Greenhouse environment self-adaptive regulation and control system based on artificial intelligence

Abstract

The present application relates to the technical field of agricultural Internet of Things and environmental intelligent control, and particularly relates to a greenhouse environment self-adaptive regulation and control system based on artificial intelligence, comprising: an environment acquisition module, which is used for acquiring multi-dimensional environment data in real time through distributed multi-source sensors; a central controller, which is used for generating an optimized regulation and control strategy; a regulation and control execution module, which is used for driving an execution device to adjust; the central controller comprises a multi-source data fusion unit, an AI decision unit and a dynamic optimization engine, which are respectively responsible for data filtering and fusion, generating an initial regulation and control strategy based on a space-time joint AI model, and reconstructing and optimizing the initial regulation and control strategy through a multi-objective optimization algorithm. The present application improves response real-time performance through multi-source sensing and data fusion, realizes predictive regulation and multi-parameter coordination by using an AI model, balances energy consumption, growth and carbon emission by combining multi-objective optimization, and supports long-term self-adaptation and strategy iteration of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural Internet of Things and environmental intelligent control technology, and particularly relates to a greenhouse environment self-adaptive regulation and control system based on artificial intelligence. BACKGROUND

[0002] In modern facility agriculture, greenhouses create suitable microenvironments for crop growth through artificial regulation and control, which plays a key role in improving the yield and quality of agricultural products. With the wide application of sensor technology and automatic control systems, the regulation and control of greenhouse environments is gradually changing from relying on traditional experience to data-driven intelligent decision-making. However, the existing regulation and control systems still have obvious shortcomings in dynamic environment adaptation, multi-parameter collaborative optimization, and long-term stable operation, which limits the further development of modern agriculture towards high efficiency and intelligence. The existing technology mainly has the following problems:

[0003] 1) Significant lag in environmental regulation: The current common system adopts a regulation and control mechanism with fixed threshold triggering, which only starts the corresponding equipment when a single parameter such as temperature or humidity exceeds the preset range. This passive control method cannot cope with dynamic environmental changes such as sudden changes in light and air disturbance, resulting in a significant lag in the regulation and control response to the actual needs of crops. Especially in large greenhouses, uneven distribution of environmental parameters further leads to problems such as overcooling, overheating, or humidity imbalance in local areas, causing stress to crop growth.

[0004] 2) Lack of multi-dimensional coordination in control strategy: Most systems still use rule-based control methods, such as simply associating temperature with ventilation and humidity with irrigation. This strategy fails to fully consider the coupling relationship between environmental parameters (e.g., temperature rise causing humidity to drop, and light supplementation operation increasing leaf transpiration) and ignores the characteristics of crops at different growth stages. As a result, the execution equipment is frequently started and stopped, resource utilization efficiency is low, and the special environmental needs of crops at different growth stages (such as seedling stage and fruit enlargement stage) cannot be met.

[0005] 3) Imbalance between energy efficiency management and growth regulation: The existing systems often separate environmental control and energy management. In order to quickly adjust temperature and humidity, high-power equipment is often used for continuous operation, resulting in high energy consumption. While emphasizing energy saving can affect environmental stability and is not conducive to crop growth, which reflects the lack of multi-objective optimization capability of the system, making it difficult to achieve the goal of green and low-carbon production.

[0006] 4) Poor long-term adaptability: Traditional systems usually rely on the initial set control logic, which cannot be optimized autonomously according to crop growth feedback, equipment performance degradation or seasonal climate changes. When facing the introduction of new varieties or extreme weather events, manual experience is required to re-adjust parameters, which is complex and requires high professional skills. This static architecture makes the system control effect decline over time, which cannot meet the development needs of the continuous upgrading of facility agriculture.

[0007] The above reasons make it difficult for the prior art to achieve precise, efficient and sustainable control of the greenhouse environment. SUMMARY

[0008] To solve the above technical problems, the present application provides a greenhouse environment adaptive control system based on artificial intelligence.

