Artificial intelligence decision system for unmanned agricultural operation

Abstract

The invention relates to the technical field of agricultural intellectualization, in particular to an artificial intelligence decision-making system for unmanned agricultural operation, which comprises an acquisition module, a construction module, a monitoring module, an AI decision-making module, a data fusion decision-making module and an execution control module which are in mutual signal connection, the semantic analysis module is used for collecting peasant household interview records and farm work log text data and carrying out semantic analysis on unstructured texts by adopting a natural language processing technology; the construction module is used for receiving the empirical feature vector and constructing a multi-dimensional associated knowledge graph by adopting a graph neural network; and the AI decision module is used for receiving the environment feature matrix, training by adopting a deep reinforcement learning model in combination with a general agricultural data set, and generating a decision scheme with a confidence score. According to the method, collaborative decision-making of regional implicit knowledge and multi-source environment data is realized by fusing a dynamic weighting mechanism of a peasant experience knowledge graph and deep reinforcement learning.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agricultural technology, and in particular to an artificial intelligence decision-making system for unmanned agricultural operations. Background Art

[0002] With the increasing shortage of rural labor and the demand for large-scale agricultural operations, unmanned operations have become an inevitable trend in the development of modern agriculture. While traditional agricultural machinery automation systems can achieve mechanical control of basic operations such as sowing and fertilizing, they lack the ability to respond in real time to dynamic changes in the farmland environment in complex decision-making scenarios such as pest control and disaster response. Artificial intelligence technology, by integrating multi-source sensor data with crop growth models, can build autonomous decision-making systems, significantly improving operational efficiency. For example, variable-rate fertilization systems based on satellite remote sensing can reduce fertilizer use by 10%-15%, and deep learning-driven pest and disease identification models can increase detection accuracy to over 85%. These technologies enable unmanned systems to replace manual labor in the complete closed loop of environmental perception, analysis, decision-making, and execution, demonstrating unique advantages in scenarios such as nighttime operations and hazardous environments.

[0003] Existing AI agricultural decision-making systems, for example, Chinese patent publication No. CN119026932A proposes an intelligent irrigation decision-making method based on artificial intelligence algorithms, which obtains crop growth data, including at least: soil moisture, ambient temperature, light intensity, wind speed, air humidity and crop growth status; normalizes the feature vectors; selects important feature data and inputs them into a machine learning model to train the model to predict the immediate water demand of crops; inputs the prediction results of the machine learning model and the selected feature data into a deep neural network model for further optimization and prediction; introduces weather forecast data as additional input to formulate the optimal irrigation strategy and determine the irrigation time, frequency and water volume; and executes irrigation operations according to the optimal irrigation strategy.

[0004] Although this invention optimizes agricultural water use and achieves precise irrigation to avoid problems such as water waste and uneven crop growth, thereby improving agricultural production efficiency and crop quality, it also has defects: First, if it relies too much on general data sets to train models, it will ignore the regional implicit knowledge accumulated by farmers over a long period of time. For example, in hilly areas, the "water accumulation warning in the southeast corner of the slope" rule summarized by farmers through experience was not adopted by the AI ​​system due to the lack of structured expression, resulting in drainage decisions being out of touch with the actual terrain. Secondly, there is a lack of implicit knowledge collection and fusion mechanisms. Existing natural language processing technology can only extract explicit information from agricultural records, such as irrigation water volume, pesticide application time and dosage, but cannot analyze spatiotemporal correlation patterns such as "a certain pest develops resistance every two years", resulting in insufficient decision-making adaptability of the model in micro-geographical difference scenarios. Therefore, there is an urgent need to develop a new decision-making system that is compatible with data-driven and empirical knowledge to solve the problem of regional adaptability in unmanned agricultural operations. Summary of the Invention

[0005] To solve the above problems, the present invention provides an artificial intelligence decision-making system for unmanned agricultural operations, which realizes collaborative decision-making of regional implicit knowledge and multi-source environmental data by integrating farmers' experience knowledge graph and the dynamic weighting mechanism of deep reinforcement learning.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows: an artificial intelligence decision-making system for unmanned agricultural operations, comprising an acquisition module, a construction module, a monitoring module, an AI decision-making module, a data fusion decision-making module, and an execution control module, all of which are signal-connected to each other. The monitoring module is signal-connected to an external sensing unit, which is used to collect environmental data from several different channels of farmland, wherein:

[0007] The acquisition module is used to collect text data from farmer interview records and agricultural logs, use natural language processing technology to perform semantic analysis on unstructured text, generate empirical feature vectors containing regional implicit knowledge, and transmit them to the construction module;

[0008] The construction module is used to receive the empirical feature vector, construct a multi-dimensional associated knowledge graph using a graph neural network, generate a structured decision rule base containing spatiotemporal constraints, and transmit it to the data fusion decision module;

[0009] The monitoring module is used to receive multi-source heterogeneous environmental data transmitted by external sensing units, use wavelet transform to perform signal noise reduction, generate a standardized environmental feature matrix, and transmit it to the AI ​​decision module;

[0010] The AI ​​decision module receives the environmental feature matrix, uses a deep reinforcement learning model combined with a general agricultural dataset for training, generates a decision plan with a confidence score, and transmits it to the data fusion decision module;

[0011] The data fusion decision module is used to receive the structured decision rule base and decision plan, and adopts a dynamic weighting algorithm to perform multimodal data fusion: when the confidence score is lower than the preset threshold, the rule matching engine is activated to prioritize the execution of the regional disposal strategy in the knowledge graph; when the confidence score is higher than the threshold, the LSTM time series prediction model is used to optimize the AI ​​decision output, generate the final control instruction, and transmit it to the execution control module;

[0012] The execution control module is used to convert the final control instructions into operating parameters of agricultural equipment and send them to the field operation terminal through the LoRaWAN protocol.

