A model-data-driven greenhouse intelligent fertilizer distribution method and system

Abstract

The present invention relates to the agricultural field and discloses a model-data-driven intelligent fertilizer distribution method and system for greenhouses. The system comprises a control and sensing module, an image analysis module, and a task decision module. The control and sensing module includes a soil sensor network and an irrigation vehicle controlled by an electronically controlled processor and a communication program. The image analysis module hardware comprises a computing device and a cloud server, while the image analysis module software comprises a growth stage detection network and a nutrient deficiency detection network. The task decision module hardware comprises an edge computing device and a cloud server, while the task decision module software comprises a fertilizer distribution expert system and a deep decision neural network based on reinforcement learning. This solution achieves high fertilizer distribution decision-making accuracy, good security and stability, and is self-upgradable. This is because the proposed decision-making method combines a machine vision crop fertilizer deficiency detection method based on a convolutional neural network with a model-data fusion-driven decision-making solution.

Description

Technical Field

[0001] The present invention relates to the field of agriculture, and in particular to a model-data jointly driven intelligent fertilizer distribution method and system for greenhouses. Background Art

[0002] With the rapid development of artificial intelligence and the Internet of Things (IoT), agriculture is becoming increasingly information-based and intelligent. Among them, greenhouse smart agriculture has seen the most rapid development, sparking extensive research and a large amount of technological output.

[0003] The foundation for smart greenhouse agriculture is the establishment of an agricultural Internet of Things (IoT). CN105137940A discloses an IoT greenhouse agricultural growth monitoring system, exemplifying a classic IoT architecture. This architecture comprises a perception system, a control system, and a communication system, responsible for information sensing, decision-making control, and wireless communication, respectively. Equipped with a large number of sensors and processors, the system enables real-time monitoring of crop growth. Once this classic network architecture was proposed, numerous subdivided functional IoTs emerged based on it, with operational decision-making IoT achieving the most success. The technology underlying this proposal is real-time fertilizer allocation decision-making.

[0004] There are two technical difficulties in fertilizer allocation decisions: one is how to assess and quantify crop fertilizer deficiencies, and the other is how to dynamically allocate fertilizer based on IoT sensor data.

[0005] For the former, i.e., quantitative determination of crop nutritional deficiencies (or nutritional abundance or deficiency), machine vision technology is usually required. The method proposed in CN114946447A is to perform regression analysis on images corresponding to nutrient abundance or deficiency in seedling cultivation. For selected seedlings, this method designs different nutrient solution component formulas, cultivates and collects images of seedling leaves under different nutrient components, establishes a database of leaf feature images of seedlings with normal nutrient deficiencies of nitrogen, phosphorus, and potassium, extracts digital eigenvalues ​​of the images, performs regression analysis, and determines the correlation between different nutrient deficiencies and the digital eigenvalues ​​of the leaf images, thereby guiding water and fertilizer management.

[0006] For the latter, that is, how to allocate fertilizers in real time, it is usually necessary to comprehensively consider the growth environment of crops and establish a mathematical model of crop fertilizer requirements. Taking into account the complexity of crop types and different greenhouse environments, it is usually necessary to introduce an expert system to integrate information and make decisions. Regarding how the expert system is deployed and decided, the journal article "Litchi Garden Information Acquisition and Intelligent Irrigation Expert Decision System Based on the Internet of Things" (doi: 10.11975 / j.issn.1002-6819.2016.20.019) has been described in detail. Among the related technical solutions, CN111108879A discloses a method for water, fertilizer and environmental management of eggplants based on an expert decision-making system. This method uses deep learning technology and growth change data of eggplants and melons to fit the growth model of crops, and matches water and fertilizer strategies for different stages according to the knowledge base of the expert system, thereby completing the fertilizer allocation decision.

[0007] Based on the above existing technology:

[0008] The crop fertilizer deficiency detection method proposed in CN114946447A suffers from limited robustness and generalization capabilities. This is because it uses regression analysis. Regression analysis cannot fully capture influencing factors and cannot fully tap into certain unpredictable latent factors, resulting in low prediction accuracy and large residual errors.

