An intelligent monitoring and control system for agricultural greenhouse

Abstract

The present application belongs to the technical field of intelligent monitoring and control, and relates to an intelligent monitoring and control system for agricultural greenhouse. The present application codes the control actions as vectors and samples to construct a sample set by collecting the data of multiple environmental factors in the greenhouse, detects the environmental fluctuation convergence state after the control is executed to construct the action-environment mapping relationship, and then generates candidate actions based on the optimal interval of the crop stage, selects the optimal control combination through the prediction model and real-time state scoring. The system solves the problems of traditional greenhouse environment control, such as dependence on manual operation, slow response, low precision, difficulty in multi-device cooperation, difficulty in dynamically adapting to the changes of crop growth stages, and low stability and energy efficiency, and can realize rapid and accurate adjustment of multiple factors such as temperature, light, water and air, significantly improving the automation level of environmental control, the quality of crop growth and the efficiency of resource utilization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and control technology, and relates to an intelligent monitoring and control system for agricultural greenhouses. Background Technology

[0002] In modern agricultural production, greenhouses serve as the core carrier of controllable environment agriculture. By artificially regulating key environmental factors such as temperature, humidity, light, and ventilation, they can overcome the limitations of the natural environment, enabling off-season crop planting, improved quality, and stable yields. As facility agriculture develops towards precision and intelligence, greenhouse environmental control technology has gradually evolved from traditional manual operation to automation. For example, the agricultural greenhouse monitoring system proposed in Chinese invention patent publication number CN106125803A uses various sensors to collect environmental parameters, which are then uploaded to a monitoring center via a wireless network. When real-time parameters exceed preset thresholds, corresponding control commands are automatically triggered.

[0003] However, this system relies on fixed-threshold control logic, which is a static decision-making mechanism and has significant limitations. On the one hand, the environmental requirements of crops change dynamically at different growth stages, and the fixed-threshold strategy is difficult to adapt to these dynamic needs during the growth process. On the other hand, there are complex coupling relationships among various environmental factors in the greenhouse. For example, ventilation not only affects temperature but also changes humidity and carbon dioxide concentration. Traditional control systems struggle to effectively coordinate the interactions between these multiple factors, easily triggering unexpected chain reactions that lead to drastic fluctuations in environmental parameters, thereby reducing control stability and affecting the optimization of crop growth.

[0004] Furthermore, the development of precise and intelligent facility agriculture not only requires automation technology, but also the ability to predict environmental changes after equipment regulation in order to effectively avoid control deviations. Moreover, existing prediction models mostly adopt a single-task learning paradigm, which only models a single environmental factor independently, such as predicting only temperature changes. This ignores the nonlinear coupling relationship and collaborative response mechanism between multiple factors such as temperature, humidity, light, and ventilation, resulting in one-sided and distorted prediction results that cannot reflect the real environmental evolution trend. In the scenario of multi-equipment collaborative regulation, the system cannot accurately predict the comprehensive impact of combined actions on multiple factors, causing errors in the selection of regulation actions, leading to the environment deviating from the optimal range and triggering crop stress or disease risks. Summary of the Invention

[0005] In view of this, in order to solve the problems mentioned in the background technology, an intelligent monitoring and control system for agricultural greenhouses is proposed.

[0006] The objective of this invention can be achieved through the following technical solution: an intelligent monitoring and control system for agricultural greenhouses, comprising: a data acquisition unit for real-time acquisition of a sequence of environmental parameters composed of multiple environmental factors within the greenhouse to be monitored.

[0007] The control sampling unit is used to encode the corresponding control actions of the environmental control equipment into control action vectors, and to construct a basic action combination sample set based on the control action vectors through a key point sampling strategy.

[0008] The mapping construction unit is used to execute each sample control action combination sequentially according to the basic action combination sample set, and combine the environmental parameter sequence to construct a mapping dataset of control action-environment state through fluctuation convergence state detection and difference calculation.

[0009] The regulation and optimization unit is used to dynamically generate a set of candidate regulation action combinations based on the optimal environmental parameter range of the current crop growth stage. Each candidate regulation action is input into the prediction model trained on the mapping dataset, and the optimal regulation action combination is selected by combining the real-time environmental state vector through quantitative scoring.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention constructs a set of candidate control action combinations based on the optimal environmental parameter range of the crop growth stage, which solves the problem that the traditional fixed threshold control strategy cannot adapt to the dynamic needs of crop growth. This enables the system to automatically adjust the control strategy according to the crop growth stage, accurately match the environmental needs of each growth stage, significantly improve the adaptability and control accuracy of the crop growth environment, and realize the intelligent and personalized environmental control.