[0009] The technical problems solved by the present application can be realized by the following technical solutions: a greenhouse environment adaptive control system based on artificial intelligence, comprising: an environment acquisition module for real-time acquisition of multi-dimensional environment data in the greenhouse by distributed deployment of multi-source sensors; a central controller connected to the environment acquisition module for receiving the multi-dimensional environment data and generating an optimized control strategy; a control execution module connected to the central controller for receiving the optimized control strategy and driving the execution equipment to adjust the environment; wherein the central controller comprises: a multi-source data fusion unit for filtering and fusion processing of the multi-dimensional environment data to obtain an environment feature matrix; an AI decision unit connected to the multi-source data fusion unit for extracting and analyzing the spatio-temporal variation characteristics of the environment feature matrix through a spatio-temporal joint AI model to generate an initial predictive control strategy; a dynamic optimization engine connected to the AI decision unit for reconstructing and optimizing the initial predictive control strategy through a multi-objective optimization algorithm to generate an optimized environment control strategy; wherein the optimization objectives of the multi-objective optimization algorithm include device energy consumption, crop growth demand satisfaction degree and carbon emission intensity.

[0010] Preferably, the multi-source data fusion unit comprises: an outlier filtering subunit for real-time detection and elimination of outliers in the multi-dimensional environment data based on a preset dynamic threshold rule; a Kalman filtering subunit connected to the outlier filtering subunit for fusing redundant data of the multi-source sensors and eliminating environmental noise using an improved Kalman filtering algorithm; a time series alignment subunit connected to the Kalman filtering subunit for synchronizing the timestamps of the multi-source sensors to the central controller clock and generating the environment feature matrix.

[0011] Preferably, the AI decision unit comprises: a spatial feature extraction subunit configured to process the environmental spatial distribution features in the environmental feature matrix through a three-dimensional convolutional neural network; a time sequence feature extraction subunit configured to process the environmental time sequence evolution rules in the environmental feature matrix through a gated recurrent unit; and a feature fusion subunit connected to the spatial feature extraction subunit and the time sequence feature extraction subunit, and configured to generate the initial predictive regulation strategy by weighting and fusing the environmental spatial distribution features and the environmental time sequence evolution rules through a multi-head attention mechanism.

[0012] Preferably, the spatio-temporal joint AI model in the AI decision unit is constructed through a model training module, and the model training module is configured to: construct a cross-season training data set covering multiple complete crop growth cycles, the cross-season training data set comprising greenhouse environmental time sequence data, outdoor weather station data, crop hyperspectral images, and yield and quality record data; design a multi-task loss function, the loss function simultaneously optimizing an environmental parameter stability index, a crop accumulated temperature satisfaction rate, and a water resource utilization efficiency; train the spatio-temporal joint AI model using a distributed training framework, and complete model convergence within a preset iteration period; and verify the trained spatio-temporal joint AI model through a verification set.

[0013] Preferably, the dynamic optimization engine is configured to: establish a multi-objective function based on a device energy consumption model, a crop growth demand model, and a carbon emission intensity factor; set a constraint condition that a daily cumulative light intensity of the crop is not less than a light saturation point of the crop; solve the initial predictive regulation strategy using a non-dominated sorting genetic algorithm to generate a Pareto optimal solution set; and select an optimal solution from the Pareto optimal solution set according to a preset preference weight to generate the optimized regulation strategy.

[0014] Preferably, the central controller further comprises a model evolution unit, and the model evolution unit comprises: an incremental data set construction subunit configured to periodically collect crop physiological indexes and construct an incremental data set in combination with historical regulation strategy execution records; and a model optimization subunit connected to the incremental data set construction subunit and configured to fine-tune and optimize the spatio-temporal joint AI model based on online transfer learning technology and using the incremental data set.

[0015] Preferably, the regulation execution module comprises: an instruction compiling unit configured to compile policy parameters in the optimized environment regulation strategy into industrial Internet of Things executable code; a security coding and checking unit connected to the instruction compiling unit and configured to add a time stamp and a cyclic redundancy check code to the executable code and encapsulate the executable code into a control instruction package; an instruction synchronization issuing unit connected to the security coding and checking unit and configured to issue the control instruction package to each execution device through a time stamp synchronization mechanism; and an execution feedback and monitoring unit connected to the instruction synchronization issuing unit and configured to receive and analyze feedback signals of the execution device through a double buffering mechanism and compare the feedback signals with the control instruction package in real time.

[0016] Preferably, the system further comprises an edge intelligent gateway module connected to the environment collecting module and the regulation execution module, wherein the edge intelligent gateway module comprises: a lightweight anomaly detection unit configured to perform real-time anomaly detection on the multi-dimensional environment data based on a lightweight neural network model and generate an emergency signal when an anomaly is identified; and an emergency regulation unit connected to the lightweight anomaly detection unit and configured to start a local emergency regulation protocol to drive the execution device to intervene and upload an event diagnosis report to the central controller after receiving the emergency signal.