[0009] The fertilizer allocation decision-making technology proposed in CN111108879A does not explicitly incorporate machine vision to assist in fertilizer allocation, and the decision-making system lacks an upgrade mechanism. Crops often exhibit visual abnormalities when they are deficient in certain nutrients. Ignoring the crucial method of "identifying disease based on images" and relying solely on other environmental sensor data for decision-making is inherently ineffective.

[0010] In summary, we proposed a model-data jointly driven intelligent fertilizer distribution method and system for greenhouses. Summary of the Invention

[0011] (1) Technical problems solved

[0012] In response to the shortcomings of the existing technology, the present invention provides a model-data jointly driven greenhouse intelligent fertilizer distribution method and system, proposes a growth stage feature detection and nutrient deficiency detection method based on a dual convolutional neural network, and performs fertilizer deficiency detection from two dimensions: the overall growth characteristics and local growth characteristics of the crop, further improving the accuracy and generalization ability of the fertilizer deficiency detection model. On this basis, a fusion fertilizer distribution decision system consisting of an expert system and a decision network based on deep reinforcement learning is proposed, which achieves optimal decision-making and autonomous upgrading of the model while ensuring the stability of the decision results.

[0013] (2) Technical solution

[0014] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a model-data jointly driven greenhouse intelligent fertilizer distribution system, comprising a control sensing module, an image analysis module and a task decision module;

[0015] The control sensor module includes a soil sensor network and an irrigation vehicle with an electronic control processor control program and a communication program;

[0016] The hardware of the image analysis module consists of the visual acquisition gripper of the irrigation vehicle, high-performance edge computing equipment, and large-capacity data storage media;

[0017] The image analysis module software includes a growth stage detection network and a nutrient deficiency detection network. The growth stage detection network inputs the overall image of the plant in a normal development state and images of the stems, leaves, flowers and fruits, and outputs confidence vectors for the plant in each growth stage. The nutrient deficiency detection network inputs the overall image of the plant in a nutrient deficiency state and images of the stems, leaves, flowers and fruits, and outputs a degree coefficient vector of the nutrients that the plant may be lacking.

[0018] The task decision module hardware includes edge computing devices and cloud servers;

[0019] The task decision module software includes a fertilizer expert system and a deep decision neural network based on reinforcement learning.

[0020] Preferably, the soil sensing network is composed of a plurality of soil integrated sensors and is linked to a cloud information system;

[0021] The irrigation vehicle is equipped with a mechanical gripper equipped with a camera and necessary lighting equipment;

[0022] The irrigation vehicle is provided with multiple irrigation sides, each of which is equipped with a mechanical gripper;

[0023] The liquid distribution tank on the irrigation vehicle should be equipped with an integrated sensor;

[0024] The feeding port of the filling truck is equipped with a pressure sensor.

[0025] Preferably, the growth stage detection network and the nutrient deficiency detection network both comprise a convolutional layer, a pooling layer, a fully connected layer and an output layer.

[0026] Preferably, the fertilizer expert system consists of a model library, a database and an inference engine. The model library establishes and stores mathematical models of fertilizer distribution for various plants based on the experience of greenhouse planting experts. The database stores knowledge information and growth requirement data of different crops. The inference engine can obtain the current optimal fertilizer material ratio and irrigation amount decision results based on the mathematical model of fertilizer distribution decision in the model library.

[0027] Preferably, the deep decision neural network based on reinforcement learning uses DQN as the network framework for fertilizer decision-making, explains the fertilizer task from the perspective of mathematical modeling, and the decision goal is to minimize the modulus of the crop undernutrition coefficient vector. The fertilizer decision is explained by the Markov process. The state is the growth characteristic solution data and the soil sensor return data. The action is the fertilizer material ratio, irrigation amount, and EC value. The reward is the opposite of the modulus of the undernutrition coefficient vector. The training of the decision network parameters is to make fertilizer decisions and obtain the rewards of this round by obtaining the crop growth status of each round. The TD algorithm is used to iterate the training until the parameters converge.

[0028] A model-data-co-driven intelligent fertilizer distribution method for greenhouses includes the following steps:

[0029] Step 1: Initialize the task data for the filling vehicle and take out the fertilizer raw materials at the same time;

[0030] Step 2: Use the irrigation vehicle to obtain crop images, and then use the growth stage detection network and the nutrient deficiency detection network to analyze the image features;

[0031] Step 3: Perform data preprocessing based on the task decision module, integrate the deep decision neural network based on reinforcement learning to make fertilizer allocation decisions, and selectively update the decision network by comparing historical decision data;

[0032] Step 4: After receiving the decision result from the task decision module, the irrigation vehicle executes the fertilizer distribution task.