[0011] (2) This invention constructs a basic action combination sample set through a key point sampling strategy and automatically divides single-device, dual-device collaborative and multi-device linkage actions, which solves the problem that traditional control systems have difficulty coordinating multi-factor coupling relationships. The system can automatically identify and eliminate logical conflict actions, effectively coordinate the collaborative work of multiple devices such as ventilation and shading, avoid unexpected environmental fluctuations caused by control actions, and greatly improve the stability of system control and the degree of crop growth optimization.

[0012] (3) By constructing a prediction model with a multi-task learning architecture, this invention solves the problem that existing single-task models cannot model the nonlinear coupling relationship of multiple factors. The model can simultaneously predict the changes of multiple environmental factors, accurately capture the complex coupling relationship between factors such as temperature, humidity, and CO2 concentration, and provide systematic and complete prediction results, providing comprehensive and reliable data support for the collaborative control decision of multiple execution systems. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram showing the connections of each unit in the system of the present invention.

[0015] Figure 2 This is a flowchart of the mapping dataset construction method in this invention.

[0016] Figure 3 This is a structural diagram of the prediction model in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, the present invention provides an intelligent monitoring and control system for agricultural greenhouses, comprising: a data acquisition unit, a control sampling unit, a mapping construction unit, and a control optimization unit. The connection relationships between the units are as follows: the data acquisition unit and the mapping construction unit are connected, the control sampling unit and the mapping construction unit are connected, and the mapping construction unit and the control optimization unit are connected.

[0019] The data acquisition unit is used to collect in real time the environmental parameter sequence consisting of multiple environmental factors inside the greenhouse to be monitored.

[0020] In one specific embodiment, the method for acquiring the multidimensional environmental parameters is as follows: the original environmental parameter sequence is acquired in real time at fixed time intervals through an integrated environmental sensor network, and the original environmental parameter sequence is timestamped and outlier detected and corrected to obtain an environmental parameter sequence, which includes environmental factors such as temperature, humidity, CO2 concentration and light intensity.

[0021] The control sampling unit is used to encode the corresponding control actions of the environmental control equipment into control action vectors, and to construct a basic action combination sample set based on the control action vectors through a key point sampling strategy.

[0022] Furthermore, the method for encoding the corresponding control actions of the environmental control equipment into control action vectors is as follows: select the corresponding control equipment for the ventilation system, shading system, supplemental lighting system, and humidification / dehumidification system according to the environment of the greenhouse to be monitored.

[0023] The control actions of the corresponding devices in the aforementioned systems are uniformly encoded into control action vectors. ,in This indicates the percentage of ventilation opening; its default value is [value missing]. ; This indicates the percentage of shade net opening; its default value is [value missing]. ; This indicates the percentage of fill light intensity; its default value is [value missing]. ; and These are binary switch signals used to control the start and stop of the humidifier and dehumidifier respectively; their default values ​​are both... .

[0024] It should be noted that the default value means that the device remains off or in a zero-output state. When a certain dimension of the control action vector is set to the default value, the device will not execute new control commands. The ventilation opening is set to 50% by default to maintain basic air circulation in the greenhouse and avoid drastic fluctuations in temperature and humidity caused by complete closure or excessive ventilation. The shade net is set to 0% by default to prevent strong light from scorching the seedlings in the early stage of system startup. The supplemental lighting and humidification / dehumidification equipment are turned off by default to save energy and wait for the system to be dynamically activated based on sensor data.

[0025] It should be explained that the aforementioned control action vector uniformly expresses the control actions of different devices such as ventilation, shading, supplemental lighting, and humidification / dehumidification in a standardized and structured manner. It integrates continuous parameters and discrete switching quantities into the same numerical vector, which not only eliminates device differences but also provides a mathematical basis for building a sample set of basic action combinations covering all control strategies. At the same time, this encoding serves as a structured input to the prediction model, ensuring the feasibility of model training and the interpretability of decisions. Furthermore, by setting default values ​​for each dimension, it clarifies the baseline state without intervention, thus avoiding device logic conflicts.