[0013] Furthermore, the acquisition module includes a natural language processing unit, a semantic analysis unit, and a rule extraction unit, wherein:

[0014] The natural language processing unit is used to receive unstructured data such as farmers' voice interview recordings and handwritten farm log image data, and uses end-to-end speech recognition and optical character recognition technology to convert the unstructured data into standardized text data and transmit it to the semantic analysis unit;

[0015] A semantic analysis unit is used to perform dependency syntax analysis and named entity recognition on standardized text data, extract triple knowledge units containing geographic identifiers, crop varieties, and abnormal events, and transmit them to the rule extraction unit;

[0016] The rule extraction unit is used to convert the triple knowledge unit into a "condition-action" decision rule by adopting a causal reasoning model and a keyword matching algorithm, and to attach regional spatiotemporal labels to generate an empirical feature vector to be transmitted to the construction module.

[0017] Furthermore, the rule extraction unit includes a temporal correlation subunit, a spatial correlation subunit, and a visualization verification subunit, wherein:

[0018] The time series association subunit uses the LSTM-Attention model to analyze the periodic patterns in agricultural logs and generate time-dependent decision rules;

[0019] The spatial association subunit extracts the association rules between micro-topographic features and agricultural operations through Voronoi diagram segmentation and generates geographic spatial constraints;

[0020] The visualization verification sub-unit generates an explanatory map of the extracted time-dependent decision rules and geographic spatial constraints for farmers to confirm. If the farmer's feedback conflicts with the rule logic, the rule weight downgrade mechanism is triggered.

[0021] Furthermore, the building block includes an entity disambiguation unit, a graph association unit, and a rule verification unit, wherein:

[0022] The entity disambiguation unit is used to receive the empirical feature vector, eliminate the semantic conflicts of homonymous agricultural operations by matching geographic coordinates and aligning crop growth cycles, and generate a standardized empirical vector with geographic tags;

[0023] The graph association unit is used to receive standardized experience vectors and use the TransR algorithm to establish cross-modal association edges between pest and disease cycle laws, meteorological factors, and agricultural operations, generating intermediate knowledge graph data containing spatiotemporal constraints.

[0024] The rule verification unit is used to compare the intermediate data of the knowledge graph with the external historical agricultural operation database, correct the associated edge weights that are inconsistent with the measured results through the back-propagation algorithm, and output a structured decision rule library that matches the actual agricultural conditions in the region. The structured decision rule library contains spatiotemporal constraints including rule priority coefficients and effective geographical scope.

[0025] Furthermore, the monitoring module includes a data alignment unit, an anomaly detection unit, and a feature encoding unit, wherein:

[0026] The heterogeneous data alignment unit is used to receive heterogeneous data transmitted by external sensing units, including drone remote sensing images, soil sensor data, and weather station data, and generate environmental monitoring data in a unified coordinate system by performing spatiotemporal registration. The data is then transmitted to the anomaly detection unit and feature encoding unit respectively.

[0027] The anomaly detection unit uses the isolation forest algorithm to identify sensor unit failures or abnormal values ​​in environmental monitoring data. If an abnormal environmental monitoring data is detected, the redundant device switching mechanism is triggered and the data is re-collected;

[0028] The feature coding unit uses wavelet transform to decompose the noise-reduced environmental monitoring data, extracts frequency domain features, and combines them with the autoencoder for dimensionality compression to generate a standardized environmental feature matrix.

[0029] Furthermore, the AI ​​decision-making module includes an environmental feature analysis unit, an incremental learning unit, a confidence assessment unit, and an adversarial training unit, where:

[0030] The environmental feature parsing unit is used to receive the environmental feature matrix, remove redundant dimensions using a recursive feature elimination algorithm, and generate an environmental state vector that matches the input dimension of the reinforcement learning model;

[0031] The incremental decision unit receives the environment state vector and the regional rules updated by the knowledge graph, encodes the regional rules as constraints of the policy network, and dynamically updates the policy network parameters through the online policy gradient algorithm, taking the environment state vector as input and the regional rules as constraints, and outputs a preliminary decision plan containing micro-geographic adaptation parameters;

[0032] The confidence evaluation unit is used to receive the environment state vector and the preliminary decision plan, perform 300 forward propagation sampling on the preliminary decision plan using the Monte Carlo Dropout method, calculate the probability distribution variance of each decision action, and generate a confidence score on a scale of 0-100%. When the score is lower than 60%, a rule base intervention flag is added to the preliminary decision plan; when the score is higher than 60%, a decision plan with a confidence score is generated and transmitted to the data fusion decision module;

[0033] The adversarial training unit is used to input the environmental state vector into the GAN generator, synthesize the adversarial sample vector with superimposed extreme climate characteristics, deduce the adversarial sample vector through the policy network, generate alternative decision plans and inject them into the confidence evaluation unit to train the dataset.

