Greenhouse water, fertilizer and pesticide optimal control system and method based on environment-crop-input multi-element coupling

Abstract

This invention provides a greenhouse water, fertilizer, and pesticide optimization control system and method based on multi-factor coupling of environment, crop, and inputs. The system includes deploying a sensor network to collect greenhouse data, processing the data to generate a spatiotemporally continuous dataset, quantifying the dynamic causal chain of environment, crop, and inputs based on the generated dataset, generating a decision rule base, and fusing the decision rule base with the real-time simulation results of a root zone three-dimensional digital twin model to generate water, fertilizer, and pesticide control commands. This invention achieves continuous spatiotemporal acquisition and fusion processing of greenhouse environment, crop, and input data through a multimodal sensor network, constructs a decision rule base containing dynamic causal chains, and combines real-time simulation of the root zone three-dimensional digital twin model with Bayesian causal inference to generate precise water, fertilizer, and pesticide control commands. This effectively quantifies the nonlinear causal relationships between multiple factors and improves the scientific nature of decision-making.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for facility agriculture, specifically to a greenhouse water, fertilizer, and pesticide optimization control system and method based on the coupling of multiple factors including environment, crops, and inputs. Background Technology

[0002] High and stable yields of greenhouse crops depend on the precise and coordinated regulation of the environment, crop physiology, and inputs (water, fertilizer, and pesticides). Traditional greenhouse control systems often employ a "threshold-feedback" or "expert rule" model.

[0003] On the environmental side: data is collected through temperature, humidity, light, and CO2 sensors, and ventilation, shading, or supplemental lighting are triggered based on empirical thresholds;

[0004] Crop side: Leaf area, nitrogen content and other indicators are measured manually or offline as a rough basis for fertilization and pesticide application;

[0005] On the input side: irrigation, fertilization, and pesticide application are carried out according to fixed formulas or periodic plans, lacking dynamic coupling with real-time environment and crop status.

[0006] Traditional greenhouse control methods have the following drawbacks:

[0007] a) Data gaps: Heterogeneous sensors and different sampling frequencies make it difficult to align environmental, crop, and input data in the spatiotemporal dimensions, forming "information silos";

[0008] b) Causality missing: Threshold rules only reflect the linear relationship between a single environmental factor and crop response, and cannot quantify the nonlinear causal chain of multiple factors such as temperature and humidity, light, CO2, and soil water-fertilizer-pesticide coupling.

[0009] c) Root zone blind zone: The existing system lacks real-time simulation of the three-dimensional diffusion process of rhizosphere water-nutrient-pesticide, and the control commands are mismatched with the actual needs of the root zone, resulting in water, fertilizer and pesticide redundancy, root salt damage or excessive pesticide residues.

[0010] d) Rigid rules: The expert rule base is based on historical experience and is difficult to adapt to different growth stages of crops and sudden climate changes, resulting in lagging or excessive regulation.

[0011] Therefore, there is an urgent need for an integrated control method for greenhouse water, fertilizer and pesticides that can integrate multimodal sensing. Summary of the Invention

[0012] To address the shortcomings of existing technologies, this invention provides a greenhouse water, fertilizer, and pesticide optimization control system and method based on the coupling of multiple factors including environment, crops, and inputs, in order to solve the problems mentioned in the background.

[0013] To achieve the above objectives, the present invention provides the following technical solution: a greenhouse water, fertilizer, and pesticide optimization control method based on the coupling of multiple factors including environment, crop, and inputs, comprising the following steps:

[0014] S1: Deploy a sensor network to collect greenhouse data and process the data to generate a spatiotemporally continuous dataset;

[0015] Greenhouse data includes environmental data, crop status data, and input data;

[0016] S2: Based on the generated spatiotemporal continuous dataset, quantify the dynamic causal chain of environment, crops and inputs, and generate a decision rule base;

[0017] S3: Integrate the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide and fertilizer regulation instructions;

[0018] S4: The integrated water, fertilizer and pesticide control commands are sent to the execution device via the IoT controller.

[0019] As a preferred embodiment, the specific steps for processing greenhouse data to generate a spatiotemporally continuous dataset include:

[0020] The processing of greenhouse data specifically includes outlier removal, time synchronization, spatial mapping, data standardization, data cleaning and spatiotemporal alignment to obtain a multimodal data stream with timestamps;

[0021] A time-series database is constructed to store timestamped multimodal data streams. The timestamped multimodal data streams are used to generate spatiotemporally continuous datasets by fusing historical and real-time data.

