Intelligent agent increment training system based on agricultural Internet of Things edge computing equipment operation feedback

Abstract

The invention discloses an agent increment training system based on agricultural Internet of Things edge computing equipment operation feedback, and relates to the technical field of agricultural artificial intelligence decision optimization. Comprising an agent decision module, an execution module, a data acquisition module, an automatic labeling engine module and a model increment training module which are sequentially connected to form a closed loop. According to the method, the problems of automatic data annotation and confidence evaluation in agricultural decision model iteration are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural artificial intelligence decision optimization, and more particularly to an agent incremental training system based on operation feedback of an agricultural Internet of Things edge computing device. BACKGROUND

[0002] Smart agriculture, also known as digital agriculture or information agriculture, involves various agricultural high-techs such as rapid acquisition of agricultural information, farmland planting, land management, pesticide utilization, pollution control, agricultural engineering equipment and its industrialization technology. Information management, as an important branch of computer application field, is an important part of smart agriculture and the basis of intelligent decision-making in agricultural production. Smart agriculture uses intelligent machines to collect the necessary data for agricultural production, which can greatly improve the efficiency of agricultural production, save labor costs and save resources.

[0003] Smart agriculture promotes the transformation of agricultural production from traditional agriculture to modern agriculture, making agriculture more intelligent and having a profound and significant impact on people's lives. The main application fields of smart agriculture are intelligent greenhouse, water-saving irrigation, intelligent cultivation control and aquaculture environment monitoring.

[0004] The current agricultural intelligent system has the following defects:

[0005] One-way decision-making defect: after the intelligent agent provides fertilization / irrigation suggestions, there is a lack of quantitative tracking of the actual execution effect;

[0006] High labeling cost: crop response data (such as growth changes) rely on manual labeling, which is inefficient and highly subjective;

[0007] Lack of confidence: environmental interference (meteorological mutations, sensor errors) makes it difficult to evaluate the credibility of feedback data, hindering model iteration.

[0008] Therefore, the agent incremental training system based on operation feedback of an agricultural Internet of Things edge computing device is proposed to solve the problems existing in the prior art, which is a problem that needs to be solved by the skilled in the art. SUMMARY

[0009] Therefore, the agent incremental training system based on operation feedback of an agricultural Internet of Things edge computing device is proposed to solve the problems existing in the prior art, which is a problem that needs to be solved by the skilled in the art.

[0010] In order to achieve the above purpose, the present application provides the following technical solutions:

[0011] An agent incremental training system based on operation feedback of an agricultural Internet of Things edge computing device, comprising an agent decision module, an execution module, a data acquisition module, an automatic annotation engine module and a model incremental training module connected in turn to form a closed loop;

[0012] The agent decision module is configured to output a farming scheme carrying model fingerprint information and operation trajectory information generated by the execution record;

[0013] The execution module is configured to execute the instructions of the agent decision module;

[0014] The data acquisition module is configured to acquire multi-source data;

[0015] The automatic annotation engine module is configured to generate high-quality training labels, multi-modal data conflict arbitration, data confidence evaluation and filtering, and sample management;

[0016] The model incremental training module is configured to realize dynamic iteration and adaptive optimization of model incremental training triggering.

[0017] Optionally, the model fingerprint information in the agent decision module includes a model version and a rule set version number, and the operation trajectory information includes spatial coordinates, a timestamp and environmental parameters.

[0018] Optionally, the automatic annotation engine module includes a rule loader, a conflict arbitrator, a confidence evaluator and a sample shunt.

[0019] The rule loader is configured to pull a special rule set from a warehouse according to a model ID;

[0020] The conflict arbitrator is configured to perform multi-level conflict arbitration including weighted voting, multi-source data review and wavelet fusion.

[0021] The confidence evaluator is based on five-dimensional confidence evaluation and filters low-quality annotations.

[0022] The sample shunt is configured to dynamically shunt a threshold, when conf is greater than or equal to 0.75, incremental training is performed, when 0.6 is less than or equal to conf and conf is less than 0.75, a buffer pool is entered, and when conf is less than 0.6, rule self-checking is performed.