[0026] Furthermore, the method for constructing the basic action combination sample set is as follows: the control action vector is divided into single-device actions, dual-device collaborative actions, and multi-device linkage actions.

[0027] The method for classifying single-device actions, dual-device collaborative actions, and multi-device linkage actions is as follows: automatic classification is performed based on the number of dimensions in the control action vector that are actively set to non-default values, and logical conflict actions that simultaneously activate humidification and dehumidification devices are automatically excluded during the classification process.

[0028] A single device action refers to a control action where only one dimension is a non-default value.

[0029] Dual-device collaborative action refers to a control action in which exactly two dimensions are non-default values ​​and the combination of dimensions does not conflict.

[0030] Multi-device linkage action refers to a control action where three or more dimensions are non-default values ​​and the combination of dimensions does not conflict.

[0031] For example, a single device can move as follows: This indicates that only the ventilation opening is reduced; This indicates that only the shade net is fully opened, or that only the humidifier is turned on. This type of action is used to assess the independent impact of a single device on environmental factors.

[0032] Examples of dual-device collaborative operation: This indicates that ventilation should be increased while the shade should be partially opened. This indicates that full-power supplemental lighting and dehumidification are activated simultaneously.

[0033] Multi-device linkage, for example: This indicates high ventilation, full shading, and moderate supplemental lighting. This indicates low ventilation, partial shading, full supplemental lighting, and humidification.

[0034] For each type of control strategy, the control value is selected through a key point sampling strategy within the parameter space consisting of its corresponding continuous parameter dimension and discrete parameter dimension.

[0035] It should be noted that the parameter space formed by the continuous parameter dimension and the discrete parameter dimension refers to the multi-dimensional action space jointly defined by the action parameters of all control devices. Among them: the continuous parameter dimension refers to the control parameters whose values ​​can vary within a certain continuous interval, such as the ventilation opening percentage, the shading net opening percentage, the supplementary light intensity percentage, etc., and its value range is a closed interval [0%, 100%]; the discrete parameter dimension refers to the control parameters whose values ​​can only be selected from a finite discrete set, such as the start / stop switch of the humidifier / dehumidifier, and its value range is a binary set {0, 1}.

[0036] The key point sampling strategy is as follows: along the continuous parameter dimension, select multiple sampling points that include the minimum value, the maximum value, and at least one intermediate representative value.

[0037] Enumerate all legal state combinations along the discrete parameter dimension.

[0038] It should be noted that the legal state combination refers to the remaining combination of switch states that can be safely and effectively executed after excluding conflict scenarios where mutually exclusive devices start simultaneously.

[0039] The continuous dimension sampling points are combined with the discrete dimension state and then subjected to a Cartesian product operation to generate a combination of control values.

[0040] The generated control value combinations are classified into single-device, dual-device, and multi-device control strategies to form a basic action combination sample set covering the corresponding control modes.

[0041] It should be noted that the basic action combination sample set has been actively filtered out invalid samples caused by equipment malfunction, execution failure or external interference.

[0042] In one specific embodiment, all dimensions are combined sequentially to form a hybrid high-dimensional parameter space. For example, if the system contains 3 continuous dimensions and 2 discrete dimensions, its parameter space can be formally represented as: Within this space, the sample generation unit employs a key point sampling strategy for different control strategy categories. It selects representative percentage values ​​in the continuous dimension, traverses effective switch combinations in the discrete dimension, and automatically excludes conflicting combinations such as humidification=1 and dehumidification=1. Finally, it generates structured basic action samples.

[0043] For example, in the 2D continuous subspace formed by ventilation and shading, the system may select four corner points (20%, 20%), (20%, 100%), (80%, 20%), and (80%, 100%); in the 2D discrete subspace formed by humidification and dehumidification, the system selects (0, 0), (1, 0), and (0, 1), and automatically excludes (1, 1).

[0044] It should be explained that the aforementioned control sampling unit, through the coordinated design of action coding standards and precise sampling strategies, systematically organizes scattered equipment control actions and transforms them into analyzable and reusable structured samples, laying the foundation for subsequent intelligent control. On the one hand, vector coding establishes a unified quantitative standard for multi-device control, solving the problem of fragmented action descriptions and providing structured input for subsequent modeling and analysis. On the other hand, key point sampling ensures coverage of core control states while avoiding redundancy and computational burden caused by full sampling. At the same time, by pre-emptively eliminating logical conflicts, it effectively prevents energy waste and equipment damage caused by ineffective operations such as simultaneous humidification and dehumidification. More importantly, this sample set serves as a bridge connecting action execution and environmental response, forming the basis for constructing the mapping relationship between control actions and environmental states, supporting subsequent predictive model training and optimization decisions, and is a key link in achieving precise and intelligent closed-loop control.