[0022] As a preferred embodiment, the specific process of step S2 includes: based on generating a spatiotemporal continuous dataset, extracting time-series data of the environment, crops and inputs in real time, quantifying the dynamic causal relationship among the three, and obtaining a dynamic causal chain;

[0023] Based on dynamic causal chains, a causal graph is constructed to extract the environment, crops, and inputs.

[0024] By using a graph neural network and an attention mechanism to model causal graphs, dynamic causal chains and a readable decision rule base are output.

[0025] The decision rule base includes water, pesticide, and fertilizer regulation rules based on causal relationships.

[0026] As a preferred embodiment, the specific process of extracting time-series data of the environment, crops, and inputs in real time based on the generation of a spatiotemporally continuous dataset, and quantifying the dynamic causal relationship among the three includes:

[0027] Filter environmental factors, crop factors, input factors, and real-time data from the data pool;

[0028] A method combining time-delay Granger causality test and transit entropy is used to quantify the dynamic causal strength among the three factors.

[0029] As a preferred embodiment, the formula for calculating the Granger causality test for the time delay is:

[0030] ;

[0031] in, This represents the value of the dependent variable at time t, and p represents the lag order. The regression coefficient represents the lagged term of the dependent variable itself. This represents the i-th lag term of the dependent variable. The regression coefficient represents the lagged term of the independent variable. This represents the j-th lag term of the independent variable. Represents the residual term;

[0032] The formula for calculating the nonlinear contribution of causal relationships by combining transfer entropy with computation is as follows:

[0033] ;

[0034] Where p(·) represents the probability density function, This represents the value of variable Y at time t+1. This represents the value of variable Y at time t. This represents the value of variable X at time t. This represents the probability of Y at time t+1, given only the value of variable Y at time t. This represents the probability of Y at time t+1, given the values ​​of variables X and Y at time t.

[0035] As a preferred embodiment, the graph neural network uses an attention mechanism to model the causal graph, and the calculation formula is as follows:

[0036] ;

[0037] ;

[0038] in, This represents the attention weight of node i to its neighbor node j. This represents the transpose of the learnable attention parameter vector. This represents the activation function. Represents a linear transformation matrix. This represents the original feature vector of node i. This indicates a splicing operation. This represents the result of a linear transformation of the features of neighbor node j. Let i represent the set of neighbors of node i. This represents the updated feature of node i.

[0039] As a preferred embodiment, the rule base generation process includes:

[0040] Extracting key causal paths from dynamic causal chains;

[0041] "If-Then" rules are generated based on the aforementioned causal path.

[0042] As a preferred embodiment, step S3 specifically comprises:

[0043] It receives environmental data, crop growth data, and input data from a multimodal sensor network in real time, and completes the missing data through a spatiotemporal interpolation algorithm to build a dynamically updated time-series database.

[0044] The system calls a pre-generated decision rule library, calculates the causal weights between nodes based on the current time series database state using a graph neural network, matches historical decision rules with a similarity greater than a preset threshold to the current greenhouse state, and outputs a set of candidate control strategies.

[0045] The constructed three-dimensional digital twin model of the root zone is dynamically simulated by inputting real-time data on current soil moisture, nutrient concentration, and drip irrigation tape zoning status, generating a three-dimensional infiltration rate field and nutrient distribution heat map of the root zone.

[0046] The candidate control strategy set and the results of the digital twin simulation are input into the Bayesian causal inference module to calculate the expected effect probability of each strategy in the current root region environment, and the strategies with probability values ​​greater than the preset threshold are selected as the initial selection instructions.

[0047] The initial selection command is applied to the greenhouse execution equipment, and the synchronously deployed soil sensors monitor changes in soil data in real time. The data is fed back to the digital twin model for parameter calibration, and the control command is iteratively optimized until the error between the measured value and the simulated value is less than the preset threshold, thus generating a control command containing specific parameters.