[0023] Optionally, the conflict arbitrator performs a multi-level conflict arbitration mechanism, including the following steps:

[0024] (1) Conflict detection and data request: when receiving a preliminary annotation result from the annotation engine, the conflict arbitrator initiates a data supplement process, including: initiating a resampling instruction to the data acquisition module to require multi-spectral image resampling at a specific angle or time; synchronously calling a historical database and a satellite remote sensing data source for cross verification;

[0025] (2) Dynamic rule loading and arbitration execution: the latest arbitration rule set is loaded in real time through the rule loader, and the arbitration rule set is dynamically adjusted based on the current crop growth period;

[0026] (3) Confidence weight generation: the output weighted confidence value after arbitration is completed, and the calculation formula is:

[0027] Arbitration confidence = data consistency score x 0.7 + rule matching degree x 0.3;

[0028] (4) Abnormal processing mechanism: for samples that have not been arbitrated for 3 times in succession, a rule self-checking process is automatically triggered and recorded to the blockchain audit module.

[0029] Optionally, the confidence evaluator filters low-quality labels based on the five-dimensional confidence evaluation comprehensive confidence:

[0030] Comprehensive confidence = 0.25 x sensor accuracy score + 0.30 x environmental interference score + 0.20 x

[0031] Data integrity score + 0.15 x crop stage score + 0.10 x historical consistency.

[0032] Optionally, the model incremental training module is used to realize the conditions for triggering model incremental training, including data-driven channel triggering, agricultural time rhythm channel triggering, and abnormality-driven channel triggering.

[0033] According to the above technical solution, compared with the prior art, the present application provides an intelligent agent incremental training system based on agricultural Internet of Things edge computing device operation feedback, which has the following beneficial effects:

[0034] (1) The embodiment of the present application in 1200 field test samples realizes the dual improvement of annotation efficiency and accuracy. Under the same sample size, the average time of automatic annotation is compressed from 3.5 hours / sample of manual annotation to 0.28 hours / sample, with an efficiency improvement of 92.0%; through the multi-source data arbitration mechanism, the confidence of the annotation result reaches 95.3%, the error rate is controlled below 4.7%, and the error rate is reduced by 63.2% compared with the traditional single-source annotation method (error rate 18.2%);

[0035] (2) The model iteration efficiency of the present application is significantly optimized, and the full-amount iteration period of the model is shortened from 3 months in the traditional mode to 2 weeks, with an iteration rate improvement of 87.0%; combined with the dynamic confidence filtering mechanism (conf> 0.75 samples are included in the training), the proportion of invalid training samples is reduced from 28.6% to 8.6%, the effective training resource utilization rate is improved by 70.0%, and the model parameter convergence speed is accelerated by 42.3%. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute a part of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of the provided drawings.

[0037] Figure 1 A structure block diagram of an intelligent agent incremental training system based on operation feedback of an agricultural Internet of Things edge computing device is provided for the present application.

[0038] Figure 2 An automatic labeling engine module structure schematic diagram is provided for the present application.

[0039] Figure 3 A system flow schematic diagram of embodiment 1 is provided for the present application.

[0040] Figure 4 A collaborative working mechanism schematic diagram of five-dimensional confidence assessment is provided for the present application.

[0041] Figure 5 An incremental training triggering mechanism schematic diagram is provided for the present application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] Referring to Figure 1 As shown in the drawings, the present application discloses an intelligent agent incremental training system based on operation feedback of an agricultural Internet of Things edge computing device, which comprises an intelligent agent decision module, an execution module, a data acquisition module, an automatic labeling engine module and a model incremental training module connected in sequence to form a closed loop.

[0044] The intelligent agent decision module is used to output operation trajectory information generated by carrying model fingerprint information and execution record when outputting a farming scheme.

[0045] The execution module is used to execute the instructions of the intelligent agent decision module.

[0046] The data acquisition module is used to acquire multi-source data.

[0047] The automatic labeling engine module is used to generate high-quality training labels, multi-modal data conflict arbitration, data confidence assessment and filtering, and sample management.

[0048] The model incremental training module is configured to implement dynamic iteration and adaptive optimization of model incremental training triggering.

[0049] Further, the model fingerprint information in the intelligent agent decision module includes model version and rule set version number, and the operation trajectory information includes spatial coordinates, time stamp and environmental parameters.