[0034] Furthermore, the data fusion decision module includes a weight allocation unit, a conflict resolution unit and a decision tracing unit, wherein:

[0035] The weight allocation unit is used to receive the structured decision rule base and the decision plan, and analyze the geographical scope constraints in the structured decision rule base and the spatial coordinates in the decision plan. When the confidence score is less than 60% and the geographical matching degree is greater than 85%, it generates a rule-dominant flag and triggers the rule weight to be increased to 80%, and outputs the fused decision intermediate data including the weight allocation coefficient;

[0036] The conflict resolution unit is used to receive the intermediate data of the fusion decision and extract the rule instructions. When it detects that the rule instructions conflict with the AI ​​decision, it extracts the historical cases of farmers that match the current geographic tag in the structured rule base and calculates the similarity of environmental features:

[0037] Similarity = body 6 × cos (environmental feature matrix, historical cases) +

[0038] Body·4×(1-|current date-historical case date| / growth cycle length);

[0039] If the similarity is greater than 75% and there is a rule-dominant flag, a forced switch instruction is generated and the AI ​​decision flow is terminated;

[0040] The decision optimization unit is used to perform differentiated processing based on the conflict resolution results. When a forced switching instruction is received, the control parameters of the corresponding entry in the structured rule base are directly called. When the confidence score is ≥60%, the preliminary decision plan is input into the LSTM time series prediction model and the current crop growth period parameters are loaded. A temporal convolutional network is used to extract the periodic characteristics of the meteorological data of the past 30 days. The attention mechanism is used to weight recent environmental characteristics and output the final control instruction with a time series optimization label.

[0041] The decision tracing unit is used to write the geographic matching degree, weight distribution coefficient and forced switching instruction status code into the blockchain while generating the final control command, and generate a verifiable audit log containing a three-dimensional geographic hash value based on the timestamp and spatial coordinates contained in the timing optimization tag in the final control command.

[0042] Furthermore, the weight distribution unit outputs a weight distribution coefficient in the following manner: constructing a dynamic influencing factor set including soil type, crop growth period and weather anomaly index, wherein the dynamic influencing factor set is constructed by defining an influencing factor F={}, where is soil type, and the weight distribution is: sandy soil=0.2, clay=0.5, loam=0.3; is crop growth period, and the weight distribution is: seedling stage=0.4, jointing stage=0.6, maturity stage=0.8; is weather anomaly index, and the weight is expressed as the number of consecutive drought days / 30;

[0043] The fuzzy comprehensive evaluation method is used to calculate the rule applicability coefficient of each factor; the applicability coefficient is mapped to the fusion weight of the knowledge rule through the sigmoid function.

[0044] Furthermore, the conflict resolution unit includes an experience priority subunit and a dynamic correction subunit, wherein:

[0045] The experience priority sub-unit is used to retrieve successful cases with the same geographic identifier in the farmer's historical operation records when the rule instruction conflicts with the AI ​​decision and the confidence score is less than 60%. It matches the current environmental feature matrix with the historical scenario data. If the matching degree is ≥ 70% and the rule-dominant flag is present, it generates an experience-forced instruction to override the AI ​​decision and sends an alarm log containing the conflict field to the decision tracing unit.

[0046] The dynamic correction subunit is used to receive field execution effect data in real time after executing the experience-based forced instructions, calculate the deviation between the actual effect and the expected value of the rule base, and send a rule weight degradation instruction to the construction module and trigger a graph reconstruction event when the deviation is greater than 15% for 3 hours.

[0047] Furthermore, the execution control module includes an instruction compilation unit, a priority unit, and a security rehearsal unit, wherein:

[0048] The instruction compilation unit is used to convert the final control instructions into ISO 11783 standard agricultural machinery control messages, which are compatible with agricultural equipment of multiple brands;

[0049] The priority unit dynamically adjusts the collaborative operation sequence of several agricultural equipment according to the response delay of agricultural equipment and the urgency of the operation;

[0050] The safety rehearsal unit uses digital twin technology to simulate the field execution effect of control instructions. If a conflict in the motion trajectory of agricultural equipment is detected, the abnormal instruction will be intercepted and the path will be replanned.

[0051] The above scheme has the following beneficial effects:

[0052] 1. This solution, by constructing a knowledge graph based on farmers' experience, transforms the scattered, regional, implicit knowledge of traditional farming practices (such as micro-terrain response strategies and pest mutation patterns) into a structured rule base, overcoming the limitations of traditional AI systems that rely solely on general datasets. The two-way interaction between the knowledge graph and a deep reinforcement learning model enables the system to absorb farmers' accumulated experience in spatiotemporal constraints and dynamically optimize decision-making logic through online learning, significantly improving decision-making adaptability in complex farmland scenarios.

[0053] 2. This solution, based on a dual threshold judgment mechanism of confidence score and geographic match, prioritizes regional response strategies in scenarios where AI decision-making confidence is low, effectively avoiding the decision-making biases often associated with traditional data-driven models that ignore micro-geographical differences. The conflict resolution unit quantifies the timeliness of farmer experience using spatiotemporal decay factors and, in combination with genetic sequencing data, tracks pest variation in real time. This creates a closed-loop "decision-verification-correction" feedback loop, ensuring the scientific and timely nature of empirical rules.

[0054] 3. This solution combines a visual verification mechanism with blockchain traceability technology, enabling farmers to intuitively participate in rule verification and decision tracing. The interpretable graph generated by the rule extraction unit and the three-dimensional geohash logs of the decision traceability unit jointly build a transparent decision chain, ensuring farmers' trust in the AI ​​system while providing agronomists with a basis for rule optimization.