[0033] Preferably, in the first step, the irrigation vehicle initializing task data includes synchronizing operation information, determining the time, location and object of task execution, obtaining the operation route or performing preliminary task route planning;

[0034] If an automatic fertilizer filling pile is installed when taking fertilizer raw materials, the filling vehicle will execute the automatic fertilizer taking program, otherwise it will prompt manual fertilizer filling, and the pressure sensor will accurately weigh the filling amount of various raw fertilizers.

[0035] Preferably, the second step includes the following contents:

[0036] Step 1: The irrigation vehicle moves to the task point corresponding to a certain crop block, obtains sensor information for that crop block, determines the crop type and the number of features to be extracted, then extends the gripper, adjusts the pose of the object to be photographed, and calls the camera to capture the image. The captured image is then sent to the edge computing device for further processing.

[0037] Step 2: The captured images are input into a convolutional neural network for feature extraction and growth state recognition. The output results of the growth stage detection network and the nutrient deficiency detection network are packaged into growth feature solution data, uploaded to the cloud database and input into the task decision module for further decision-making.

[0038] Step 3: Repeat the above steps until all features to be extracted are completed.

[0039] Preferably, the third step includes the following:

[0040] Step 1: Remove abnormal data, fill in missing values, and normalize. The data includes crop growth characteristic calculation data and soil moisture, salt ion concentration, pH value, EC value, and temperature information uploaded in real time by the soil sensor network.

[0041] Step 2: The decision result of the deep decision neural network based on reinforcement learning will be used as the final decision output. However, before the output, the decision result of the expert system needs to be used as a supervision value to limit the range to eliminate unreasonable decisions caused by accidental factors.

[0042] Step 3: The system memorizes the crop growth and nutritional deficiencies of this round of fertilizer allocation decisions. If the results are better than the historical optimal decisions, the results of this round of decisions are stored in the experience database.

[0043] Preferably, the fourth step includes the following:

[0044] Step 1: After receiving the decision result from the task decision module, the irrigation vehicle opens the leakage port of the fertilizer raw material, controls the fertilizer raw material to leak into the liquid mixing chamber and starts to inject water. The motor drives the abrasive cutter to rotate and crush the granular fertilizer to accelerate the dissolution, forming a water-fertilizer nutrient solution within the set running time;

[0045] Step 2: The irrigation vehicle evenly sprays the prepared water and fertilizer solution into the soil near the roots of the target plants, completing the system process.

[0046] (3) Beneficial effects

[0047] Compared with the existing technology, the present invention provides a model-data-driven intelligent fertilizer distribution method and system for greenhouses, which has the following beneficial effects:

[0048] 1. This model-data-driven intelligent greenhouse fertilizer allocation method and system achieves high fertilizer allocation decision-making accuracy, excellent security and stability, and self-upgradability. This is achieved by combining the proposed decision-making method with a convolutional neural network-based machine vision crop fertilizer deficiency detection method and designing a model-data fusion-driven decision-making solution.

[0049] 2. This model-data-driven intelligent greenhouse fertilizer distribution method and system demonstrates the high robustness and enhanced generalization capabilities of convolutional neural network models compared to regression analysis. Furthermore, this invention proposes a dual-network detection mechanism that deeply mines crop nutrient deficiencies from two dimensions of crop imagery, significantly improving detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a schematic diagram of an overview of the system of the present invention;

[0051] Figure 2 This is a schematic diagram of the control sensor module;

[0052] Figure 3 This is a schematic diagram of the image parsing module;

[0053] Figure 4 This is a schematic diagram of the task decision module;

[0054] Figure 5 Schematic diagram of the system task initialization process;

[0055] Figure 6 Schematic diagram of the process for acquiring crop images and analyzing features;

[0056] Figure 7 This is a schematic diagram of the fertilizer allocation task decision-making process;

[0057] Figure 8 This is a schematic diagram of the fertilizer distribution task execution process;

[0058] Figure 9 Schematic diagram of a mechanical gripper with a camera. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0060] See also Figures 1-4 as well as Figure 9 ,A model-data jointly driven greenhouse intelligent fertilizer distribution system includes a control sensing module, an image parsing module, and a task decision module.

