above information (including unit unification and sliding window standardization);

[0227] Obtain the forest resource monitoring task set, and perform structured processing based on parameter tags such as granularity, priority, and complexity;

[0228] Establish a schedulable relationship scoring table between tasks and nodes, and output a task resource mapping structure.

[0229] The task scheduling module 20 is used to perform scheduling optimization based on the task resource mapping and generate a matching allocation structure.

[0230] Construct a multi-constraint bipartite graph structure (node - task);

[0231] Introduce a three-dimensional adaptation scoring weight mechanism for computing power, storage, and communication;

[0232] Perform preliminary matching and functional optimization, and eliminate unstable nodes;

[0233] Output a matching allocation structure containing deployment priority scores as the basis for compression and deployment of subsequent modules.

[0234] The model compression and deployment module 30 is used to generate a lightweight model according to the matching allocation structure and deploy it to the edge node.

[0235] Perform model compression operations based on knowledge distillation to control the model size not to exceed the set resource threshold;

[0236] Generate multi-instance lightweight models according to node resource configurations and complete local format adaptation;

[0237] Generate a deployment structure (including model identification, node number, interface path, version information);

[0238] Complete registration binding, model index construction, and version control table registration.

[0239] The communication configuration module 40 is used to obtain the model deployment structure and communication status information, construct a communication configuration structure, and complete channel configuration.

[0240] Select modes corresponding to high, medium, and low communication levels (such as Wi-Fi, TDMA, LoRa);

[0241] Construct a communication configuration structure and configure channel numbers, time slot strategies, and retransmission mechanisms;

[0242] The edge node that executes communication configuration synchronizes and writes to the communication configuration table;

[0243] Establish a three-dimensional linkage relationship among tasks, models, and communication parameters to support dynamic channel reconstruction.

[0244] The local inference and feedback module 50 is used to execute local inference tasks, complete the feedback of inference results, and generate a feedback data structure.

[0245] Load the corresponding model to execute local inference tasks;

[0246] Perform normalization, quantization, and data packet encapsulation processing on the monitoring results;

[0247] Call the communication configuration for asynchronous data feedback and automatic retransmission;

[0248] Generate a feedback data structure, including fields such as inference tags, confidence levels, and communication parameters, for use by the feedback module.

[0249] The data feedback module 60 is used to generate a task weight policy and a node resource score based on the feedback data structure, and update the task resource mapping.

[0250] Execute feature extraction such as inference confidence levels, multi-node consistency, and communication success rates;

[0251] Aggregate task tags and model response metrics to identify abnormal nodes;

[0252] Output a task priority adjustment policy and a node capability level tag;

[0253] Write the updated content into the task resource mapping structure to drive the next round of task-resource rescheduling.

[0254] Through the integrated design of the above system structure and functional modules, the present invention achieves the following remarkable beneficial effects in practical applications:

[0255] Full-process closed-loop control: The system constructs a full-life-cycle closed-loop structure of "task modeling → optimization scheduling → model deployment → communication configuration → local inference → feedback → mapping update", ensuring that each round of task execution can be adaptively evolved according to the feedback, improving the system operation stability and resource utilization efficiency.

[0256] Computing power and communication collaborative scheduling optimization: By introducing an improved Hungarian algorithm and a stability partial derivative mechanism in the S200 module, a global optimal matching of tasks and nodes is achieved under multi-dimensional constraints; the S400 module supports heterogeneous communication modes such as TDMA and LoRa, significantly reducing bandwidth occupancy and supporting large-scale multi-node data collaborative feedback.

[0257] Edge Intelligence Distributed Inference: Implement GNN model compression through the S300 module and edge deployment of the S500 module to achieve in-situ processing and intelligent inference of forest resource data, minimizing the central load and latency to meet the timeliness requirements of data analysis under the complex terrain of forest areas.

[0258] Adaptive Feedback and Dynamic Reconfiguration Mechanism: The S600 module constructs a scoring feedback mechanism with tasks and nodes as two dimensions, combined with a mapping version update mechanism, to ensure that task priorities and resource scheduling can be dynamically adjusted according to the environment, improving the system robustness.

[0259] Obviously, the embodiments described above are only a part of the embodiments of this application, rather than all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosed content of this application more thorough and comprehensive. Although this application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions described in the foregoing specific embodiments, or perform equivalent replacements for some of the technical features. Any equivalent structures directly or indirectly using the content of the specification and drawings of this application in other related technical fields are equally within the scope of the patent protection of this application.