[0045] This invention constructs a basic action combination sample set through a key point sampling strategy and automatically classifies single-device, dual-device collaborative, and multi-device linkage actions. This solves the problem that traditional control systems have difficulty coordinating multi-factor coupling relationships. The system can automatically identify and eliminate logically conflicting actions, effectively coordinate the collaborative work of multiple devices such as ventilation and shading, avoid unexpected environmental fluctuations caused by control actions, and significantly improve the stability of system control and the degree of crop growth optimization.

[0046] The mapping construction unit is used to execute each sample control action combination sequentially according to the basic action combination sample set, and combine the environmental parameter sequence to construct a mapping dataset of control action-environment state through fluctuation convergence state detection and difference calculation.

[0047] For further details, please refer to Figure 2As shown, the method for constructing the mapping dataset is as follows: based on the basic action combination sample set, each sample control action combination is executed sequentially.

[0048] Based on the environmental parameter sequence output by the data acquisition unit, the standard deviation of each environmental factor is calculated within a sliding time window. When the standard deviation is less than a preset threshold, the environmental factor is determined to have entered a fluctuation convergence state.

[0049] It should be noted that the sliding time window needs to be set in conjunction with the agricultural greenhouse environmental monitoring scenario. The window duration should cover the minimum period from when the environmental factor is disturbed to when it tends to stabilize. For example, for factors that change slowly, such as temperature and humidity, it is usually set to 5–30 minutes. The specific duration can be adjusted according to the sensitivity of the crop to environmental fluctuations. The sliding step size is set to 1 / 2 or 1 / 3 of the window duration, for example, a 10-minute window and a 5-minute step size, in order to ensure real-time monitoring while avoiding data redundancy and ensuring continuous capture of the fluctuation trend of environmental factors.

[0050] The preset threshold needs to be set according to the characteristics of environmental factors and crop requirements. On the one hand, the natural fluctuation range of the factors themselves should be considered, such as setting the temperature threshold to ±0.5℃ and the humidity threshold to ±2%. On the other hand, it should be dynamically adjusted in combination with the requirements of the crop growth stage. For example, the temperature threshold during the seedling stage can be tightened to ±0.3℃. This threshold is the core basis for determining whether the environment is stable. If the standard deviation of a factor within the sliding window is lower than the preset threshold, it indicates that it is in an acceptable stable range, thus providing a basis for subsequent calculation of the environmental state vector.

[0051] The fluctuation convergence state refers to the state in which the magnitude of changes in environmental factors continuously decreases and tends to stabilize. This state is a sign that the system has determined that the environment has stabilized. Only when all environmental factors have entered this state will the system collect the average value of the factors during that period as the baseline environmental state vector or the response environmental state vector to ensure the accurate mapping relationship between the control action and environmental changes.

[0052] Before implementing the control actions for each sample, when the parameter sequences of each environmental factor have entered the fluctuation convergence state, the mean value of each factor during that period is taken as its baseline environmental state vector.

[0053] After the control action is executed, when the parameter sequence of each environmental factor is detected to have entered the fluctuation convergence state again, and the fluctuation convergence condition is continuously met within the subsequent preset time period, the mean value of each factor during that period is taken as its response environmental state vector.

[0054] It should be explained that by detecting the convergence state of environmental fluctuations, the interference caused by the large fluctuations in environmental factors is eliminated, ensuring that the obtained baseline and response values ​​can truly reflect the stable environment before regulation and the new stable state reached after regulation. At the same time, by extracting the stable environment vectors of the baseline before regulation and the response after regulation, a precise causal relationship is established, thereby clearly quantifying the specific impact of a single regulation action on each environmental factor, laying a data foundation for subsequent training of prediction models and achieving precise regulation optimization.

[0055] For each environmental factor, the difference between its response environmental state vector and the baseline environmental state vector is calculated independently to obtain the change vector of each environmental factor caused by the regulatory action.