[0061] Control sensor module:

[0062] The main function of the control sensing module is to sense soil nutrient data in real time, control the irrigation vehicle to receive the fertilizer allocation decision information from the task decision module, and execute the fertilizer allocation process and irrigation operations.

[0063] like Figure 2 As shown in FIG, the module hardware includes a soil sensor network and an irrigation vehicle.

[0064] First, to optimize the system's fertilizer allocation decisions, integrated soil sensors are deployed in the corresponding crop planting areas within the greenhouse to obtain real-time information about the fertilizer solution's moisture content, nutrient ion concentration (including at least nitrogen, phosphorus, and potassium), pH, and temperature. Multiple integrated soil sensors are networked and connected to a cloud-based information system, enabling duplex information transmission. This network is called a soil sensor network.

[0065] Secondly, fertilizer allocation requires a fertilizer application vehicle that can execute fertilizer allocation decisions. The necessary hardware for the fertilizer application vehicle is as follows:

[0066] (1) Assemble a mechanical gripper with a camera and necessary lighting equipment (hereinafter referred to as the "gripper"). Figure 9 The gripper ensures that the camera extends a certain distance from the vehicle body, and can separate overlapping stems and leaves, flatten leaves, and clamp fruits, so that the camera can clearly capture images of the stems, leaves, flowers, and fruits of the crops.

[0067] (2) Multiple irrigation sides are set up to complete irrigation work in multiple directions at the same time. Each irrigation side is equipped with a "claw" to facilitate multi-threaded collection of real-time crop growth status information.

[0068] (3) The liquid distribution tank should be equipped with sensors. Irrigation should be equipped with sensors that can obtain real-time information such as the humidity, nutrient ion content (including at least nitrogen, phosphorus and potassium), EC value (soluble salt concentration), pH value, temperature, etc. of the fertilizer solution. This information can be expressed using a comprehensive sensor.

[0069] (4) The feed port is equipped with a pressure sensor. The irrigation vehicle needs to complete real-time fertilizer allocation according to crop demand. For example, if high-nitrogen, high-phosphorus, and high-potassium fertilizers are used as fertilizer raw materials and mixed according to a certain mass ratio, a pressure sensor is required at the feed port to determine the actual fertilizer input quality to ensure the accurate execution of the fertilizer allocation decision results.

[0070] (5) Electronic control processor. The processor runs the whole vehicle control program and communication program of the irrigation vehicle.

[0071] Image parsing module:

[0072] The main function of the image analysis module is to extract and analyze the crop images uploaded by the irrigation vehicle, and to fit the current crop growth status and nutrient deficiency detection results through a well-trained convolutional neural network.

[0073] The hardware of this module mainly consists of the visual acquisition gripper of the irrigation vehicle, high-performance edge computing equipment and large-capacity data storage media, such as Figure 3 shown.

[0074] The software for this module deploys two convolutional neural networks, functionally called a growth stage detection network and a nutrient deficiency detection network. The network structure must include convolutional layers, pooling layers, fully connected layers, and an output layer.

[0075] The growth stage detection network takes as input an image of the plant as a whole and of its stems, leaves, flowers, and fruits in a normal growth state; the output is a confidence vector (i.e., probability) of the plant being in each growth stage (e.g., seedling, active growth, flowering, and maturity). (For example, if the output vector is [0.00, 0.05, 0.05, 0.90], the probabilities of the plant being in the seedling, active growth, flowering, and maturity stages are 0.00, 0.05, 0.05, and 0.90, respectively. Therefore, the crop can be considered mature with a high degree of confidence.) This network is primarily used to identify the target crop's growth stage and determine whether it is ahead of or behind the theoretical growth stage, thereby determining whether the crop is experiencing developmental delay or premature maturity.