[0017] Preferably, the system further comprises an augmented reality interaction terminal connected to the regulation execution module, wherein the augmented reality interaction terminal comprises: a digital twin visualization unit configured to dynamically render an environment parameter cloud map and a virtual state of the execution device based on a three-dimensional model of the greenhouse; and a gesture recognition and interaction unit connected to the digital twin visualization unit and configured to capture and identify a user's predefined gesture action, generate an artificial correction instruction for controlling the virtual execution device according to a preset mapping rule, and trigger a multi-expert decision voting mechanism and generate a strategy evaluation report when a deviation between the artificial correction instruction and the optimized regulation strategy exceeds a safety threshold.

[0018] Preferably, the system further comprises a fault diagnosis and self-recovery module connected to the environment collecting module and the regulation execution module, wherein the fault diagnosis and self-recovery module comprises: a device health monitoring unit configured to monitor working state parameters of the sensor group and the execution device; an intelligent compensation and redundancy switching unit connected to the device health monitoring unit and configured to automatically start a data compensation strategy based on multi-source information fusion or a device switching strategy based on hardware redundancy when detecting abnormal sensor data or execution device failure; and an operation and maintenance unit connected to the intelligent compensation and redundancy switching unit and configured to automatically generate a device maintenance work order according to the fault device information and push the device maintenance work order to a remote management terminal.

[0019] Beneficial effects: through real-time collection and data fusion of distributed multi-source sensors, the comprehensiveness and real-time response of environmental perception are improved; through the extraction of environmental space-time change characteristics by the space-time joint AI model, predictive regulation and multi-parameter collaborative optimization are realized; through multi-objective optimization algorithm collaborative optimization of energy consumption, growth demand and carbon emission, the balance of energy efficiency and regulation is realized; through the combination of the space-time joint AI model and the multi-objective optimization algorithm, long-term self-adaptation and strategy iteration of the system are realized. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a system architecture diagram of the present application;

[0021] Figure 2 is a system module schematic diagram of the present application;

[0022] Figure 3 is a block diagram of the central controller 200 of the present application;

[0023] Figure 4 is a block diagram of the multi-source data fusion unit 210 of the present application;

[0024] Figure 5 is a block diagram of the AI decision unit 220 of the present application;

[0025] Figure 6 is a block diagram of the model evolution unit 240 of the present application;

[0026] Figure 7 is a block diagram of the regulation execution module 300 of the present application;

[0027] Figure 8 is a block diagram of the edge intelligent gateway module 500 of the present application;

[0028] Figure 9 is a block diagram of the augmented reality interactive terminal 600 of the present application;

[0029] Figure 10 is a block diagram of the fault diagnosis and self-recovery module 700 of the present application;

[0030] Figure 11 is a system closed-loop regulation flowchart of the present application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0032] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0033] The present application will be further described below in conjunction with the drawings and specific embodiments, but not as a limitation of the present application.

[0034] With reference to Figures 1 to 3 The present application provides an artificial intelligence-based greenhouse environment adaptive control system 10, comprising: an environment acquisition module 100 for acquiring multi-dimensional environment data in the greenhouse in real time through distributed deployment of multi-source sensors; a central controller 200 connected to the environment acquisition module 100, for receiving multi-dimensional environment data and generating an optimized control strategy; a control execution module 300 connected to the central controller 200, for receiving the optimized control strategy and driving the execution equipment to adjust the environment; wherein the central controller 200 comprises: a multi-source data fusion unit 210 for filtering and fusing the multi-dimensional environment data to obtain an environment feature matrix; an AI decision unit 220 connected to the multi-source data fusion unit 210, for extracting and analyzing the spatio-temporal variation characteristics of the environment feature matrix through a spatio-temporal joint AI model to generate an initial predictive control strategy; a dynamic optimization engine 230 connected to the AI decision unit 220, for reconstructing and optimizing the initial predictive control strategy through a multi-objective optimization algorithm to generate an optimized environment control strategy; wherein the optimization objectives of the multi-objective optimization algorithm include device energy consumption, crop growth demand satisfaction degree, and carbon emission intensity.

[0035] Specifically, in the embodiments of the present application, in order to effectively solve the problems of environmental control lag, lack of multi-parameter coordination, imbalance between energy efficiency and growth, and poor long-term adaptability of the system, the AI decision unit with spatio-temporal perception ability and the dynamic multi-objective optimization engine are introduced, the dynamic changes of environment parameters, crop growth stage characteristics, device energy consumption and carbon emission intensity are included in the unified optimization framework, and the advanced prediction and collaborative control of the greenhouse environment are realized. The system can generate an optimized strategy that takes into account response speed, resource efficiency and low-carbon target based on real-time sensing data and historical operation experience, thereby avoiding the problems of response delay, strategy rigidity and high energy consumption in traditional methods, and significantly improving the precision, efficiency and sustainability of greenhouse environment control.