[0056] The control action vector, baseline environmental state vector, response environmental state vector, and environmental factor change vector are combined to form a structured sample record.

[0057] By performing logical consistency checks on structured sample records, invalid samples are automatically identified and removed, while valid samples are retained to form a mapping dataset of regulatory actions and environmental states.

[0058] It should be noted that the logical consistency check includes device execution feedback consistency verification and change direction rationality verification.

[0059] Equipment execution feedback consistency verification refers to comparing whether the actual position / state of the equipment is consistent with the command. For example, when the ventilation opening command is 50%, if the feedback is 48%, it passes; if the feedback is 0%, it fails.

[0060] The test for the rationality of the change direction indicates that the humidity should rise when the humidifier is turned on; if it falls, it is marked as invalid.

[0061] It should be explained that the constructed mapping dataset provides a data foundation for the precise regulation and optimization of the intelligent monitoring and control system for agricultural greenhouses. The prediction model needs to rely on the correspondence between the regulation action vector and the environmental state change in this dataset to learn the pattern, so as to predict the effect of different regulation actions in practical applications. At the same time, the system can use this to clarify the causal relationship between regulation actions and environmental changes, and select the regulation scheme that approaches the optimum by combining the optimal environmental range of crops.

[0062] The construction method has three main aspects: First, it is based on data collection in a steady state, requiring that environmental factors before and after regulation are in a state of fluctuating convergence to avoid large fluctuations and ensure clear causal relationships. Second, the data dimensions are complete, with each sample containing the regulation action, baseline environment, response environment, and change vector, covering the entire chain from before regulation to action, after regulation, and change. Third, it automatically removes conflicting or invalid samples through logical consistency checks to ensure that the dataset is real and effective, providing reliable support for model training and decision optimization.

[0063] The regulation and optimization unit is used to dynamically generate a set of candidate regulation action combinations based on the optimal environmental parameter range of the current crop growth stage. Each candidate regulation action is input into the prediction model trained on the mapping dataset, and the optimal regulation action combination is selected by combining the real-time environmental state vector through quantitative scoring.

[0064] Furthermore, the method for obtaining the candidate set of regulatory actions is to obtain the optimal environmental parameter range corresponding to the current crop growth stage;

[0065] Based on the optimal environmental parameter range, candidate actions that meet the following conditions are selected from the control action vector space:

[0066] (a) Actions with only one dimension that is not a default value.

[0067] (b) Actions where any two dimensions are non-default values ​​and there is no logical conflict in the combination of dimensions.

[0068] (c) Actions where three or more dimensions are non-default values ​​and there is no logical conflict in the combination of dimensions.

[0069] Actions that satisfy conditions (a), (b), and (c) are combined to form a candidate set of regulatory actions.

[0070] For example, when the system identifies that the current crop is in the flowering stage, it obtains the associated optimal environmental parameter range, such as temperature. ,humidity In the control action vector space, priority is given to action dimension combinations that can reduce temperature, such as ventilation and shading, and actions that can increase humidity, such as humidification, to generate candidate actions such as [40,0,0,1,0] (ventilation + humidification), [0,60,0,1,0] (shading + humidification), and [30,50,0,1,0] (ventilation + shading + humidification). Conflicting combinations such as [0,0,0,1,1] (humidification + dehumidification) are automatically filtered to form a candidate control action combination set.

[0071] This invention solves the problem that traditional fixed threshold control strategies cannot adapt to the dynamic needs of crop growth by constructing a set of candidate control action combinations based on the optimal environmental parameter range of crop growth stages. This enables the system to automatically adjust the control strategy according to the crop growth stage, accurately match the environmental needs of each growth stage, significantly improve the adaptability and control accuracy of crop growth environment, and realize intelligent and personalized environmental control.

[0072] Furthermore, the training method of the prediction model is as follows: using the combination of the control action vector and the environmental state vector in the mapping dataset as input features, and the corresponding environmental factor change vector as supervision label, the machine learning model is trained to obtain a prediction model for predicting the changes of each environmental factor.

[0073] For further details, please refer to Figure 3 As shown, the prediction model training process adopts a multi-task learning architecture. The model internally sets up a shared feature extraction layer and multiple independent output branches. Each output branch corresponds to the prediction output of the change in an environmental factor. All output branches receive gradient updates synchronously during training and share the parameters of the shared feature extraction layer.