[0076] For the nutrient deficiency detection network, the input is an overall image of a plant in a nutrient deficiency state (such as nitrogen, phosphorus, or potassium deficiency) and images of its stems, leaves, flowers, and fruits (e.g., yellowing leaves, stunted plants, etc.); the output is a coefficient vector indicating the degree of nutrient deficiency in the plant. The vector elements range from 0 to 1. For example, if the output layer outputs the vector [0.3, 0.2, 0.2], this indicates that the crop's nitrogen, phosphorus, and potassium deficiencies are 0.3, 0.2, and 0.2, respectively. The closer the coefficient is to 1, the greater the deficiency in that particular nutrient. The dimensionality of the output vector can be expanded by the technical implementer based on the nutrient categories to be examined.

[0077] Regarding the training of the weight parameters of the convolutional neural network, it is necessary to prepare two types of images, overall and local, of various greenhouse crops at various growth stages and their corresponding labels. The labels should be calibrated by professional greenhouse operation practitioners.

[0078] Two convolutional neural networks were trained using the two types of images and their corresponding labels. The resulting networks were then used to determine the growth stage of a crop plant and identify the nutrient deficiencies associated with abnormal traits.

[0079] Task decision module:

[0080] The main function of the task decision module is to complete fertilizer allocation decisions.

[0081] like Figure 4As shown in the figure, the hardware foundation of this module is a high-performance edge computing device and a cloud server. The former needs to be deployed near the greenhouse. In addition to solving the crop growth characteristics in the image analysis module, it also performs real-time fertilizer allocation and material decision-making. The latter is used to record the growth status and coordinate data of crops in various areas of multiple greenhouses, perform macro-scheduling and overall control of the system, and provide a human-computer interaction interface for remote monitoring and operation of the system by practitioners.

[0082] The software part of this module consists of a fertilizer expert system and a deep decision neural network based on reinforcement learning (hereinafter referred to as decision neural network or decision network).

[0083] The fertilizer expert system consists of a model library, a database, and an inference engine. The model library, based on the experience of greenhouse cultivation experts, establishes and stores mathematical models for fertilizer allocation for various plants. The database stores knowledge and growth requirements for different crops. The inference engine, using the mathematical models for fertilizer allocation decisions in the model library, determines the optimal fertilizer ratio and irrigation rate.

[0084] A deep decision neural network based on reinforcement learning uses DQN (Deep Q Learning) as the network framework for fertilizer allocation decisions. Explaining the fertilizer allocation task from a mathematical modeling perspective, the decision-making goal is to minimize the modulus of the crop undernutrition coefficient vector (i.e., to keep the value of each element of the vector as close to 0 as possible). Fertilization decisions are explained using a Markov process: the states are growth characteristic solution data and soil sensor feedback data; the actions are fertilizer material ratios, irrigation rates, and EC values; and the reward is the inverse of the modulus of the undernutrition coefficient vector. To train the decision network parameters, the system obtains the crop growth status for each round, makes fertilizer allocation decisions (fertilizer material ratios, fertilizer rates, and EC values), and receives rewards for each round. Training is then iterated using the TD algorithm (temporal difference algorithm) until the parameters converge.

[0085] When the system actually makes fertilizer allocation decisions, the fertilizer expert system and the decision neural network simultaneously make model-based and data-based decisions, respectively. The expert system's results serve as initialization data during initial network training to save exploration time for the reinforcement learning algorithm. They also serve as supervisory values ​​to evaluate the network's subsequent decision-making results. If the decision result data exceeds a certain range of supervisory values, the current decision is rejected and the decision is recalculated. (The specific ranges and thresholds can be determined by technical implementers based on actual circumstances and are not limited by this proposal.)

[0086] The fertilizer expert system mainly relies on model decision-making, and the decision network mainly relies on data-driven and training to make fitting decisions. Therefore, the entire system is called a fusion decision-making system jointly driven by models and data.

[0087] See Figure 5-Figure 8 A greenhouse intelligent fertilizer distribution method driven by a model-data combination includes the following steps:

[0088] Step 1: System task initialization

[0089] (1) The irrigation vehicle initializes task data. This process includes synchronizing operation information, determining the time, location, and object of task execution, and obtaining the operation route or performing preliminary task route planning.

[0090] (2) Fertilizer raw material collection. If the system is equipped with an automatic fertilizer filling pile, the filling truck will automatically perform the fertilizer collection process. Otherwise, it will prompt manual fertilizer filling. The pressure sensor accurately weighs the amount of various raw fertilizers (such as high nitrogen fertilizer, high phosphorus fertilizer, high potassium fertilizer, etc.).