[0036] Specifically, in the above technical solution, a plurality of high-precision sensors are distributedly deployed in the greenhouse to acquire multi-dimensional environment data in real time, including but not limited to air temperature and humidity sensors, photosynthetic active radiation sensors, carbon dioxide concentration sensors, soil multi-parameter composite sensors, and leaf surface microenvironment monitors, wherein the sampling frequency is not less than 1 time / minute.

[0037] The following is a specific embodiment:

[0038] Embodiment 1: Environmental data collection

[0039] Inside a 400 square meter standard greenhouse, a high-density sensor network consisting of 32 monitoring nodes was deployed, with each node evenly distributed to comprehensively capture environmental variations. Each node integrates multiple high-precision sensors to continuously collect multi-dimensional environmental data at a sampling frequency of no less than once per minute. Specifically, it includes:

[0040] Air temperature and humidity monitoring: PT1000 temperature sensors and capacitive humidity sensors installed at 1.5 meters above ground level, data acquisition every 30 seconds, temperature measurement accuracy up to ±0.3 degrees Celsius, humidity range covers 0 to 100% RH.

[0041] Light monitoring: Quantum-type photosynthetically active radiation sensors suspended from the greenhouse roof, accurately capturing photosynthetically active radiation intensity in the 400 to 700 nanometer spectral range, data output in units of micromoles per square meter per second (μmol / m² / s).

[0042] Carbon dioxide concentration monitoring: monitored by a non-dispersive infrared sensor (NDIR carbon dioxide sensor), with a range of 0 to 5000 ppm.

[0043] Soil environment monitoring: sensed by FDR three-in-one probes (capacitive multi-parameter composite sensors) buried in the crop root zone, installed at a depth of 10 centimeters, capable of simultaneously detecting soil volumetric water content (unit %), temperature (unit °C), and electrical conductivity (unit mS / cm) at depths of 5 to 20 centimeters.

[0044] Crop monitoring: microenvironment monitors clamped on tomato leaves, recording stomatal conductance and leaf condensation status every 5 minutes, data in units of millimoles per square meter per second (mmol / m² / s).

[0045] The entire sensor network integrates air, light, gas, soil, and crop physiological multi-dimensional data sources, forming a real-time, high-precision, and comprehensive greenhouse environmental perception system, providing a reliable data foundation for subsequent intelligent decision-making.

[0046] As a preferred embodiment of the present application, with reference to Figure 4 , the multi-source data fusion unit 210 includes: an outlier filtering sub-unit 211 for real-time detection and elimination of outliers in multi-dimensional environmental data based on preset dynamic threshold rules; a Kalman filter sub-unit 212 connected to the outlier filtering sub-unit 211 for fusing redundant data of multi-source sensors and eliminating environmental noise using an improved Kalman filter algorithm; a time alignment sub-unit 213 connected to the Kalman filter sub-unit 212 for synchronizing the timestamps of multi-source sensors to the central controller clock and generating an environmental feature matrix.

[0047] Specifically, in the embodiments of the present application, in order to effectively improve the reliability and consistency of environmental data and provide high-quality input for subsequent AI decision-making, the multi-source data fusion unit 210 of the central controller adopts a systematic data purification and fusion process. The specific implementation steps are as follows:

[0048] First, based on the dynamic threshold rule, the original data stream is monitored in real time. For example, if it is monitored that the temperature value of a certain node jumps by more than ±5℃ within 10 seconds, it is automatically determined as a sensor failure or transient interference, and the monitoring data of the spatial adjacent node is immediately used for interpolation replacement, effectively avoiding the pollution of invalid or error data to the system.

[0049] Next, an improved Kalman filter algorithm is used to deeply fuse and eliminate noise of the cleaned multi-source sensor data with spatial redundancy; the algorithm calculates and assigns different confidence weights to the data of adjacent sensors (such as 3 adjacent temperature and humidity probes), and performs weighted average fusion, thereby significantly improving the estimation accuracy and anti-interference ability of key environmental parameters (such as temperature and humidity).