[0050] Further, the automatic labeling engine module includes a rule loader, a conflict arbitrator, a confidence evaluator and a sample shunt;

[0051] The rule loader is configured to pull a special rule set from the warehouse according to the model ID;

[0052] The conflict arbitrator is configured to perform multi-level conflict arbitration including weighted voting, multi-source data review and wavelet fusion;

[0053] The confidence evaluator is based on five-dimensional confidence evaluation to filter low-quality labeling;

[0054] The sample shunt is configured to dynamically shunt the threshold, when conf is greater than or equal to 0.75, incremental training is performed, 0.6 is less than or equal to conf and conf is less than 0.75, and the rule self-check is performed.

[0055] Further, the conflict arbitrator performs a multi-level conflict arbitration mechanism, including the following steps:

[0056] The conflict arbitrator performs a multi-level conflict arbitration mechanism, including the following steps:

[0057] (1) Conflict detection and data request: when receiving the preliminary labeling result from the labeling engine, the conflict arbitrator initiates a data supplement process, including: initiating a resampling instruction to the data acquisition module, requiring multi-spectral image resampling at a specific angle or time; synchronously calling the historical database and satellite remote sensing data source for cross verification;

[0058] (2) Dynamic rule loading and arbitration execution: the latest arbitration rule set is loaded in real time through the rule loader, and the arbitration rule set is dynamically adjusted based on the current crop growth period;

[0059] (3) Confidence weight generation: after arbitration, the weighted confidence value is output, and the calculation formula is:

[0060] Arbitration confidence = data consistency score x 0.7 + rule matching degree x 0.3;

[0061] (4) Abnormal processing mechanism: for samples that have not been decided for three times in succession, the rule self-check process is automatically triggered and recorded to the blockchain audit module.

[0062] Specifically, the blockchain audit module is a security audit and traceability module built on distributed ledger technology, used to immutably record arbitration anomalies and subsequent processing.

[0063] The blockchain audit module includes a data on-chain structure, automated smart contract processing, cross-chain verification mechanisms, and audit traceability interfaces.

[0064] In the data on-chain structure, the audit event is first encapsulated by encapsulating the complete context parameters of the abnormal arbitration sample into a structured audit event. The parameters include the sample ID, the set of conflict annotation results, the trigger timestamp, the data supplementation record, and the number of historical arbitrations. The event hash value is then calculated as a unique digital fingerprint. Subsequently, the audit event is broadcast to the verification nodes of the agricultural data audit consortium chain through the Practical ByzantineFault Tolerance consensus algorithm. After verification by a majority of nodes, it is packaged into a new block to achieve distributed evidence storage.

[0065] In the automated processing of smart contracts, a self-inspection process is deployed to trigger the contract. When three consecutive pending arbitration events are detected on the chain, the pre-set smart contract logic is automatically executed. This includes sending a rule self-inspection instruction to the rule management platform and locking the subsequent arbitration process for that sample. At the same time, an expert review task sheet with a priority identifier is generated and pushed to the manual audit queue. In addition, the key nodes of the self-inspection process are recorded by tracking the contract through the audit trajectory, including rule verification results, manual review conclusions, and system parameter adjustment records, forming a verifiable processing trajectory chain.

[0066] In the cross-chain verification mechanism, the Merklegen hash value of the blockchain audit module is first periodically anchored to the public blockchain through the time anchoring service to enhance the legal validity of the audit data; at the same time, it supports third-party audit institutions to verify the compliance of the arbitration exception handling process through the zk-SNARK protocol, so as to prove the correct execution of the processing logic without disclosing the original data.

[0067] The audit tracing interface provides a standardized API query interface, which supports the retrieval of historical audit records by multiple dimensions such as sample ID, time range, and confidence threshold, and generates an encrypted and verifiable audit report. All query results are accompanied by digital signatures and timestamps, which comply with the judicial electronic evidence preservation standards.

[0068] Specifically, the system adopts a dynamic weight rebalancing mechanism. When any of the modules of weighted voting, multi-source data verification, and wavelet fusion fails, the system will dynamically redistribute the original fixed weight value of the failed module to the remaining effective modules according to a preset ratio. For example, if wavelet fusion fails, weighted voting will absorb 60% of the wavelet weight, and multi-source data verification will absorb 40% of the wavelet weight.

[0069] Furthermore, the confidence estimator evaluates the overall confidence level of filtering low-quality labels based on five-dimensional confidence levels as follows:

[0070] Overall confidence level = 0.25 × sensor accuracy score + 0.30 × environmental interference score + 0.20 ×

[0071] Data integrity score + 0.15 × crop stage score + 0.10 × historical consistency score.