[0091] Step 2: Acquire crop images and parse features

[0092] (1) The irrigation vehicle moves to the task point corresponding to a certain crop block, obtains the sensor information of the crop block, determines the crop type and the number of features to be extracted, and then extends the gripper to adjust the posture of the object to be photographed (such as moving away overlapping stems and leaves, flattening the leaves to be photographed, etc.) and then calls the camera to shoot the image. The shooting results will be sent to the edge computing device for further processing.

[0093] (2) The captured images are input into a convolutional neural network to perform feature extraction and growth state recognition. The output results of the growth stage detection network and the undernutrition detection network are packaged into growth feature solution data, uploaded to the cloud database and input into the task decision module for further decision-making.

[0094] (3) Repeat steps (1) and (2) until all features to be extracted are completed.

[0095] Step 3: Fertilizer task decision

[0096] (1) Data preprocessing. The input of the task decision module is divided into two parts. One is the crop growth feature solution data uploaded by the image analysis module. This part of the data has been normalized in the output link of the neural network in process 2. The other is the soil moisture, salt ion concentration, pH value, EC value and temperature information uploaded in real time by the soil sensor network. This part of the data requires data preprocessing, including removing abnormal data, supplementing missing values ​​and normalization.

[0097] (2) The fusion decision system is used to make fertilizer allocation decisions. The decision result of the decision neural network will be used as the final decision output, but before the output, the decision result of the expert system needs to be used as a supervisory value to limit the range in order to eliminate unreasonable decisions under the influence of accidental factors.

[0098] (3) Selectively update the decision network by comparing historical decision data. The system memorizes the crop growth and nutrient deficiencies of the current round of fertilizer allocation decisions. If the decision is better than the historical optimal decision, the decision is stored in the experience database. This data will be used to train and update network parameters or support related research by practitioners.

[0099] Step 4: Fertilizer task execution

[0100] (1) After receiving the decision result of the task decision module, the irrigation vehicle opens the leakage port of the fertilizer raw material, controls the fertilizer raw material to leak into the liquid mixing chamber and starts to inject water. The motor drives the abrasive tool to rotate and crush the granular fertilizer to accelerate the dissolution, forming a water-fertilizer nutrient solution within the set operating time.

[0101] (2) The irrigation vehicle evenly sprays the prepared water and fertilizer solution into the soil near the roots of the target plants, completing the system process.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A greenhouse intelligent fertilizer distribution system driven by model and data, characterized by: Includes control sensing module, image analysis module and task decision module; The control sensing module includes a soil sensing network and an irrigation vehicle including an electronic control processor control program and a communication program; The hardware of the image analysis module consists of the visual acquisition gripper of the irrigation vehicle, high-performance edge computing equipment, and large-capacity data storage media; The image analysis module software includes a growth stage detection network and a nutrient deficiency detection network. The growth stage detection network inputs the overall image of the plant in a normal development state and images of the stems, leaves, flowers and fruits, and outputs confidence vectors for the plant in each growth stage. The nutrient deficiency detection network inputs the overall image of the plant in a nutrient deficiency state and images of the stems, leaves, flowers and fruits, and outputs a degree coefficient vector of the nutrients that the plant may be lacking. The task decision module hardware includes edge computing devices and cloud servers; The task decision module software includes a fertilizer expert system and a deep decision neural network based on reinforcement learning; The soil sensing network includes a plurality of soil integrated sensors and is linked to a cloud information system; The irrigation vehicle is equipped with a mechanical gripper equipped with a camera and necessary lighting equipment; The irrigation vehicle is provided with multiple irrigation sides, each of which is equipped with a mechanical gripper; The liquid distribution tank on the irrigation vehicle should be equipped with an integrated sensor; The feeding port of the filling vehicle is equipped with a pressure sensor; The fertilizer expert system consists of a model library, a database, and an inference engine. The model library establishes and stores mathematical models for fertilizer allocation for various plants based on the experience of greenhouse planting experts. The database stores knowledge information and growth demand data for different crops. The inference engine, based on the mathematical models for fertilizer allocation decisions in the model library, can determine the current optimal fertilizer material ratio and irrigation amount decision results. The reinforcement learning-based deep decision neural network uses DQN as the network framework for fertilizer allocation decisions. The fertilizer allocation task is explained from the perspective of mathematical modeling. The decision-making goal is to minimize the modulus of the crop undernutrition coefficient vector. The fertilizer allocation decision is explained by the Markov process. The state is the growth characteristic solution data and the data returned by the soil sensor. The action is the fertilizer material ratio, irrigation amount, and EC value. The reward is the opposite of the modulus of the undernutrition coefficient vector. The training of the decision network parameters is to make fertilizer allocation decisions and obtain the rewards of this round by obtaining the crop growth status in each round. The TD algorithm is used for repeated iterative training until the parameters converge.