[0048] As a preferred embodiment, a greenhouse water, fertilizer, and pesticide optimization control system based on the coupling of multiple factors of environment, crop, and inputs is provided to implement the above-mentioned greenhouse water, fertilizer, and pesticide optimization control method based on the coupling of multiple factors of environment, crop, and inputs, comprising:

[0049] Data acquisition module: Collects greenhouse data through a multimodal sensor network, processes the greenhouse data, and generates a spatiotemporal continuous dataset;

[0050] Decision rule generation module: used to quantify the dynamic causal chain of environment, crops and inputs based on the generated spatiotemporal continuous dataset, and generate a decision rule base;

[0051] Regulation command generation module: used to integrate the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide and fertilizer regulation commands;

[0052] Command execution module: Sends the integrated water, fertilizer and pesticide regulation commands to the execution device through the IoT controller.

[0053] This invention provides a greenhouse water, fertilizer, and pesticide optimization control system and method based on the coupling of multiple factors including environment, crop, and inputs, which has the following beneficial effects: Greenhouse data is collected by deploying a sensor network and processed to generate a spatiotemporally continuous dataset. Based on this dataset, the dynamic causal chain of environment, crop, and inputs is quantified to generate a decision rule base. This decision rule base is then fused with the real-time simulation results of a three-dimensional digital twin model of the root zone to generate water, fertilizer, and pesticide control commands. These commands are then sent to the execution device via an IoT controller. The spatiotemporal connection of greenhouse environment, crop, and input data is achieved through a multimodal sensor network. Continuing to collect and fuse data, a decision rule base containing dynamic causal chains is constructed. Combining real-time simulation of the root zone's three-dimensional digital twin model with Bayesian causal inference, precise water, fertilizer, and pesticide regulation commands are generated. Execution and feedback optimization are utilized via the Internet of Things to form a closed-loop control mechanism, effectively quantifying the nonlinear causal relationships among multiple factors and improving the scientific nature of decision-making. The digital twin model dynamically simulates the water-nutrient diffusion process in the root zone, achieving precise matching and real-time calibration of regulation commands. Based on reinforcement learning, the rule base is dynamically optimized to adapt to environmental changes and crop growth needs, reducing redundant input of water, fertilizer, and pesticides, improving resource utilization efficiency, and simultaneously reducing pesticide residue risks, thus promoting intelligent and sustainable greenhouse production. Attached Figure Description

[0054] Figure 1 This is a flowchart of the greenhouse water, fertilizer and pesticide optimization control method based on the coupling of multiple factors of environment, crop and inputs according to the present invention;

[0055] Figure 2 This is a block diagram of the greenhouse water, fertilizer and pesticide optimization control system based on the coupling of multiple factors such as environment, crop and inputs according to the present invention. Detailed Implementation

[0056] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0057] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0058] like Figure 1 As shown, this embodiment of the invention provides a greenhouse water, fertilizer, and pesticide optimization control method based on the coupling of multiple factors including environment, crop, and inputs, comprising the following steps:

[0059] S1: Deploy a sensor network to collect greenhouse data and process the data to generate a spatiotemporally continuous dataset;

[0060] Greenhouse data includes environmental data, crop status data, and input data;

[0061] Specifically, by deploying environmental sensors (temperature, humidity, light, CO2), soil sensors (moisture, EC value, pH), crop growth sensors (leaf area index, stem flow rate), and pesticide and fertilizer residue detection devices in the greenhouse, a spatially distributed multimodal sensing network is formed, and greenhouse data is collected through the multimodal sensing network.

[0062] In this embodiment, the processing of greenhouse data specifically includes outlier removal, time synchronization, spatial mapping, data standardization, data cleaning, and spatiotemporal alignment to obtain a multimodal data stream with timestamps.

[0063] A time-series database is constructed to store timestamped multimodal data streams. The timestamped multimodal data streams are used to generate spatiotemporally continuous datasets by fusing historical and real-time data.

[0064] Specifically, outlier removal employs the sliding Forest algorithm to filter sensor noise; time synchronization uses the DTW algorithm to align environmental data with crop image data; spatial mapping unifies root fiber optic data and soil moisture sensor coordinates into a three-dimensional root zone model through ICP point cloud registration; and data standardization uses Z-Score normalization to process heterogeneous indicators such as soil EC value (μS / cm) and plant nitrogen content (g / kg). A spatiotemporally continuous dataset is generated by stitching together historical data from the past 7 days with current data through a sliding window.

[0065] It should be noted that time-series databases store spatiotemporally continuous datasets.