[0050] Finally, the timestamps of all heterogeneous sensors are synchronized to the high-precision clock of the central controller, solving the problem of data asynchrony caused by slight differences in sampling time, and finally generating a normalized environmental feature matrix with a period of 10 minutes, which completely contains the fusion information of six dimensions of temperature, humidity, light, soil moisture, electrical conductivity and stomatal conductance.

[0051] As can be seen, the multi-source data fusion unit 210 builds a complete preprocessing chain from data cleaning, noise suppression to space-time alignment, which fundamentally guarantees the data quality of the input AI model. Not only does it enhance the fault tolerance of the system to sensor failures and transient abnormalities through dynamic threshold and data interpolation mechanism, but also significantly improves the accuracy and stability of the monitoring data through the improved Kalman filter algorithm using data redundancy, while time alignment ensures the consistency of multi-source data on the time axis, laying a solid and reliable data foundation for subsequent spatio-temporal joint AI model for accurate prediction and decision-making, and finally improving the regulation precision and reliability of the whole system.

[0052] As a preferred embodiment of the present application, refer to Figure 5, the AI decision unit 220 comprises: a spatial feature extraction subunit 221, configured to process environmental spatial distribution features in the environmental feature matrix through a three-dimensional convolutional neural network; a time sequence feature extraction subunit 222, configured to process environmental time sequence evolution rules in the environmental feature matrix through a gated recurrent unit; and a feature fusion subunit 223 connected with the spatial feature extraction subunit 221 and the time sequence feature extraction subunit 222, configured to generate an initial predictive regulation strategy by fusing the environmental spatial distribution features and the environmental time sequence evolution rules through a multi-head attention mechanism.

[0053] Specifically, in the embodiment of the present application, the AI decision unit 220 adopts a deep deterministic policy gradient framework as a learning basis, and the environmental feature matrix is mapped and compressed in a high dimension through a three-layer fully connected network to provide a structured representation for subsequent spatio-temporal analysis. The spatio-temporal joint architecture of this unit contains two parallel processing paths, which specifically include: the spatial feature extraction subunit 221 adopts a three-dimensional convolutional neural network to process a 10x10 spatial grid constructed by 20 monitoring points, and extracts three-dimensional spatial distribution features of parameters such as temperature field and humidity field layer by layer through a 3x3x3 convolution kernel, which can effectively identify complex spatial patterns such as persistent low-temperature area in the corner of the greenhouse and high-humidity layer near the ground; the time sequence feature extraction subunit 222 relies on a gated recurrent unit network to analyze historical light intensity and CO2 concentration data for 240 hours, and uses the gating mechanism to accurately capture dynamic patterns such as diurnal light cycle changes and CO2 concentration consumption rules caused by crop photosynthesis.

[0054] Further, the high-dimensional features extracted by the double-path are adaptively weighted and fused by the feature fusion subunit 223 through a multi-head attention mechanism, usually with a weight of 0.6 for spatial features and a weight of 0.4 for time sequence features, to highlight the dominant role of spatial heterogeneity while taking into account the evolution trend, and finally generate a fine device control instruction sequence with a unit of minutes, for example: 08:30 open the east side sunshade curtain to 40%, 09:15 start 1-zone drip irrigation for 120 seconds, and the instruction time resolution reaches 15 seconds.

[0055] The following is a specific embodiment:

[0056] Embodiment 2: Environmental spatio-temporal feature analysis

[0057] 1) Input data preparation stage:

[0058] The historical environmental data for 240 hours (i.e. 10 days) is organized in time sequence into a tensor structure with a dimension of 60x32x6, which contains 60 time points, 32 monitoring nodes and 6 environmental parameters on each node.

[0059] 2) Spatio-temporal joint model running stage:

[0060] The three-dimensional convolutional network in the spatial feature extraction subunit 221 uses a 3×3×3 convolutional kernel system to scan the data cube, effectively identifying the spatial correlation and abnormal regions of environmental parameters. For example, it discovers a significant spatial correlation between the low-temperature area in the northwest corner and the high-humidity area in the southeast corner. The gated recurrent unit in the temporal feature extraction subunit 222 analyzes the data according to the time step, learns and memorizes key time patterns, such as accurately capturing the pattern that the light intensity reaches its peak at 14 pm every day. In the feature fusion subunit 223, the system assigns a weight of 0.6 to spatial features and a weight of 0.4 to temporal features for weighted fusion, and finally outputs a vector containing four key decision values: [target temperature, target humidity, supplemental light intensity, irrigation amount] = [25.3℃, 70%RH, 180μmol / m² / s, 12L / m²]. This vector constitutes the environmental control target based on multi-dimensional prediction for the current period.