[0055] 4. This solution, through standardized interfaces and containerized deployment, enables seamless integration of the system with existing agricultural IoT devices. The digital twin preview and priority arbitration mechanism for the execution control module ensures the safety and efficiency of multi-machine collaborative operations without hardware modification. This solution is particularly suitable for unmanned operations in complex mountainous terrain, terraced fields, and other areas.

[0056] 5. This solution synthesizes regional extreme climate samples through adversarial training units and combines them with the variation response rules in the knowledge graph. This enables the system to generate alternative strategies that take into account both scientific models and local experience under abnormal conditions such as drought, floods, or outbreaks of pests and diseases, significantly reducing the risk of traditional unmanned systems losing control in sudden agricultural situations.

[0057] 6. This solution forms a spiraling intelligent evolutionary system through a closed technical loop of "experience structuring → knowledge mapping → decision fusion → execution verification." The semantic analysis of the acquisition module provides empirical elements for the construction module. The spatiotemporal constraints output by the construction module, combined with the real-time environmental data from the monitoring module, drive AI decisions. The data fusion module achieves a flexible balance between "artificial intelligence" and "human experience" through dynamic weight allocation. Ultimately, decisions are implemented through twin verification in the execution module. This multi-level collaborative mechanism fundamentally resolves the dual contradictions of agricultural AI systems: excessive data dependence and insufficient regional adaptability.

[0058] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of the system framework of an embodiment of the artificial intelligence decision-making system for unmanned agricultural operations of the present invention. DETAILED DESCRIPTION

[0060] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0061] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0062] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0063] The following is further described in detail through specific implementation methods:

[0064] Example:

[0065] As attached Figure 1 As shown: An artificial intelligence decision-making system for unmanned agricultural operations includes an acquisition module, a construction module, a monitoring module, an AI decision module, a data fusion decision module, and an execution control module that are interconnected. The monitoring module is signal-connected to an external sensor unit, which is used to collect environmental data from several different channels of farmland, including:

[0066] The acquisition module is used to collect text data from farmer interview records and agricultural logs, use natural language processing technology to perform semantic analysis on unstructured text, generate empirical feature vectors containing regional implicit knowledge, and transmit them to the construction module;

[0067] The acquisition module includes a natural language processing unit, a semantic analysis unit, and a rule extraction unit, among which:

[0068] The natural language processing unit is used to receive unstructured data such as farmers' voice interview recordings and handwritten farm log image data, and uses end-to-end speech recognition and optical character recognition technology to convert the unstructured data into standardized text data and transmit it to the semantic analysis unit;

[0069] A semantic analysis unit is used to perform dependency syntax analysis and named entity recognition on standardized text data, extract triple knowledge units containing geographic identifiers, crop varieties, and abnormal events, and transmit them to the rule extraction unit;

[0070] The rule extraction unit is used to transform the triple knowledge unit into a "condition-action" decision rule using a causal reasoning model and a keyword matching algorithm, and to add regional spatiotemporal labels to generate an empirical feature vector that is transmitted to the construction module. The rule extraction unit includes a temporal association subunit, a spatial association subunit, and a visualization verification subunit, where:

[0071] The time series association subunit uses the LSTM-Attention model to analyze the periodic patterns in agricultural logs and generate time-dependent decision rules;

[0072] The spatial association subunit extracts the association rules between micro-topographic features and agricultural operations through Voronoi diagram segmentation and generates geographic spatial constraints;

[0073] The visualization verification sub-unit generates an explanatory map of the extracted time-dependent decision rules and geographic spatial constraints for farmers to confirm. If the farmer's feedback conflicts with the rule logic, the rule weight downgrade mechanism is triggered.

[0074] The construction module is used to receive the empirical feature vector, construct a multi-dimensional associated knowledge graph using a graph neural network, generate a structured decision rule base containing spatiotemporal constraints, and transmit it to the data fusion decision module;

[0075] The building blocks include entity disambiguation unit, graph association unit, and rule verification unit, where:

[0076] The entity disambiguation unit is used to receive the empirical feature vector, eliminate the semantic conflicts of homonymous agricultural operations by matching geographic coordinates and aligning crop growth cycles, and generate a standardized empirical vector with geographic tags;

[0077] The graph association unit is used to receive standardized experience vectors and use the TransR algorithm to establish cross-modal association edges between pest and disease cycle laws, meteorological factors, and agricultural operations, generating intermediate knowledge graph data containing spatiotemporal constraints.

[0078] The rule verification unit is used to compare the intermediate data of the knowledge graph with the external historical agricultural operation database, correct the associated edge weights that are inconsistent with the measured results through the back-propagation algorithm, and output a structured decision rule library that matches the actual agricultural conditions in the region. The structured decision rule library contains spatiotemporal constraints including rule priority coefficients and effective geographical scope.

[0079] The monitoring module is used to receive multi-source heterogeneous environmental data transmitted by external sensing units, use wavelet transform to perform signal noise reduction, generate a standardized environmental feature matrix, and transmit it to the AI ​​decision module;

[0080] The monitoring module includes a data alignment unit, an anomaly detection unit, and a feature encoding unit, where:

[0081] The heterogeneous data alignment unit is used to receive heterogeneous data transmitted by external sensing units, including drone remote sensing images, soil sensor data, and weather station data, and generate environmental monitoring data in a unified coordinate system by performing spatiotemporal registration. The data is then transmitted to the anomaly detection unit and feature encoding unit respectively.