[0066] Specifically, by using the sliding Forest algorithm to remove sensor noise, the DTW algorithm to align heterogeneous time-series data, and the ICP point cloud registration to a unified spatial coordinate system, combined with Z-Score standardization to process heterogeneous indicators, a spatiotemporally continuous dataset with timestamps was constructed. This solved the data discontinuity problem caused by the heterogeneity of the greenhouse environment and provided a high-quality data foundation for accurate decision-making.

[0067] S2: Based on the generated spatiotemporal continuous dataset, quantify the dynamic causal chain of environment, crops and inputs, and generate a decision rule base;

[0068] In this embodiment, the specific process of step S2 includes: based on generating a spatiotemporal continuous dataset, extracting time-series data of the environment, crops and inputs in real time, quantifying the dynamic causal relationship among the three, and obtaining a dynamic causal chain;

[0069] Based on dynamic causal chains, a causal graph is constructed to extract the environment, crops, and inputs.

[0070] By using a graph neural network and an attention mechanism to model causal graphs, dynamic causal chains and a readable decision rule base are output.

[0071] The decision rule base includes water, pesticide, and fertilizer regulation rules based on causal relationships.

[0072] Specifically, the process of generating a spatiotemporally continuous dataset, extracting time-series data of the environment, crops, and inputs in real time, and quantifying the dynamic causal relationships among the three includes:

[0073] Filter environmental factors from the data pool, such as temperature, humidity, light, and CO2 concentration; crop factors, such as leaf area index, stem flow rate, root activity, and nutrient content; and input factors, such as irrigation amount, fertilizer concentration, and pesticide dosage, both historical and real-time data.

[0074] A method combining time-delay Granger causality test and transit entropy is used to quantify the dynamic causal strength among the three factors.

[0075] The formula for calculating the Granger causality test with time delay is as follows:

[0076] ;

[0077] in, This represents the value of the dependent variable at time t, and p represents the lag order. The regression coefficient represents the lagged term of the dependent variable itself. This represents the i-th lag term of the dependent variable. The regression coefficient represents the lagged term of the independent variable. This represents the j-th lag term of the independent variable. This represents the residual term.

[0078] The nonlinear contribution of causal relationships is calculated by combining transfer entropy with computation, and the formula is as follows:

[0079] ;

[0080] Where p(·) represents the probability density function, This represents the value of variable Y at time t+1. This represents the value of variable Y at time t. This represents the value of variable X at time t. This represents the probability of Y at time t+1, given only the value of variable Y at time t. This represents the probability of Y at time t+1, given the values ​​of variables X and Y at time t.

[0081] It should be noted that the cause-effect graph G=(V,E) is a graph where the set of nodes V represents the state variables of the environment, crops, and inputs, and the set of edges E represents the causal relationships between the variables.

[0082] In this embodiment, the dynamic causal chain representation formula is:

[0083] ;

[0084] in, This represents the environment state vector at time t. Let represent the crop state vector at time t. Denotes the input vector at time t. The nonlinear function representing the parameterization of GNN, Indicates the causal delay time. This indicates the noise term.

[0085] Specifically, the nonlinear causal quantification method based on time-delay Granger causality test and transfer entropy breaks through the limitations of traditional single environmental factor regulation. It realizes multi-dimensional dynamic correlation analysis between environmental factors such as temperature, humidity, light, and CO2 concentration and crop leaf area index, stem flow rate, nutrient content, and water, fertilizer and pesticide input, significantly improving the scientific nature of regulation strategies.

[0086] Specifically, graph neural networks use an attention mechanism to model causal graphs, and their calculation formula is as follows:

[0087] ;

[0088] ;

[0089] in, This represents the attention weight of node i to its neighbor node j. This represents the transpose of the learnable attention parameter vector. This represents the activation function. Represents a linear transformation matrix. This represents the original feature vector of node i. This indicates a splicing operation. This represents the result of a linear transformation of the features of neighbor node j. Let i represent the set of neighbors of node i. This represents the updated feature of node i.

[0090] Specifically, the process of generating the rule base includes:

[0091] Extracting key causal paths from dynamic causal chains;

[0092] "If-Then" rules are generated based on the aforementioned causal path.

[0093] Specifically, by using graph neural networks to model causal graphs, key causal paths are extracted in real time to generate an "If-Then" rule base. Compared with static expert rule bases, this can adapt to crop growth stages and environmental changes, and dynamically adjust water, fertilizer and pesticide regulation parameters.