[0074] In one specific embodiment, input features are paired with supervision labels and fed into a machine learning model architecture for end-to-end training. The machine learning model preferably employs a multi-task learning (MTL) architecture, which internally includes:

[0075] A shared feature extraction layer, such as a fully connected neural network or a Transformer encoder, is used to extract general regulatory-environment interaction representations from input features. Multiple independent output branches are used, each corresponding to a task of predicting the change in environmental factors such as temperature, humidity, light intensity, and CO2 concentration. All output branches receive gradient updates synchronously during training and share the parameters of the shared feature extraction layer to improve the model's generalization ability and training efficiency.

[0076] The training objective is to minimize the loss function between the predicted and actual changes of each environmental factor. Common algorithms such as Adam and SGD can be used as optimizers. Hyperparameters such as learning rate, batch size, and number of training rounds can be adjusted according to the actual data scale and convergence.

[0077] The prediction model, given any regulatory action vector and the current environmental state vector, can simultaneously output the expected changes of each environmental factor under the intervention of that action, providing accurate and efficient prediction support for the subsequent strategy optimization unit.

[0078] In other embodiments, the machine learning model is not limited to a neural network structure. It can adopt a regression model with multiple output capabilities, such as support vector regression (SVR), random forest, gradient boosting tree such as XGBoost, and LightGBM, as long as it can receive the combination of regulatory actions and environmental states as input and output multiple environmental factor changes.

[0079] This invention solves the problem that existing single-task models cannot model the nonlinear coupling relationship of multiple factors by constructing a prediction model with a multi-task learning architecture. The model can simultaneously predict the changes of multiple environmental factors, accurately capture the complex coupling relationship between factors such as temperature, humidity, and CO2 concentration, and provide systematic and complete prediction results, providing comprehensive and reliable data support for the collaborative control decision-making of multiple execution systems.

[0080] In a further embodiment, the fusion method of the input features can be vector concatenation, weighted fusion, attention mechanism fusion, etc., and is not limited to a specific form.

[0081] Furthermore, the method for obtaining the optimal combination of control actions is as follows: input each candidate control action in the candidate control action combination set into the prediction model unit to obtain the corresponding predicted values ​​of the changes in each environmental factor.

[0082] Collect the real-time environmental state vector corresponding to the current crop growth stage, and perform vector calculation between the predicted value of the change and the real-time environmental state vector to obtain the expected environmental state vector after executing each candidate action.

[0083] Based on the expected environmental state vector and the optimal environmental parameter range, the quantitative score of each candidate action is calculated, and the candidate control action with the lowest score is selected as the optimal control action combination.

[0084] Furthermore, the method for calculating the quantitative score of each candidate action is as follows: based on the optimal environmental parameter range corresponding to the current crop growth stage, for each candidate action, calculate the minimum deviation between each environmental factor in its expected environmental state vector and the upper and lower boundaries of the corresponding environmental factor range in the optimal environmental parameters.

[0085] If the expected environmental factor value is lower than the lower boundary of the interval, then the difference between the lower boundary and the value is taken.

[0086] If the value is higher than the upper boundary of the interval, then the difference between the value and the upper boundary is taken.

[0087] If it is within the interval, the deviation value is zero.

[0088] Normalize each deviation value and sum them according to preset weights to obtain the quantitative score of the candidate action.

[0089] It should be noted that the preset weights can be customized by the user, adjusted by the system through self-learning, or dynamically configured according to crop type / season. For example, the system has a built-in weight table for tomato crops: seedling stage (temperature 0.5, humidity 0.3...), flowering stage (temperature 0.4, humidity 0.3...), fruiting stage (light 0.4...). Users can also manually adjust the weights through the interface, and the system can automatically optimize the weights based on historical yield data.

[0090] For example, the system obtains the optimal environmental parameter range for the tomato flowering period: temperature 22-26℃, humidity 60-70%, light intensity 8000-10000 lux, and CO2 concentration 800-1000 ppm. The current real-time environmental state vector is: [temperature=28.5℃, humidity=55%, light intensity=7500 lux, CO2=700 ppm]. Based on the optimal environmental parameter range for the tomato flowering period, a set of candidate control action combinations is dynamically generated.