[0061] This demonstrates that the AI ​​decision-making unit 220, through parallel processing and deep fusion of spatial and temporal pathways, not only enables the system to perceive the global spatial heterogeneity of the greenhouse environment in real time but also to deeply understand the dynamic evolution of environmental factors, thereby achieving accurate characterization and advanced prediction of environmental conditions. The weighted fusion strategy based on an attention mechanism can dynamically adjust the emphasis of decision-making criteria according to different scenarios, enabling the generated control strategy to possess spatial specificity, temporal foresight, and multi-device collaboration, significantly improving the accuracy, response speed, and resource utilization efficiency of environmental regulation, creating a more stable and suitable growth environment for crops.

[0062] In a preferred embodiment of the present invention, the spatiotemporal joint AI model in the AI ​​decision-making unit 220 is constructed through the model training module 400, which is used to perform the following operations: construct a cross-seasonal training dataset covering multiple complete crop growth cycles, including greenhouse environmental time-series data, outdoor weather station data, crop hyperspectral images, and yield and quality record data; design a multi-task loss function that simultaneously optimizes environmental parameter stability indicators, crop accumulated temperature satisfaction rate, and water resource utilization efficiency; train the spatiotemporal joint AI model using a distributed training framework and complete model convergence within a preset iteration period; and validate the trained spatiotemporal joint AI model using a validation set.

[0063] Specifically, in the embodiment of the present application, the training process of the AI decision unit 220 is based on a cross-season multi-source data set covering the entire growth period of crops for five years, which deeply integrates environmental data such as temperature, humidity, and wind speed provided by outdoor weather stations, high-spectral images of crops collected by drones (vegetation characteristics such as NDVI index are obtained every ten days), and key quality parameters such as fruit sugar content and single fruit weight recorded during the harvesting process.

[0064] Correspondingly, the training process innovatively defines a multi-task loss function that simultaneously optimizes three core objectives, including: an environmental stability term that evaluates the stability of regulation and control by calculating the root mean square error between actual temperature and humidity and the optimal interval for crops; an accumulated temperature satisfaction term that ensures the growth rhythm by evaluating the matching degree between daily effective accumulated temperature (i.e., the cumulative value of periods greater than 10°C) and the physiological needs of crops; and a water efficiency term that optimizes water resource utilization efficiency by counting the water consumption per unit yield.

[0065] Further, at the hardware level, a server equipped with 4 Tesla V100 GPUs is used, and a data parallel distributed training framework is used with a batch size of 256. After sufficient training for 200 iterations, the model converges efficiently and achieves an environmental control accuracy of 94.3% on the validation set, which means that all environmental parameters are accurately maintained within the optimal growth interval for crops for 94.3% of the time.

[0066] This training process integrates cross-season, multi-modal agricultural big data and designs multi-task optimization objectives that are consistent with the actual agricultural production, so that the trained spatio-temporal AI model not only has high environmental prediction and control accuracy, but also deeply understands the complex coupling relationship between crop growth and development and environmental factors. Its ultimate manifestation is the significant improvement in the stability of greenhouse environmental regulation and control, the precise controllability of growth progress, and the efficiency of resource utilization, providing a core intelligent driving force for achieving the production goals of high yield, high quality, and low consumption of crops.

[0067] As a preferred embodiment of the present application, the dynamic optimization engine 230 performs the following operations: establishing a multi-objective function based on a device energy consumption model, a crop growth demand model, and a carbon emission intensity factor; setting the constraint condition that the daily cumulative light intensity of crops is not less than the light saturation point of crops; using a non-dominated sorting genetic algorithm to solve the initial predictive regulation and control strategy to generate a Pareto optimal solution set; and selecting the optimal solution from the Pareto optimal solution set according to the preset preference weight to generate the optimized regulation and control strategy.

[0068] Specifically, in the embodiments of the present application, the dynamic optimization engine 230 combines future 72-hour weather forecasts and current crop growth stage models to output high-precision environmental control strategies. For example, for the growth requirement of maintaining a daily temperature of 23-26°C during the flowering period of tomatoes, the engine can generate zoned temperature control target values with a control accuracy of ±0.5°C, set the upper limit of humidity fluctuation to 85% RH, dynamically adjust the spectral formula of LED light supplementation to optimize the red-to-blue light ratio, and develop an irrigation decision table that precisely allocates water according to crop root depth.