[0094] S3: Integrate the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide and fertilizer regulation instructions;

[0095] Specifically, it receives environmental data, crop growth data, and input data from a multimodal sensor network in real time, and completes the missing data through a spatiotemporal interpolation algorithm to build a dynamically updated time-series database.

[0096] The system calls a pre-generated decision rule library, calculates the causal weights between nodes based on the current time series database state using a graph neural network, matches historical decision rules with a similarity greater than a preset threshold to the current greenhouse state, and outputs a set of candidate control strategies.

[0097] The constructed three-dimensional digital twin model of the root zone is dynamically simulated by inputting real-time data on current soil moisture, nutrient concentration, and drip irrigation tape zoning status, generating a three-dimensional infiltration rate field and nutrient distribution heat map of the root zone.

[0098] The candidate control strategy set and the results of the digital twin simulation are input into the Bayesian causal inference module to calculate the expected effect probability of each strategy in the current root region environment, and the strategies with probability values ​​greater than the preset threshold are selected as the initial selection instructions.

[0099] The initial selection command is applied to the greenhouse execution equipment, and the soil sensors deployed simultaneously monitor changes in soil data in real time. The data is fed back to the digital twin model for parameter calibration, and the control command is iteratively optimized until the error between the measured value and the simulated value is less than the preset threshold.

[0100] Generate control instructions containing specific parameters.

[0101] It should be noted that the control instructions include: drip irrigation zone switch status, fertilizer injection rate and pesticide spraying dosage, and expected root zone humidity distribution and nutrient concentration curve.

[0102] S4: The integrated water, fertilizer and pesticide control commands are sent to the execution device via the IoT controller.

[0103] In this embodiment, after S4, the method further includes: continuously collecting environmental parameters, crop physiological indicators, root zone status, and input residue data after execution through a deployed sensor network to form a dynamic feedback dataset;

[0104] Real-time feedback data is input into the root zone 3D digital twin model, the model parameters are dynamically corrected, the virtual space and physical space are synchronously mapped, and the corrected environment-crop response prediction is output.

[0105] Based on feedback data and model correction results, a reinforcement learning algorithm is used to dynamically optimize the decision rule base generated by S2.

[0106] Optimizing the decision rule base generated by S2 includes: adjusting the weights of the environment-crop causal chain, updating the input effect function, and introducing constraints.

[0107] This embodiment provides a greenhouse water, fertilizer, and pesticide optimization control method based on the coupling of multiple factors including environment, crop, and inputs. It utilizes a multimodal sensor network to achieve continuous spatiotemporal acquisition and fusion processing of greenhouse environment, crop, and input data. This constructs a decision rule base containing dynamic causal chains. Combined with real-time simulation of a three-dimensional digital twin model of the root zone and Bayesian causal inference, precise water, fertilizer, and pesticide control commands are generated. The method leverages the Internet of Things (IoT) for execution and feedback optimization, forming a closed-loop control mechanism. This effectively quantifies the nonlinear causal relationships among multiple factors, enhancing the scientific nature of decision-making. The method dynamically simulates the water-nutrient diffusion process in the root zone using a digital twin model, achieving precise matching and real-time calibration of control commands. Based on reinforcement learning, the method dynamically optimizes the rule base, adapting to environmental changes and crop growth needs, reducing redundant water, fertilizer, and pesticide inputs, improving resource utilization efficiency, and simultaneously reducing pesticide residue risks, thus promoting intelligent and sustainable greenhouse production.

[0108] like Figure 2As shown, this embodiment also provides a greenhouse water, fertilizer, and pesticide optimization control system based on the coupling of multiple factors of environment, crop, and inputs, used to implement the above-mentioned greenhouse water, fertilizer, and pesticide optimization control method based on the coupling of multiple factors of environment, crop, and inputs, including:

[0109] Data acquisition module: Collects greenhouse data through a multimodal sensor network, processes the greenhouse data, and generates a spatiotemporal continuous dataset;

[0110] Decision rule generation module: used to quantify the dynamic causal chain of environment, crops and inputs based on the generated spatiotemporal continuous dataset, and generate a decision rule base;

[0111] Regulation command generation module: used to integrate the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide and fertilizer regulation commands;

[0112] Command execution module: Sends the integrated water, fertilizer and pesticide regulation commands to the execution device through the IoT controller.