[0072] Specifically, the weighting of the five-dimensional scoring system is based on the results of multiple regression analysis of 2,000 sets of field trial data. The analysis fits the mathematical relationship between each indicator and the target output, and quantifies the contribution and weight of each dimension accordingly.

[0073] Furthermore, the incremental model training module is used to implement the conditions for triggering incremental model training, including data-driven channel triggering, agricultural rhythm channel triggering, and anomaly-driven channel triggering.

[0074] Specifically, for the data-driven channel: the formula for calculating the sample size threshold is as follows:

[0075] Agricultural rhythm channel: Training is forcibly triggered when the seedling emergence rate reaches 90%, the leaf age index is >3.5, within 24 hours after flowering, and the moisture content drops to 23%.

[0076] Anomaly-driven channel: Emergency training is initiated when the model prediction error rate is >15% for 3 consecutive days or when there is continuous high temperature / heavy rain.

[0077] Example 1

[0078] This invention includes an agent decision-making module, an execution module, a data acquisition module, an automated annotation engine module, and a model incremental training module, which are connected in sequence to form a closed loop.

[0079] like Figure 3 As shown, the agent generates an agricultural action plan for the current agricultural scenario and outputs the plan along with its model ID to the execution system. The execution system executes the plan, and the data acquisition module then collects the environmental feedback data after execution and stores it in association with the model ID.

[0080] The collected feedback data is sent to the annotation engine for automated preliminary annotation, producing annotation results that may contain conflicts. When the conflict arbitrator determines that the current data quality is insufficient or that annotation conflicts are difficult to resolve automatically, it will initiate a resampling or supplementary data request. The data acquisition module responds to this request and provides multi-source verification data.

[0081] The annotation engine and arbitrator submit the processed samples to the confidence estimator. The confidence estimator performs a comprehensive analysis of the samples and calculates an accurate confidence score.

[0082] The system sets a confidence threshold and splits the samples based on the evaluation results:

[0083] High-confidence samples (conf≥0.75): The system considers these samples to have extremely high annotation quality and strong reliability. These samples are directly added to the training pool of the training system for subsequent incremental model training.

[0084] Medium / low confidence samples (conf < 0.75): The system considers the automatically labeled results for these samples unreliable. To avoid introducing incorrect labels that contaminate the model, the system automatically initiates a manual verification process. This process pushes the samples to a manual labeling platform for final expert review. High-quality samples that have been manually verified and corrected are then added to the training pool.

[0085] The training system uses newly added high-quality samples (from high-confidence samples and manually validated samples) in the training pool to optimize old models in the model repository through incremental training and active learning strategies, generating new versions of models with stronger performance. The new models are deployed back to the agent, thus starting the next round of optimization cycle, ultimately achieving continuous, reliable, and efficient autonomous evolution of the agent's decision-making capabilities.

[0086] like Figure 2 As shown, the automated annotation engine module is used to generate high-quality training labels, arbitrate multimodal data conflicts, evaluate and filter data confidence, and manage samples. The automated annotation engine module includes a rule loader, a conflict arbitrator, a confidence evaluator, and a sample splitter.

[0087] like Figure 4 As shown, the confidence evaluator, based on five-dimensional confidence assessment, filters out low-quality labeled content as follows:

[0088] Confidence evaluators include sensor accuracy, environmental interference, data integrity, crop stage, and historical consistency.

[0089] The calculation process for sensor accuracy is as follows: Multispectral camera basic accuracy 0.9 (factory error 7%) × operating condition coefficient 1.0 (temperature 28℃), score 0.90;

[0090] The calculation process for environmental disturbance is as follows: In the absence of extreme weather, env_score() returns 1.0, resulting in a score of 1.00;

[0091] The data integrity calculation process was as follows: missing soil pH values ​​were filled in using the Sentinel-2 satellite, with a score of 0.8;

[0092] The calculation process for crop stages is as follows: the current stage is the tasseling stage (the model requires the jointing stage), the stage score is 0.6 × parameter correlation score 1.0 × conversion accuracy score 0.8, and the score is 0.48;

[0093] The historical consistency calculation process is as follows: current NDVI increase Δ = 0.15, historical average Δ = 0.12 (Z-score = 1.2), segment score 0.4, and score 0.40.