Claims

1. A multi-level forest resource monitoring method, characterized in that, including: Construct a task-resource mapping by normalizing the status information of each edge node and the forest resource monitoring task set; Construct a scheduling optimization model and perform iterative matching by obtaining the task-resource mapping to generate a matching allocation structure; The expression of the scheduling optimization model is: Among them, is the task and the comprehensive adaptation score of the edge node ; is the task-node matching set formed by the current scheduling strategy; is the scheduling conflict cost function; is the conflict penalty factor; is the comprehensive matching score of the scheduling strategy; The matching allocation structure includes a model deployment priority score: Among them, is the deployment priority score of the task-node pair; is the node current scheduling offset risk indicator; Compress the central GNN model according to the matching allocation structure and deploy it to the edge nodes, construct a model deployment structure, and complete registration marking and version binding; Select a communication mode and configure a communication channel according to the model deployment structure and communication status information, and complete the synchronization and storage of the communication configuration structure; Execute the local inference task of the model deployment structure, complete the encoding and asynchronous transmission back of the inference result, and generate a transmission-back data structure; Extract features and aggregate data from the transmission-back data structure to generate a task weight policy and a node resource score, and feedback them to the task-resource mapping to update the configuration policy.

2. The multi-level forest resource monitoring method according to claim 1, wherein, Constructing a task-resource mapping includes: Normalize the computing power status information, storage status information, and communication status information of each edge node to generate edge node status information; Structurally process the forest resource monitoring task set according to task granularity, task priority, and computational complexity to generate task set information; Construct a task-resource mapping based on the edge node status information and the task set information.

3. The multi-level forest resource monitoring method according to claim 1, wherein Constructing a scheduling optimization model includes: Obtain the task-resource mapping and construct a scheduling optimization model based on the edge node status information and the task set information; Execute the task scheduling matching operation to generate a scheduling intermediate result; Perform iterative optimization and constraint rearrangement based on the scheduling intermediate result to update the scheduling result.

4. The multi-level forest resource monitoring method according to claim 3, wherein, Generating a matching allocation structure includes: Extract the optimal allocation relationship based on the iteratively optimized scheduling result to construct a matching allocation structure; Generate an edge node task execution list according to the matching allocation structure.

5. The multi-level forest resource monitoring method according to claim 1, wherein Constructing a model deployment structure includes: Obtain the matching allocation structure, perform compression processing on the central graph neural network model to generate a lightweight model; Deploy the lightweight model to the edge nodes according to the matching allocation structure to construct a model deployment structure; Perform registration marking on the model deployment structure and bind version information.

6. The multi-level forest resource monitoring method according to claim 1, wherein Configuring a communication channel includes: Obtain the model deployment structure and communication status information, and select a communication mode according to the communication quality level; Construct a communication configuration structure and allocate communication channel resources; Complete the edge node synchronization process of the communication configuration structure and store the configuration result in the communication configuration table.

7. The multi-level forest resource monitoring method according to claim 1, wherein Executing the local inference task includes: Obtain the model deployment structure deployed on the edge nodes, and call the lightweight model to execute the local inference task; Obtain the inference result data and encode the data according to the communication configuration structure; Asynchronously transmit back the inference result through the communication configuration structure to generate a transmission-back data structure.

8. The multi-level forest resource monitoring method according to claim 1, wherein Updating the configuration policy includes: Obtain the transmission-back data structure, perform global feature extraction and multi-node data aggregation; Generate a task weight policy and a node resource score; Feedback the task weight policy and the node resource score to the task-resource mapping to complete policy update.

9. The multi-level forest resource monitoring method according to claim 5, wherein The graph neural network model is compressed by the knowledge distillation method, and the communication configuration structure is constructed by the time division multiple access mechanism.

10. A multi-level forest resource monitoring system, applied to the multi-level forest resource monitoring method described in any one of claims 1-9, characterized in that, including: A task resource construction module, which is used to obtain edge node status information and a forest resource monitoring task set, and construct a task resource mapping; A task scheduling module, which is used to perform scheduling optimization based on the task resource mapping and generate a matching allocation structure; A model compression and deployment module, which is used to generate a lightweight model according to the matching allocation structure and deploy it to edge nodes; A communication configuration module, which is used to obtain a model deployment structure and communication status information, construct a communication configuration structure and complete channel configuration; A local inference and feedback module, which is used to execute local inference tasks and complete the feedback of inference results, and generate a feedback data structure; A data feedback module, which is used to generate a task weight policy and a node resource score based on the feedback data structure, and update the task resource mapping.

Images