[0082] The anomaly detection unit uses the isolation forest algorithm to segment high-dimensional remote sensing images into 16×16 pixel blocks. The NDVI index (normalized difference vegetation index) of each block is calculated as an input feature. The anomaly threshold τ is dynamically adjusted according to the crop growth period (τ = 0.5 in the seedling stage and τ = 0.6 in the mature stage) to avoid misjudgment caused by changes in the growth period. This is used to identify sensor unit failures or anomalies in environmental monitoring data. If an anomaly in environmental monitoring data is detected, a redundant device switching mechanism is triggered and data is re-collected. For example, if an anomaly in soil moisture sensor data is detected, the surface temperature data of the infrared thermal imaging sensor in the same area is preferentially switched to for cross-validation.

[0083] The feature coding unit uses wavelet transform to decompose the noise-reduced environmental monitoring data, extracts frequency domain features, and combines them with the autoencoder for dimensionality compression to generate a standardized environmental feature matrix.

[0084] The AI ​​decision module receives the environmental feature matrix, uses a deep reinforcement learning model combined with a general agricultural dataset for training, generates a decision plan with a confidence score, and transmits it to the data fusion decision module;

[0085] The AI ​​decision-making module includes an environmental feature analysis unit, an incremental learning unit, a confidence assessment unit, and an adversarial training unit, among which:

[0086] The environmental feature parsing unit is used to receive the environmental feature matrix and use a recursive feature elimination algorithm to remove redundant dimensions. The implementation of the recursive feature algorithm includes: using the L1 regularized logistic regression model as the feature sorting base model, the regularization coefficient λ = 0.01, initializing the feature set, and the target feature number k = 32; looping the following operations until F current =k:

[0087] a. Train the base model and obtain the feature weight w i ;

[0088] b. Eliminate the 10% features with the smallest absolute weights and update F current ;

[0089] Evaluate the performance of feature subsets through cross-validation, select the subset that maximizes the reinforcement learning model reward function R, and generate an environment state vector that matches the reinforcement learning model input dimension. The dimension of the environment state vector is consistent with the number of neurons in the policy network input layer.

[0090] The incremental decision unit receives the environment state vector and the regional rules updated by the knowledge graph, encodes the regional rules as constraints of the policy network, and dynamically updates the policy network parameters through the online policy gradient algorithm, taking the environment state vector as input and the regional rules as constraints, and outputs a preliminary decision plan containing micro-geographic adaptation parameters;

[0091] The confidence evaluation unit is used to receive the environment state vector and the preliminary decision plan, perform 300 forward propagation sampling on the preliminary decision plan using the Monte Carlo Dropout method, calculate the probability distribution variance of each decision action, and generate a confidence score on a scale of 0-100%. When the score is lower than 60%, a rule base intervention flag is added to the preliminary decision plan; when the score is higher than 60%, a decision plan with a confidence score is generated and transmitted to the data fusion decision module;

[0092] The adversarial training unit is used to input the environmental state vector into the GAN generator, synthesize the adversarial sample vector with superimposed extreme climate characteristics, deduce the adversarial sample vector through the policy network, generate alternative decision plans and inject them into the confidence evaluation unit to train the dataset.

[0093] The data fusion decision module is used to receive the structured decision rule base and decision plan, and adopts a dynamic weighting algorithm to perform multimodal data fusion: when the confidence score is lower than the preset threshold, the rule matching engine is activated to prioritize the execution of the regional disposal strategy in the knowledge graph; when the confidence score is higher than the threshold, the LSTM time series prediction model is used to optimize the AI ​​decision output, generate the final control instruction, and transmit it to the execution control module;

[0094] The data fusion decision module includes a weight allocation unit, a conflict resolution unit, and a decision tracing unit, among which:

[0095] The weight allocation unit is used to receive a structured decision rule base and a decision plan, and analyze the geographical range constraints in the structured decision rule base and the spatial coordinates in the decision plan. When the confidence score is less than 60% and the geographical matching degree is greater than 85%, a rule-dominant flag is generated and the rule weight is triggered to be increased to 80%, and fusion decision intermediate data containing the weight allocation coefficient is output; the weight allocation unit outputs the weight allocation coefficient in the following manner: constructing a dynamic influencing factor set including soil type, crop growth period and weather anomaly index, wherein the dynamic influencing factor set is constructed by defining an influencing factor F = {}, where is soil type, the weight distribution is: sandy soil = 0.2, clay = 0.5, loam = 0.3; is crop growth period, the weight distribution is: seedling stage = 0.4, jointing stage = 0.6, maturity stage = 0.8; is weather anomaly index, the weight is expressed as the number of consecutive drought days / 30;

[0096] The fuzzy comprehensive evaluation method is used to calculate the rule applicability coefficient of each factor; the applicability coefficient is mapped to the fusion weight of the knowledge rule through the sigmoid function.

[0097] The conflict resolution unit is used to receive the intermediate data of the fusion decision and extract the rule instructions. When it detects that the rule instructions conflict with the AI ​​decision, it extracts the historical cases of farmers that match the current geographic tag in the structured rule base and calculates the similarity of environmental features:

[0098] Similarity = body 6 × cos (environmental feature matrix, historical cases) +

[0099] Body·4×(1-|current date-historical case date| / growth cycle length);

[0100] If the similarity is greater than 75% and there is a rule-dominant flag, a forced switch instruction is generated and the AI ​​decision flow is terminated;

[0101] The conflict resolution unit includes an experience priority subunit and a dynamic correction subunit, where:

[0102] The experience priority sub-unit is used to retrieve successful cases with the same geographic identifier in the farmer's historical operation records when the rule instruction conflicts with the AI ​​decision and the confidence score is less than 60%. It matches the current environmental feature matrix with the historical scenario data. If the matching degree is ≥ 70% and the rule-dominant flag is present, it generates an experience-forced instruction to override the AI ​​decision and sends an alarm log containing the conflict field to the decision tracing unit.