[0094] Overall confidence level = 0.25 × 0.90 + 0.30 × 1.00 + 0.20 × 0.80 + 0.15 × 0.48 + 0.10 × 0.40 = 0.782

[0095] like Figure 5 As shown, incremental training is triggered when the confidence level is ≥0.75 (0.782), and the sample enters the training pool. When the data accumulation reaches a threshold (e.g., 200 field crop data), lightweight training is triggered.

[0096] Example 2

[0097] An embodiment of the present invention on closed-loop optimization of pest and disease identification and response in the context of rice pest and disease control includes the following:

[0098] In this embodiment, the intelligent agent decision-making module outputs a spraying plan (model version v2.3, rule set RS rice v1.2) for a certain decision, carrying an operational trajectory digital twin (coordinates: 30.5°N, 120.3°E; timestamp: 2025-06-15 10:00; environmental parameters: temperature 28°C, humidity 85%).

[0099] Afterwards, the execution module controls the drone to spray fungicide, covering the target rice paddy area.

[0100] The data acquisition module collects multi-source data 72 hours after spraying:

[0101] (1) Multispectral image (resolution 0.5m, bands including Red Edge and NIR);

[0102] (2) Soil moisture sensor data (accuracy ±3%);

[0103] (3) Real-time rainfall records from meteorological stations.

[0104] The automated annotation engine module detected a multimodal data conflict (visible light images showed a reduction in lesions, but thermal infrared images showed an abnormally high leaf temperature), and the conflict arbitrator triggered a multi-level arbitration mechanism:

[0105] (1) Send a resampling command to the data acquisition module to re-acquire multispectral images at noon;

[0106] (2) Cross-validation was performed by calling the historical database (data from the same period last year) and Sentinel-2 satellite data, and it was found that humidity interference caused thermal infrared errors;

[0107] (3) Dynamically load the special rule set for rice heading period, and calculate the arbitration confidence level as: data consistency score 0.9×0.7 + rule matching degree 1.0×0.3=0.93.

[0108] The confidence estimator calculates the overall confidence level:

[0109] Sensor accuracy score 0.85 (camera operating condition coefficient 0.95), environmental interference score 0.7 (rainfall impact), data integrity score 1.0, crop stage score 0.9 (heading stage), historical consistency score 0.8 → overall confidence score = 0.25×0.85 + 0.30×0.7 + 0.20×1.0 + 0.15×0.9 + 0.10×0.8 = 0.815. The sample splitter determines that conf = 0.815 ≥ 0.75, and the sample enters the training pool.

[0110] The incremental training module for the model was triggered by the data-driven channel (the sample size reached the threshold Nthreshold = 100 + 50 × 0.6 ≈ 138), and started lightweight training.

[0111] Update the weights of the pest and disease identification model to reduce the false positive rate due to humidity interference.

[0112] Results: The system reduced the annotation error rate from the initial 12% to 4.5%, shortened the model iteration cycle to 10 days, and improved efficiency by 90% compared with traditional methods.

[0113] Example 3

[0114] Another embodiment of the present invention, targeting the triggering mechanism of an abnormal driving channel under a sudden drought scenario in the wheat-growing area of ​​the Huang-Huai Plain, includes the following:

[0115] In this embodiment, the agent decision-making module generates an irrigation plan (model fingerprint: v3.1, rule set RS wheat v2.0) for a certain decision, and the digital twin of the operation trajectory includes the soil moisture content threshold (warning value <18%).

[0116] Subsequently, the execution module started the drip irrigation system, but the data acquisition module detected that the soil moisture content dropped to 15% for three consecutive days (sensor error ±2%).

[0117] The automated annotation engine module detected an anomaly:

[0118] The confidence evaluator assessed that the environmental interference score dropped sharply to 0.4 (due to sensor drift caused by high temperature of 40℃). The overall confidence level = 0.25×0.8 + 0.30×0.4 + 0.20×0.9 + 0.15×0.7 + 0.10×0.6 = 0.625 → the sample enters the buffer pool.

[0119] The conflict arbitrator used satellite remote sensing data (Landsat-9) for verification, and the arbitration confidence level increased to 0.78. The model incremental training module was triggered by an abnormal driving channel (model prediction error rate >18% for 3 consecutive days), and emergency training was initiated: dynamically adjusting the irrigation decision model and optimizing the water stress response rules.