[0091] Candidate action A: Open the ventilation window to 40%. Input action A into the prediction model and output the vector of changes in each factor: [ΔT=-3.2℃, ΔH=-2%, ΔL=0lux, ΔC=0ppm]. Expected environmental state vector: [28.5-3.2, 55-2, 7500+0, 700+0]=[25.3℃, 53%, 7500lux, 700ppm]. The deviation calculation formulas for each factor are: Temperature 25.3℃∈[22,26], distance=0; Humidity 53%<60%, distance=60-53=7; Light intensity 7500<8000; distance= 8000-7500=500; CO2700<800, distance=800-700=100, set weights: [temperature 0.3, humidity 0.2, light 0.3, CO2 0.2], normalize each deviation value and sum the weights = 0×0.3+0.7×0.2+0.25×0.3+0.5×0.2=0.315.

[0092] Candidate action B: Open the ventilation window by 20% + turn on the supplementary light by 60%. Using the same calculation method as action A, the final quantitative score of B is 0.2875.

[0093] Candidate action C: Ventilation window 10% + supplementary lighting 40% + CO2 generator turned on. Using the same calculation method as action A, the final quantitative score C is 0.2475.

[0094] The action C with the lowest score was selected as the optimal control action combination. After action C was executed, the environment will be closest to the ideal range of the tomato flowering period, and the overall control effect will be optimal.

[0095] The parameters involved in the above formula are all dimensionless and calculated numerically. The formula is a formula obtained from the most recent real situation by collecting a large amount of data and simulating it with software. The preset parameters in the formula are set by those skilled in the art according to the actual situation.

[0096] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0097] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0098] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0099] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0100] Finally, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent monitoring and control system for an agricultural greenhouse, characterized by, The application relates to a method for optimizing the environment of a greenhouse, and comprises the following steps: a data acquisition unit is used to acquire an environment parameter sequence composed of multiple environment factors in a greenhouse in real time; a regulation sampling unit is used to encode corresponding regulation actions of an environment regulation device into a regulation action vector, and to construct a basic action combination sample set through a key point sampling strategy according to the regulation action vector; a mapping construction unit is used to sequentially execute each sample regulation action combination according to the basic action combination sample set, and to construct a mapping data set of regulation action-environment state through fluctuation convergence state detection and difference calculation in combination with the environment parameter sequence; a regulation optimization unit is used to dynamically generate a candidate regulation action combination set according to an optimal environment parameter interval of a current crop growth stage, to input each candidate regulation action into a prediction model trained based on the mapping data set, and to select an optimal regulation action combination through quantitative scoring in combination with a real-time environment state vector.

2. The intelligent monitoring and control system for agricultural greenhouse according to claim 1, wherein, The method for encoding the corresponding regulation actions of the environment regulation device into the regulation action vector is as follows: corresponding regulation devices of a ventilation system, a sunshade system, a light supplementing system and a humidifying and dehumidifying system are selected according to the environment of the greenhouse to be monitored; The regulation actions of the corresponding devices of the several systems are uniformly coded as a regulation action vector , wherein represents the percentage of the ventilation opening degree, and the default value is ; ; ; ; ; and are binary switch quantities respectively controlling the start and stop of the humidification and dehumidification devices, and the default values are both .

3. The intelligent monitoring and control system for agricultural greenhouse according to claim 1, wherein, the construction method of the basic action combination sample set is as follows: the regulation action vector is divided into single-device actions, double-device collaborative actions and multi-device linkage actions; for each type of regulation strategy, a regulation value is selected through a key point sampling strategy in a parameter space composed of a continuous parameter dimension and a discrete parameter dimension, as follows: on the continuous parameter dimension, multiple sampling points containing a minimum value, a maximum value and at least one intermediate representative value are selected along the value range thereof; on the discrete parameter dimension, all legal state combinations thereof are enumerated; the continuous dimension sampling points and the discrete dimension state combinations are subjected to Cartesian product operation to generate regulation value combinations; the generated regulation value combinations are classified according to single-device, double-device and multi-device regulation strategies to constitute a basic action combination sample set covering corresponding regulation modes.

4. The intelligent monitoring and control system for agricultural greenhouse according to claim 3, characterized in that, The division method of the single-device action, the double-device collaborative action and the multi-device linkage action is as follows: automatic division is performed according to the number of dimensions in the regulation action vector that are actively set to non-default values, and logical conflict actions of simultaneously activating a humidifying device and a dehumidifying device are automatically excluded in the division process: a single-device action refers to a regulation action with only one dimension being a non-default value; a double-device collaborative action refers to a regulation action with exactly two dimensions being non-default values and with no conflict in the dimension combination; a multi-device linkage action refers to a regulation action with three or more dimensions being non-default values and with no conflict in the dimension combination.