[0103] The dynamic correction subunit is used to receive field execution effect data in real time after executing the experience-based forced instructions, calculate the deviation between the actual effect and the expected value of the rule base, and send a rule weight degradation instruction to the construction module and trigger a graph reconstruction event when the deviation is greater than 15% for 3 hours.

[0104] The decision optimization unit is used to perform differentiated processing based on the conflict resolution results. When a forced switching instruction is received, the control parameters of the corresponding entry in the structured rule base are directly called. When the confidence score is ≥60%, the preliminary decision plan is input into the LSTM time series prediction model and the current crop growth period parameters are loaded. A temporal convolutional network is used to extract the periodic characteristics of the meteorological data of the past 30 days. The attention mechanism is used to weight recent environmental characteristics and output the final control instruction with a time series optimization label.

[0105] The decision tracing unit is used to write the geographic matching degree, weight distribution coefficient and forced switching instruction status code into the blockchain while generating the final control command, and generate a verifiable audit log containing a three-dimensional geographic hash value based on the timestamp and spatial coordinates contained in the timing optimization tag in the final control command.

[0106] The execution control module is used to convert the final control instructions into operating parameters of agricultural equipment and send them to the field operation terminal via the LoRaWAN protocol;

[0107] The execution control module includes an instruction compilation unit, a priority unit, and a safety rehearsal unit, among which:

[0108] The instruction compilation unit is used to convert the final control instructions into ISO 11783 standard agricultural machinery control messages, which are compatible with agricultural equipment of multiple brands;

[0109] The priority unit dynamically adjusts the collaborative operation sequence of several agricultural equipment according to the response delay of agricultural equipment and the urgency of the operation;

[0110] The safety rehearsal unit uses digital twin technology to simulate the field execution effect of control instructions. If a conflict in the motion trajectory of agricultural equipment is detected, the abnormal instruction will be intercepted and the path will be replanned.

[0111] The specific implementation process is as follows: This example uses a rice-growing region in Jiangsu Province as an example. Due to its overreliance on a general pest and disease database, the traditional AI plant protection system was unable to identify the periodic resistance variations of the local rice planthopper population, resulting in inaccurate pesticide spraying decisions (historical misjudgment rate >35%). Furthermore, the system failed to leverage farmers' accumulated experience, such as the need to dig ditches and drain water in advance on southeastern slopes during the rainy season to prevent insect egg breeding.

[0112] This solution deployed mobile terminals in the field through the collection module, recording a total of 120 hours of dialect interviews with 20 farmers. It also scanned approximately 3,000 pages of handwritten farm logs from the past five years. These logs were converted into structured text using the dialect ASR model and optical character recognition (OCR) technology in the natural language processing unit. The semantic analysis unit extracted key triples, such as "southeast slope - three consecutive days of rainfall - ditching for drainage." The rule extraction unit, using the temporal association subunit, discovered the periodic pattern of "rice planthoppers showing pesticide resistance mutations every two years in June." The visual verification subunit then generated a map for farmers to verify.

[0113] Because traditional systems rely solely on satellite meteorological data and are unable to correlate micro-topographic features, this system uses Voronoi diagram segmentation to identify the proportion of waterlogged areas on the southeast slope. The entity disambiguation unit in the construction module analyzes the differences in "drainage" operations between slopes and flat land. When the slope is greater than 15°, the drainage volume increases by 30%. The atlas association unit establishes a cross-modal correlation edge of "precipitation → soil moisture content → insect egg hatching rate", with a correlation strength of 0.87.

[0114] The monitoring module integrates drone multispectral imagery and soil sensors, detecting that the moisture content of the southeast slope has reached 50%, with a moisture content threshold of 35%. When the threshold is exceeded by 15%, the feature encoding unit extracts high-frequency water seepage signals through wavelet transform and generates an environmental feature matrix. The adversarial training unit of the AI ​​decision module synthesizes extreme rainfall samples, simulates five consecutive days of heavy rain, and outputs a "dual-mode strategy":

[0115] Mode A: conventional spray (confidence 58%);

[0116] Model B: drainage + biopesticide (confidence 72%);

[0117] The confidence assessment unit marks pattern A with a "low confidence" flag (score 56%).

[0118] The data fusion module detects that the confidence level of pattern A is less than 60% and the geographic matching degree is 91%, triggering the rule-dominant flag. The conflict resolution unit retrieves the farmer's historical case: "Waterlogging on the southeast slope → Drainage + Bacillus thuringiensis spraying", with a similarity of 83%. At this time, a forced switching instruction is generated.

[0119] The execution control module compiled the "drainage + biopesticide" instruction into an ISO 11783 message and dispatched three unmanned agricultural machines via LoRaWAN. The safety rehearsal unit detected a 23% probability of agricultural machinery path conflict and avoided it by replanning the operation mode to an echelon operation mode.

[0120] The dynamic correction subunit maintains the rule weight by monitoring that the rate of insect egg disappearance is 18% faster than expected.