[0120] Results: Based on statistical analysis of three typical drought monitoring areas (sample size n=120 per area), the system completed model iteration within 48 hours, improving irrigation efficiency by 35% compared to before optimization (95% confidence interval: 31.2%–38.8%, p<0.01). By introducing a dynamic threshold screening mechanism, the proportion of invalid training decreased from 28.6% before optimization to 7.1%, a relative reduction of 75% (χ2=42.36, p<0.001). Combined with the improved accuracy of soil moisture prediction (RMSE decreased from 2.3% to 1.7%), water waste decreased by 22% compared to the baseline (95% confidence interval: 18.5%–25.5%, t=6.82, p<0.001). All indicators passed the statistical significance test, verifying the effectiveness of incremental model training.

[0121] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0122] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An incremental training system for intelligent agents based on operational feedback from edge computing devices in agricultural Internet of Things (IoT) systems, characterized in that, It includes an agent decision-making module, an execution module, a data acquisition module, an automated labeling engine module, and a model incremental training module that are connected in sequence to form a closed loop; The intelligent agent decision-making module is used to output agricultural plans by carrying model fingerprint information and operation trajectory information generated from execution records; The execution module is used to execute the instructions of the agent's decision-making module; The data acquisition module is used to collect data from multiple sources. The automated annotation engine module is used to generate high-quality training labels, arbitrate multimodal data conflicts, evaluate and filter data confidence, and manage samples. The incremental model training module is used to implement dynamic iteration and adaptive optimization triggered by incremental model training.

2. The intelligent agent incremental training system based on operational feedback from agricultural IoT edge computing devices according to claim 1, characterized in that, The model fingerprint information in the intelligent agent decision-making module includes the model version and rule set version number, while the operation trajectory information includes spatial coordinates, timestamps, and environmental parameters.

3. The intelligent agent incremental training system based on operational feedback from agricultural IoT edge computing devices according to claim 1, characterized in that, The automated annotation engine module includes a rule loader, a conflict arbitrator, a confidence evaluator, and a sample splitter; A rule loader is used to pull a specific set of rules from the repository based on the model ID; A conflict arbitrator for multi-level conflict arbitration, including weighted voting, multi-source data verification, and wavelet fusion; The confidence estimator, based on five-dimensional confidence assessment, filters out low-quality labels. The sample splitter is used for dynamic splitting thresholds. When conf≥0.75, incremental training is performed; when 0.6≤conf<0.75, the sample enters the buffer pool; and when conf<0.6, rule self-checking is performed.

4. The intelligent agent incremental training system based on operational feedback from agricultural IoT edge computing devices according to claim 3, characterized in that, The conflict arbitrator implements a multi-level conflict arbitration mechanism, including the following steps: (1) Conflict detection and data request: When receiving the preliminary annotation results from the annotation engine, the conflict arbitrator actively initiates the data supplementation process, including: sending a resampling command to the data acquisition module to request the resampling of multispectral images at a specific angle or time; and simultaneously calling historical databases and satellite remote sensing data sources for cross-validation. (2) Dynamic rule loading and arbitration execution: The latest arbitration rule set is loaded in real time through the rule loader, and the arbitration rule set is dynamically adjusted based on the current crop growth period; (3) Confidence weight generation: After arbitration is completed, the weighted reset confidence value is output. The calculation formula is: Arbitration confidence = Data consistency score × 0.7 + Rule matching degree × 0.3; (4) Anomaly handling mechanism: For samples that have been unresolved for three consecutive arbitrations, the rule self-check process is automatically triggered and recorded in the blockchain audit module.

5. The intelligent agent incremental training system based on operational feedback from edge computing devices in agricultural Internet of Things according to claim 3, characterized in that, The confidence estimator evaluates the overall confidence level of filtering low-quality labels based on a five-dimensional confidence level as follows: Overall confidence level = 0.25 × sensor accuracy score + 0.30 × environmental interference score + 0.20 × data integrity score + 0.15 × crop stage score + 0.10 × historical consistency score.

6. The intelligent agent incremental training system based on operational feedback from agricultural IoT edge computing devices according to claim 1, characterized in that, The incremental model training module is used to implement the conditions for triggering incremental model training, including data-driven channel triggering, agricultural rhythm channel triggering, and anomaly-driven channel triggering.

Images