[0113] In this embodiment, the processing of greenhouse data specifically includes outlier removal, time synchronization, spatial mapping, data standardization, data cleaning, and spatiotemporal alignment to obtain a multimodal data stream with timestamps.

[0114] A time-series database is constructed to store timestamped multimodal data streams. The timestamped multimodal data streams are used to generate spatiotemporally continuous datasets by fusing historical and real-time data.

[0115] Specifically, outlier removal employs the sliding Forest algorithm to filter sensor noise; time synchronization uses the DTW algorithm to align environmental data with crop image data; spatial mapping unifies root fiber optic data and soil moisture sensor coordinates into a three-dimensional root zone model through ICP point cloud registration; and data standardization uses Z-Score normalization to process heterogeneous indicators such as soil EC value (μS / cm) and plant nitrogen content (g / kg). A spatiotemporally continuous dataset is generated by stitching together historical data from the past 7 days with current data through a sliding window.

[0116] It should be noted that time-series databases store spatiotemporally continuous datasets.

[0117] In this embodiment, the specific process of generating a decision rule base by quantifying the dynamic causal chain of environment, crop and input based on the generated spatiotemporal continuous dataset includes: extracting time series data of environment, crop and input in real time based on the generated spatiotemporal continuous dataset, quantifying the dynamic causal relationship among the three, and obtaining the dynamic causal chain;

[0118] Based on dynamic causal chains, a causal graph is constructed to extract the environment, crops, and inputs.

[0119] By using a graph neural network and an attention mechanism to model causal graphs, dynamic causal chains and a readable decision rule base are output.

[0120] The decision rule base includes water, pesticide, and fertilizer regulation rules based on causal relationships.

[0121] Specifically, the process of generating a spatiotemporally continuous dataset, extracting time-series data of the environment, crops, and inputs in real time, and quantifying the dynamic causal relationships among the three includes:

[0122] Filter environmental factors from the data pool, such as temperature, humidity, light, and CO2 concentration; crop factors, such as leaf area index, stem flow rate, root activity, and nutrient content; and input factors, such as irrigation amount, fertilizer concentration, and pesticide dosage, both historical and real-time data.

[0123] A method combining time-delay Granger causality test and transit entropy is used to quantify the dynamic causal strength among the three factors.

[0124] The formula for calculating the Granger causality test with time delay is as follows:

[0125] ;

[0126] in, This represents the value of the dependent variable at time t, and p represents the lag order. The regression coefficient represents the lagged term of the dependent variable itself. This represents the i-th lag term of the dependent variable. The regression coefficient represents the lagged term of the independent variable. This represents the j-th lag term of the independent variable. This represents the residual term.

[0127] The nonlinear contribution of causal relationships is calculated by combining transfer entropy with computation, and the formula is as follows:

[0128] ;

[0129] Where p(·) represents the probability density function, This represents the value of variable Y at time t+1. This represents the value of variable Y at time t. This represents the value of variable X at time t. This represents the probability of Y at time t+1, given only the value of variable Y at time t. This represents the probability of Y at time t+1, given the values ​​of variables X and Y at time t.

[0130] It should be noted that the cause-effect graph G=(V,E) is a graph where the set of nodes V represents the state variables of the environment, crops, and inputs, and the set of edges E represents the causal relationships between the variables.

[0131] In this embodiment, the dynamic causal chain representation formula is:

[0132] ;

[0133] in, This represents the environment state vector at time t. Let represent the crop state vector at time t. Denotes the input vector at time t. The nonlinear function representing the parameterization of GNN, Indicates the causal delay time. This indicates the noise term.

[0134] Specifically, graph neural networks use an attention mechanism to model causal graphs, and their calculation formula is as follows:

[0135] ;

[0136] ;

[0137] in, This represents the attention weight of node i to its neighbor node j. This represents the transpose of the learnable attention parameter vector. This represents the activation function. Represents a linear transformation matrix. This represents the original feature vector of node i. This indicates a splicing operation. This represents the result of a linear transformation of the features of neighbor node j. Let i represent the set of neighbors of node i. This represents the updated feature of node i.

[0138] Specifically, the process of generating the rule base includes:

[0139] Extracting key causal paths from dynamic causal chains;

[0140] "If-Then" rules are generated based on the aforementioned causal path.