[0121] Field experiments were conducted to verify the effects of the traditional AI system and this system, and the comparison results are as follows:

[0122]

[0123] As can be seen from the above table, this system, by building a farmer experience knowledge graph and a dynamic fusion mechanism, comprehensively surpasses traditional AI systems in key indicators. The accuracy of pest and disease identification has increased by 15% to 93%, the regional decision-making adaptability has increased by 26 percentage points, the extreme weather misjudgment rate has decreased by 22%, the pesticide use has been reduced by 22%, and the rule update response time has been shortened from 72 hours to 4 hours, providing a reliable AI solution for complex farmland scenarios.

[0124] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. An artificial intelligence decision-making system for unmanned agricultural operations, characterized in that: It includes a signal-connected acquisition module, a construction module, a monitoring module, an AI decision module, a data fusion decision module, and an execution control module. The monitoring module is signal-connected to an external sensor unit, which is used to collect environmental data from several different channels of farmland, including: The acquisition module is used to collect text data from farmer interview records and agricultural logs, use natural language processing technology to perform semantic analysis on unstructured text, generate empirical feature vectors containing regional implicit knowledge, and transmit them to the construction module; The construction module is used to receive the empirical feature vector, construct a multi-dimensional associated knowledge graph using a graph neural network, generate a structured decision rule base containing spatiotemporal constraints, and transmit it to the data fusion decision module; The monitoring module is used to receive multi-source heterogeneous environmental data transmitted by external sensing units, use wavelet transform to perform signal noise reduction, generate a standardized environmental feature matrix, and transmit it to the AI ​​decision module; The AI ​​decision module receives the environmental feature matrix, uses a deep reinforcement learning model combined with a general agricultural dataset for training, generates a decision plan with a confidence score, and transmits it to the data fusion decision module; The data fusion decision module is used to receive the structured decision rule base and decision plan, and adopts a dynamic weighting algorithm to perform multimodal data fusion: when the confidence score is lower than the preset threshold, the rule matching engine is activated to prioritize the execution of the regional disposal strategy in the knowledge graph; when the confidence score is higher than the threshold, the LSTM time series prediction model is used to optimize the AI ​​decision output, generate the final control instruction, and transmit it to the execution control module; The execution control module is used to convert the final control instructions into operating parameters of agricultural equipment and send them to the field operation terminal through the LoRaWAN protocol.

2. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 1, characterized in that: The acquisition module includes a natural language processing unit, a semantic analysis unit, and a rule extraction unit, among which: The natural language processing unit is used to receive unstructured data such as farmers' voice interview recordings and handwritten farm log image data, and uses end-to-end speech recognition and optical character recognition technology to convert the unstructured data into standardized text data and transmit it to the semantic analysis unit; A semantic analysis unit is used to perform dependency syntax analysis and named entity recognition on standardized text data, extract triple knowledge units containing geographic identifiers, crop varieties, and abnormal events, and transmit them to the rule extraction unit; The rule extraction unit is used to convert triple knowledge units into "condition-action" decision rules using a causal reasoning model and keyword matching algorithm, and to attach regional spatiotemporal labels to generate empirical feature vectors that are transmitted to the construction module.

3. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 2, characterized in that: The rule extraction unit includes a temporal association subunit, a spatial association subunit, and a visualization verification subunit, wherein: The time series association subunit uses the LSTM-Attention model to analyze the periodic patterns in agricultural logs and generate time-dependent decision rules; The spatial association subunit extracts the association rules between micro-topographic features and agricultural operations through Voronoi diagram segmentation and generates geographic spatial constraints; The visualization verification sub-unit generates an explanatory map of the extracted time-dependent decision rules and geographic spatial constraints for farmers to confirm. If the farmer's feedback conflicts with the rule logic, the rule weight downgrade mechanism is triggered.

4. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 3, characterized in that: The building blocks include entity disambiguation unit, graph association unit, and rule verification unit, where: The entity disambiguation unit is used to receive the empirical feature vector, eliminate the semantic conflicts of homonymous agricultural operations by matching geographic coordinates and aligning crop growth cycles, and generate a standardized empirical vector with geographic tags; The graph association unit is used to receive standardized experience vectors and use the TransR algorithm to establish cross-modal association edges between pest and disease cycle laws, meteorological factors, and agricultural operations, generating intermediate knowledge graph data containing spatiotemporal constraints. The rule verification unit is used to compare the intermediate data of the knowledge graph with the external historical agricultural operation database, correct the associated edge weights that are inconsistent with the measured results through the back-propagation algorithm, and output a structured decision rule library that matches the actual agricultural conditions in the region. The structured decision rule library contains spatiotemporal constraints including rule priority coefficients and effective geographical scope.

5. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 4, characterized in that: The monitoring module includes a data alignment unit, an anomaly detection unit, and a feature encoding unit, where: The heterogeneous data alignment unit is used to receive heterogeneous data transmitted by external sensing units, including drone remote sensing images, soil sensor data, and weather station data, and generate environmental monitoring data in a unified coordinate system by performing spatiotemporal registration. The data is then transmitted to the anomaly detection unit and feature encoding unit respectively. The anomaly detection unit uses the isolation forest algorithm to identify sensor unit failures or abnormal values ​​in environmental monitoring data. If an abnormal environmental monitoring data is detected, the redundant device switching mechanism is triggered and the data is re-collected; The feature coding unit uses wavelet transform to decompose the noise-reduced environmental monitoring data, extracts frequency domain features, and combines them with the autoencoder for dimensionality compression to generate a standardized environmental feature matrix.

6. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 5, characterized in that: The AI ​​decision-making module includes an environmental feature analysis unit, an incremental learning unit, a confidence assessment unit, and an adversarial training unit, among which: The environmental feature parsing unit is used to receive the environmental feature matrix, remove redundant dimensions using a recursive feature elimination algorithm, and generate an environmental state vector that matches the input dimension of the reinforcement learning model; The incremental decision unit receives the environment state vector and the regional rules updated by the knowledge graph, encodes the regional rules as constraints of the policy network, and dynamically updates the policy network parameters through the online policy gradient algorithm, taking the environment state vector as input and the regional rules as constraints, and outputs a preliminary decision plan containing micro-geographic adaptation parameters; The confidence evaluation unit is used to receive the environment state vector and the preliminary decision plan, perform 300 forward propagation sampling on the preliminary decision plan using the Monte Carlo Dropout method, calculate the probability distribution variance of each decision action, and generate a confidence score on a scale of 0-100%. When the score is lower than 60%, a rule base intervention flag is added to the preliminary decision plan; when the score is higher than 60%, a decision plan with a confidence score is generated and transmitted to the data fusion decision module; The adversarial training unit is used to input the environmental state vector into the GAN generator, synthesize the adversarial sample vector with superimposed extreme climate characteristics, deduce the adversarial sample vector through the policy network, generate alternative decision plans and inject them into the confidence evaluation unit to train the dataset.

7. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 6, characterized in that: The data fusion decision module includes a weight allocation unit, a conflict resolution unit, and a decision tracing unit, among which: The weight allocation unit is used to receive the structured decision rule base and the decision plan, and analyze the geographical scope constraints in the structured decision rule base and the spatial coordinates in the decision plan. When the confidence score is less than 60% and the geographical matching degree is greater than 85%, it generates a rule-dominant flag and triggers the rule weight to be increased to 80%, and outputs the fused decision intermediate data including the weight allocation coefficient; The conflict resolution unit is used to receive the intermediate data of the fusion decision and extract the rule instructions. When it detects that the rule instructions conflict with the AI ​​decision, it extracts the historical cases of farmers that match the current geographic tag in the structured rule base and calculates the similarity of environmental features: Similarity = uli × cos (environmental feature matrix, historical cases) + ul4×(1-|current date-historical case date| / growth cycle length); If the similarity is greater than 75% and there is a rule-dominant flag, a forced switch instruction is generated and the AI ​​decision flow is terminated; The decision optimization unit is used to perform differentiated processing based on the conflict resolution results. When a forced switching instruction is received, the control parameters of the corresponding entry in the structured rule base are directly called. When the confidence score is ≥60%, the preliminary decision plan is input into the LSTM time series prediction model and the current crop growth period parameters are loaded. A temporal convolutional network is used to extract the periodic characteristics of the meteorological data of the past 30 days. The attention mechanism is used to weight recent environmental characteristics and output the final control instruction with a time series optimization label. The decision tracing unit is used to write the geographic matching degree, weight distribution coefficient and forced switching instruction status code into the blockchain while generating the final control command, and generate a verifiable audit log containing a three-dimensional geographic hash value based on the timestamp and spatial coordinates contained in the timing optimization tag in the final control command.

8. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 7, characterized in that: The weight distribution unit output includes the weight distribution coefficient in the following manner: construct a dynamic influencing factor set including soil type, crop growth period and weather anomaly index, wherein the dynamic influencing factor set is constructed by defining the influencing factor F = {f1, f2, f3}, f1 is the soil type, and the weight distribution is: sandy soil = 0.2, clay = 0.5, loam = 0.3; f2 is the crop growth period, and the weight distribution is: seedling stage = 0.4, jointing stage = 0.6, maturity stage = 0.8; f3 is the weather anomaly index, and the weight is expressed as the number of consecutive drought days / 30; The fuzzy comprehensive evaluation method is used to calculate the rule applicability coefficient of each factor; the applicability coefficient is mapped to the fusion weight of the knowledge rule through the sigmoid function.

9. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 8, characterized in that: The conflict resolution unit includes an experience priority subunit and a dynamic correction subunit, where: The experience priority sub-unit is used to retrieve successful cases with the same geographic identifier in the farmer's historical operation records when the rule instruction conflicts with the AI ​​decision and the confidence score is less than 60%. It matches the current environmental feature matrix with the historical scenario data. If the matching degree is ≥ 70% and the rule-dominant flag is present, it generates an experience-forced instruction to override the AI ​​decision and sends an alarm log containing the conflict field to the decision tracing unit. The dynamic correction subunit is used to receive field execution effect data in real time after executing the experience-based forced instructions, calculate the deviation between the actual effect and the expected value of the rule base, and send a rule weight degradation instruction to the construction module and trigger a graph reconstruction event when the deviation is greater than 15% for 3 hours.

10. The artificial intelligence decision-making system for unmanned agricultural operations according to claim 9, characterized in that: The execution control module includes an instruction compilation unit, a priority unit, and a safety rehearsal unit, among which: The instruction compilation unit is used to convert the final control instructions into ISO 11783 standard agricultural machinery control messages, which are compatible with agricultural equipment of multiple brands; The priority unit dynamically adjusts the collaborative operation sequence of several agricultural equipment according to the response delay of agricultural equipment and the urgency of the operation; The safety rehearsal unit uses digital twin technology to simulate the field execution effect of control instructions. If a conflict in the motion trajectory of agricultural equipment is detected, the abnormal instruction will be intercepted and the path will be replanned.

Images