[0069] Further specifically, the dynamic optimization engine 230 takes a multi-objective function as the core, where the device energy consumption item is collected in real time by a smart meter, the carbon emission item is converted according to 0.86 kg CO2 per kWh, and the crop growth priority item is set differently according to key growth periods (such as setting the weight to 0.7 during the fruit setting period). Under the physiological constraint of ensuring that the daily cumulative photosynthetic photon flux is not less than 12 mol / m², the NSGA-II algorithm is used to solve the Pareto optimal solution set, achieving a coordinated reduction in energy consumption and emissions.

[0070] The actual optimization effect shows that the daily average operating time of the ventilation equipment can be reduced from 8 hours to 5.2 hours, with a reduction of 35%; the peak power of the irrigation system is also reduced from 8.5 kW to 6.2 kW, with a reduction of 27%. This not only significantly reduces system operating costs and carbon intensity, but also fully meets the environmental needs of high-quality crop growth, embodying the high integration of intelligent decision-making and green production.

[0071] The following is a specific embodiment:

[0072] Embodiment 3: Multi-objective strategy optimization

[0073] 1) Establish a target function system covering three core optimization objectives:

[0074] Device energy consumption = ventilation power × time + water pump power × flow;

[0075] Growth quality = (actual temperature - target temperature)² + (actual humidity - target humidity)²;

[0076] Carbon emissions = total electricity consumption × 0.86 (kg CO2 / kWh).

[0077] 2) Use the non-dominated sorting genetic algorithm (NSGA-II) for multi-objective optimization solution:

[0078] The algorithm first generates 100 sets of candidate regulation strategies with differences, and the strategy parameters include key variables such as ventilation duration (increased by 0.5 hours in the range of 4-8 hours) and irrigation amount; then, the system calculates the device energy consumption, crop growth quality score and carbon emission score of each set of strategies; then, the algorithm strictly implements the constraint condition, automatically eliminates invalid schemes whose crop growth quality score exceeds the set threshold or device energy consumption is higher than the allowed upper limit, and ensures that all candidate strategies meet the basic needs of crop growth and system operation restrictions.

[0079] 3) Multi-objective Pareto sorting analysis of candidate strategies screened by constraints:

[0080] Based on the performance in the three dimensions of device energy consumption, growth quality and carbon emission, non-dominated solutions that are better in any target and not worse in other targets are identified, forming a Pareto optimal solution set.

[0081] 4) Output optimal solution:

[0082] The system selects the regulation strategy with the best overall performance from the Pareto optimal solution set according to the preset comprehensive evaluation criteria, for example: ventilation duration 5.2 hours, irrigation amount 10.5 L / m².

[0083] The optimization results show that this strategy significantly improves resource utilization efficiency while ensuring environmental regulation accuracy, achieving a balance between energy consumption, emissions and crop growth needs.

[0084] As a preferred embodiment of the present application, with reference to Figure 3 , Figure 6 and Figure 11 , the central controller 200 further comprises a model evolution unit 240, which comprises: an incremental data set construction sub-unit 241 for periodically collecting crop physiological indicators and constructing an incremental data set in combination with historical regulation strategy execution records; a model optimization sub-unit 242 connected to the incremental data set construction sub-unit 241, for fine-tuning and optimizing the spatio-temporal joint AI model based on online transfer learning technology using the incremental data set.

[0085] Specifically, in the embodiment of the present application, the central controller 200 integrates the model evolution unit 240, which significantly improves the long-term adaptability and regulation accuracy of the system through a continuous learning mechanism. The model evolution unit 240 operates regularly, collects key physiological indicators of crops including stem flow rate and chlorophyll fluorescence index every month through high-precision sensors (such as stem flow meters with measurement accuracy of ±0.1 ml / min and chlorophyll fluorometers that can detect photosystem II quantum efficiency), and combines these data with historical environmental regulation strategies and their execution records to construct an incremental data set for model iteration.

[0086] Subsequently, the model optimization subunit 242 uses the incremental data to efficiently fine-tune and optimize the spatio-temporal joint AI decision model based on online transfer learning technology. This process adopts a feature decoupling transfer learning method, and the specific optimization process is as follows: the three-dimensional convolution network weights in the model responsible for extracting spatial features are frozen to retain their existing strong feature extraction capabilities and prior knowledge, and only the last two layers of network parameters of the gated recurrent unit are fine-tuned. The optimization process is guided by the environmental control error and the deviation of the crop physiological index, specifically by fusing 70% of the environmental error and 30% of the physiological index deviation to construct a loss function, and using a small batch gradient descent algorithm with a learning rate of 0.001 for training, when the test set accuracy breaks through 85%, the new model is deployed.