[0141] Specifically, the detailed process of integrating the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide, and fertilizer regulation instructions includes:

[0142] It receives environmental data, crop growth data, and input data from a multimodal sensor network in real time, and completes the missing data through a spatiotemporal interpolation algorithm to build a dynamically updated time-series database.

[0143] The system calls a pre-generated decision rule library, calculates the causal weights between nodes based on the current time series database state using a graph neural network, matches historical decision rules with a similarity greater than a preset threshold to the current greenhouse state, and outputs a set of candidate control strategies.

[0144] The constructed three-dimensional digital twin model of the root zone is dynamically simulated by inputting real-time data on current soil moisture, nutrient concentration, and drip irrigation tape zoning status, generating a three-dimensional infiltration rate field and nutrient distribution heat map of the root zone.

[0145] The candidate control strategy set and the results of the digital twin simulation are input into the Bayesian causal inference module to calculate the expected effect probability of each strategy in the current root region environment, and the strategies with probability values ​​greater than the preset threshold are selected as the initial selection instructions.

[0146] The initial selection command is applied to the greenhouse execution equipment, and the soil sensors deployed simultaneously monitor changes in soil data in real time. The data is fed back to the digital twin model for parameter calibration, and the control command is iteratively optimized until the error between the measured value and the simulated value is less than the preset threshold.

[0147] Generate control instructions containing specific parameters.

[0148] It should be noted that the control instructions include: drip irrigation zone switch status, fertilizer injection rate and pesticide spraying dosage, and expected root zone humidity distribution and nutrient concentration curve.

[0149] Furthermore, the greenhouse water, fertilizer and pesticide optimization control system based on the coupling of multiple factors of environment, crop and input in this application also includes a closed-loop optimization module, which is used to optimize the decision rule base.

[0150] Specifically, the optimization process includes: continuously collecting environmental parameters, crop physiological indicators, root zone status, and input residue data after execution through a deployed sensor network to form a dynamic feedback dataset;

[0151] Real-time feedback data is input into the root zone 3D digital twin model, the model parameters are dynamically corrected, the virtual space and physical space are synchronously mapped, and the corrected environment-crop response prediction is output.

[0152] Based on feedback data and model correction results, a reinforcement learning algorithm is used to dynamically optimize the generated decision rule base.

[0153] The optimized decision rule base includes: adjusting the weights of the environment-crop causal chain, updating the input effect function, and introducing constraints.

[0154] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A greenhouse water, fertilizer, and pesticide optimization control method based on multi-factor coupling of environment, crop, and inputs, characterized in that, Includes the following steps: S1: Deploy a sensor network to collect greenhouse data and process the data to generate a spatiotemporally continuous dataset; Greenhouse data includes environmental data, crop status data, and input data; S2: Based on the generated spatiotemporal continuous dataset, quantify the dynamic causal chain of environment, crops and inputs, and generate a decision rule base; S3: Integrate the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide and fertilizer regulation instructions; S4: The integrated water, fertilizer, and pesticide control commands are sent to the execution device via the IoT controller; The specific process for processing greenhouse data to generate a spatiotemporally continuous dataset includes: The processing of greenhouse data specifically includes outlier removal, time synchronization, spatial mapping, data standardization, data cleaning and spatiotemporal alignment to obtain a multimodal data stream with timestamps; Among them, the sliding Forest algorithm is used to filter sensor noise; the DTW algorithm is used to align environmental data and crop image data; and ICP point cloud registration is used to unify root fiber data and soil sensor coordinates into a three-dimensional root zone model. A time-series database is constructed to store timestamped multimodal data streams. The timestamped multimodal data streams are used to generate a spatiotemporally continuous dataset by fusing historical and real-time data. The specific process of step S2 includes: based on the generation of a spatiotemporal continuous dataset, extracting time-series data of the environment, crops and inputs in real time, quantifying the dynamic causal relationship among the three, and obtaining a dynamic causal chain; Based on dynamic causal chains, a causal graph is constructed to extract the environment, crops, and inputs. By using a graph neural network and an attention mechanism to model causal graphs, dynamic causal chains and a readable decision rule base are output. Among them, the decision rule base includes water, pesticide, and fertilizer regulation rules based on causal relationships; The specific steps of S3 are as follows: It receives environmental data, crop growth data, and input data from a multimodal sensor network in real time, and completes the missing data through a spatiotemporal interpolation algorithm to build a dynamically updated time-series database. The pre-generated decision rule library is invoked. Based on the current time series database state, the causal weights between nodes are calculated through a graph neural network. Historical decision rules with a similarity greater than a preset threshold to the current greenhouse state are matched, and a set of candidate control strategies is output. The constructed three-dimensional digital twin model of the root zone is dynamically simulated by inputting real-time data on current soil moisture, nutrient concentration, and drip irrigation tape zoning status, generating a three-dimensional infiltration rate field and nutrient distribution heat map of the root zone. The candidate control strategy set and the results of the digital twin simulation are input into the Bayesian causal inference module to calculate the expected effect probability of each strategy in the current root region environment, and the strategies with probability values ​​greater than the preset threshold are selected as the initial selection instructions. The initial selection command is applied to the greenhouse execution equipment, and the synchronously deployed soil sensors monitor changes in soil data in real time. The data is fed back to the digital twin model for parameter calibration, and the control command is iteratively optimized until the error between the measured value and the simulated value is less than the preset threshold, thus generating a control command containing specific parameters.