5. The intelligent monitoring and control system for agricultural greenhouse according to claim 1, wherein, The construction method of the mapping data set is as follows: each sample regulation action combination is sequentially executed according to the basic action combination sample set: based on the environment parameter sequence output by the data acquisition unit, the standard deviation of each environment factor is calculated in a sliding time window, and when the standard deviation is less than a preset threshold value, it is determined that the environment factor enters a fluctuation convergence state; before each sample regulation action is executed, when the parameter sequences of all the environment factors are detected to enter the fluctuation convergence state, the mean values of the factors in the period are taken as baseline environment state vectors of the factors, respectively. After performing the regulation action, when it is detected that the parameter sequence of each environmental factor again enters the fluctuation convergence state, and the fluctuation convergence condition is continuously met within the subsequent preset time length, the mean value of each factor in the time period is taken as the response environmental state vector of the factor respectively; For each environmental factor, the difference between its response environmental state vector and baseline environmental state vector is independently calculated to obtain the change amount vector of each environmental factor caused by the regulation action; The regulation action vector, baseline environmental state vector, response environmental state vector, and environmental factor change amount vector are combined to form a structured sample record; Through logical consistency checking on the structured sample record, invalid samples are automatically identified and removed, and valid samples are retained to constitute a mapping data set of regulation action-environmental state.

6. The intelligent monitoring and control system for agricultural greenhouse according to claim 5, wherein, The training method of the prediction model is: The combination of the regulation action vector and the environmental state vector in the mapping data set is taken as the input feature, and the corresponding environmental factor change amount vector is taken as the supervision label to train the machine learning model to obtain the prediction model for predicting the change amount of each environmental factor.

7. The intelligent monitoring and control system for agricultural greenhouse according to claim 6, wherein, The training process of the prediction model includes: A multi-task learning architecture is adopted, and a shared feature extraction layer and multiple independent output branches are set inside the model. Each output branch corresponds to the change amount prediction output of one environmental factor, and all output branches synchronously receive gradient updates and share the parameters of the shared feature extraction layer during training.

8. The intelligent monitoring and control system for agricultural greenhouse according to claim 1, wherein, The acquisition method of the candidate regulation action combination set is: An optimal environmental parameter interval corresponding to the current crop growth stage is obtained; Based on the optimal environmental parameter interval, candidate actions that meet the following conditions are selected from the regulation action vector space: (a) only one dimension is a non-default value action; (b) any two dimensions are non-default values and the dimension combination has no logical conflict; (c) three or more dimensions are non-default values and the dimension combination has no logical conflict; Actions that meet conditions (a), (b), and (c) are combined to form a candidate regulation action combination set.

9. The intelligent monitoring and control system for agricultural greenhouse according to claim 1, wherein, The acquisition method of the optimal regulation action combination is: Each candidate regulation action in the candidate regulation action combination set is input into the prediction model unit to obtain the corresponding environmental factor change amount prediction value; The real-time environmental state vector corresponding to the current crop growth stage is collected, and the change amount prediction value is calculated with the real-time environmental state vector to obtain the expected environmental state vector after performing each candidate action; According to the expected environmental state vector and the optimal environmental parameter interval, the quantitative score of each candidate action is calculated, and the candidate regulation action with the lowest score is selected as the optimal regulation action combination.

10. The intelligent monitoring and control system for agricultural greenhouse according to claim 9, wherein, The method for calculating the quantitative score of each candidate action is: According to the optimal environmental parameter interval corresponding to the current crop growth stage, for each candidate action, the minimum deviation value of each environmental factor in the expected environmental state vector from the upper and lower boundaries of the corresponding environmental factor interval in the optimal environmental parameter is calculated respectively: If the expected environmental factor value is lower than the lower boundary of the interval, the difference between the lower boundary and the value is taken; If it is higher than the upper boundary of the interval, the difference between the value and the upper boundary is taken; If it is within the interval, the deviation value is zero; The deviation values are normalized and summed up by preset weights to obtain a quantitative score of the candidate action.

Images