2. The greenhouse intelligent fertilizer distribution system driven by model and data according to claim 1 is characterized by: The growth stage detection network and the undernutrition detection network both include a convolutional layer, a pooling layer, a fully connected layer and an output layer.

3. A greenhouse intelligent fertilizer distribution method driven by model and data, characterized in that: The system according to any one of claims 1 to 2 comprises the following steps: Step 1: Initialize the task data for the filling vehicle and take out the fertilizer raw materials at the same time; Step 2: Use the irrigation vehicle to obtain crop images, and then use the growth stage detection network and the nutrient deficiency detection network to analyze the image features; Step 3: Perform data preprocessing based on the task decision module, integrate the deep decision neural network based on reinforcement learning to make fertilizer allocation decisions, and selectively update the decision network by comparing historical decision data; Step 4: After receiving the decision result of the task decision module, the irrigation vehicle performs the fertilizer distribution task.

4. The greenhouse intelligent fertilizer distribution method driven by model and data combination according to claim 3 is characterized by: In the first step, the irrigation vehicle initializes task data including synchronizing operation information, determining the time, location and object of task execution, obtaining the operation route or performing preliminary task route planning; If an automatic fertilizer filling pile is installed when taking fertilizer raw materials, the filling vehicle will execute the automatic fertilizer taking program, otherwise it will prompt manual fertilizer filling, and the pressure sensor will accurately weigh the filling amount of various raw fertilizers.

5. The greenhouse intelligent fertilizer distribution method driven by model and data combination according to claim 4 is characterized by: The second step includes the following: Step 2.1: The irrigation vehicle moves to the task point corresponding to a certain crop block, obtains sensor information for that crop block, determines the crop type and the number of features to be extracted, then extends the gripper, adjusts the pose of the object to be photographed, and calls the camera to capture the image. The captured image is then sent to the edge computing device for further processing; Step 2.2: The captured images are fed into a convolutional neural network for feature extraction and growth state recognition. The outputs of the growth stage detection network and the undernutrition detection network are packaged into growth feature solution data, which are then uploaded to a cloud database and fed into the task decision module for further decision-making. Step 2.3: Repeat steps 2.1-2.3 above until all features to be extracted are completed.

6. The greenhouse intelligent fertilizer distribution method driven by model and data combination according to claim 5 is characterized by: The third step includes the following: Step 3.1: Remove abnormal data, fill in missing values ​​and normalize. The data includes crop growth characteristic calculation data and soil moisture, salt ion concentration, pH value, EC value and temperature information uploaded in real time by the soil sensor network; Step 3.2: The decision result of the deep decision neural network based on reinforcement learning will be used as the final decision output. However, before the output, the decision result of the expert system needs to be used as a supervision value to limit the range to eliminate unreasonable decisions caused by accidental factors. Step 3.3: The system memorizes the crop growth and nutritional deficiencies of this round of fertilizer allocation decision-making results. If they are better than the historical optimal decision-making results, the decision-making results of this round will be stored in the experience database.

7. The greenhouse intelligent fertilizer distribution method driven by model and data combination according to claim 6 is characterized by: The fourth step includes the following: Step 4.1: After receiving the decision result from the task decision module, the irrigation vehicle opens the fertilizer raw material leakage port, controls the fertilizer raw material to leak into the liquid mixing chamber, and begins to inject water. The motor drives the abrasive cutter to rotate and crush the granular fertilizer to accelerate its dissolution, forming a water-fertilizer nutrient solution within the set running time; Step 4.2: The irrigation vehicle evenly sprays the prepared water and fertilizer solution into the soil near the roots of the target plants, completing the system process.

Images