2. The greenhouse water, fertilizer, and pesticide optimization control method based on multi-factor coupling of environment, crop, and inputs as described in claim 1, is characterized in that... The specific process for extracting time-series data of the environment, crops, and inputs in real time based on the generated spatiotemporal continuous dataset, and quantifying the dynamic causal relationship among the three, includes: Filter environmental factors, crop factors, input factors, and real-time data from the data pool; A method combining time-delay Granger causality test and transit entropy is used to quantify the dynamic causal strength among the three factors.

3. The greenhouse water, fertilizer, and pesticide optimization control method based on multi-factor coupling of environment, crop, and inputs as described in claim 2, is characterized in that... The formula for calculating the Granger causality test with time delay is as follows: ； in, This represents the value of the dependent variable at time t, and p represents the lag order. The regression coefficient represents the lagged term of the dependent variable itself. This represents the i-th lag term of the dependent variable. The regression coefficient represents the lagged term of the independent variable. This represents the j-th lag term of the independent variable. Represents the residual term; The formula for calculating the nonlinear contribution of causal relationships by combining transfer entropy with computation is as follows: ； Where p(·) represents the probability density function, This represents the value of variable Y at time t+1. This represents the value of variable Y at time t. This represents the value of variable X at time t. This represents the probability of Y at time t+1, given only the value of variable Y at time t. This represents the probability of Y at time t+1, given the values ​​of variables X and Y at time t.

4. The greenhouse water, fertilizer, and pesticide optimization control method based on multi-factor coupling of environment, crop, and inputs as described in claim 1, characterized in that, The graph neural network uses an attention mechanism to model the causal graph. The calculation formula is as follows: ； ； in, This represents the attention weight of node i to its neighbor node j. This represents the transpose of the learnable attention parameter vector. This represents the activation function. Represents a linear transformation matrix. This represents the original feature vector of node i. This indicates a splicing operation. This represents the result of a linear transformation of the features of neighbor node j. Let i represent the set of neighbors of node i. This represents the updated feature of node i.

5. The greenhouse water, fertilizer, and pesticide optimization control method based on multi-factor coupling of environment, crop, and inputs as described in claim 4, is characterized in that... The process of generating the rule base includes: Extracting key causal paths from dynamic causal chains; "If-Then" rules are generated based on the key causal path.

6. A greenhouse water, fertilizer, and pesticide optimization control system based on multi-factor coupling of environment, crop, and inputs, used to implement the greenhouse water, fertilizer, and pesticide optimization control method based on multi-factor coupling of environment, crop, and inputs as described in any one of claims 1-5, characterized in that, include: Data acquisition module: Collects greenhouse data through a multimodal sensor network, processes the greenhouse data, and generates a spatiotemporal continuous dataset; Decision rule generation module: used to quantify the dynamic causal chain of environment, crops and inputs based on the generated spatiotemporal continuous dataset, and generate a decision rule base; Regulation command generation module: used to integrate the decision rule base with the real-time simulation results of the root zone three-dimensional digital twin model to generate water, pesticide and fertilizer regulation commands; Command execution module: Sends the integrated water, fertilizer and pesticide regulation commands to the execution device through the IoT controller.

Images