[0087] This operation brings significant benefits: it enables the system to have excellent evolutionary learning and self-adaptive ability. The system can continuously optimize its decision model using the continuously collected crop physiological response data, so as to quickly adjust the strategy when facing the introduction of new varieties or long-term changes in the environment.

[0088] A typical case is that the system only needs to use 72 hours of continuous monitoring data (sampling interval of 5 minutes) for incremental learning to complete the regulation and adaptation of the newly introduced strawberry variety, greatly shortening the two-week adaptation period required by traditional methods to three days, and the initial control accuracy reaches 89%. This not only greatly improves the adaptability of the system to different crops and growth stages, but also realizes the deep and precise coupling between the environmental control strategy and the internal physiological response of the crop, ensuring the forward-looking and scientific nature of the regulation strategy.

[0089] As a preferred embodiment of the present application, with reference to Figure 7 The regulation execution module 300 includes: an instruction compiling unit 310, configured to compile the strategy parameters in the optimized environmental regulation strategy into industrial Internet of Things executable code; a security coding and checking unit 320 connected to the instruction compiling unit 310, configured to add a time stamp and a cyclic redundancy check code to the executable code, and encapsulate it into a control instruction package; wherein the control instruction package contains sunshade curtain opening and closing angle control code, drip irrigation electromagnetic valve opening and closing time sequence code, ventilation unit speed curve code and heating sheet duty cycle parameters; an instruction synchronous issuing unit 330 connected to the security coding and checking unit 320, configured to issue the control instruction package to each execution device through a time stamp synchronization mechanism; and an execution feedback and monitoring unit 340 connected to the instruction synchronous issuing unit 330, configured to receive and analyze the feedback signals of the execution device through a double buffering mechanism, and compare the consistency of the feedback signals and the control instruction package in real time.

[0090] Specifically, in the embodiment of the present application, the regulation execution module 300 has the following specific working process:

[0091] The instruction compiling unit 310 first converts the parameters in the optimization strategy into industrial Internet of Things executable code, for example, the target opening angle of the sunshade curtain is encoded into an 8-bit binary control code, where 01011011 corresponds to 70% opening; at the same time, the on-off timing of the drip irrigation solenoid valve is converted into a hexadecimal instruction conforming to the RS485 communication protocol, such as 0x3A01 indicating that the control of the No. 1 valve is opened for 120 seconds.

[0092] The secure coding and checking unit 320 subsequently adds high-precision timestamps to these instruction data packets and calculates CRC-16 cyclic redundancy check codes, using the polynomial 0x8005, to ensure the integrity and security of the instructions during transmission.

[0093] The instruction synchronization issuing unit 330 uses a timestamp synchronization mechanism to ensure that all control instruction packets are synchronized and issued to the execution devices distributed in each area of the greenhouse within 50 milliseconds, and the instruction content covers complex control commands such as speed curve encoding of the ventilator set and duty cycle parameters of the heating sheet.

[0094] The execution feedback and monitoring unit 340 receives and analyzes the feedback signals of each execution device in real time through a double buffering mechanism, and continuously compares the actual feedback such as valve state and curtain position with the consistency of the control instruction packet.

[0095] The following is a specific embodiment:

[0096] Embodiment 4: Device instruction compiling and issuing

[0097] 1) Instruction conversion stage:

[0098] The control of the sunshade curtain is given in the form of opening percentage, which is linearly mapped by the system into an 8-bit binary control code, with 0% corresponding to 00000000 for full-closed state and 100% corresponding to 11111111 for full-open state; the control instruction of the drip irrigation system is calculated based on the target irrigation amount, according to the formula "opening time (seconds) = (irrigation amount x area) / single valve flow rate", the electromagnetic valve needs to be kept open for 1680 seconds, for example, to achieve an irrigation amount of 10.5 L / m² under the conditions of a total area of 400 square meters and a single valve flow rate of 2.5 L / s; for the ventilation equipment, the system uses a PID control algorithm to convert the target temperature into real-time fan speed adjustment instructions, and the calculation formula is:

[0099] Speed adjustment proportion = Kp x temperature difference + Ki x cumulative temperature difference + Kd x temperature difference rate, where Kp, Ki, and Kd are pre-adjusted empirical parameters, with values of 0.8, 0.05, and 0.1, respectively.

[0100] 2) Instruction safety and issuing stage:

[0101] The system adds a CRC-16 cyclic redundancy check code based on the 0x8005 polynomial to each instruction and encapsulates the instruction data into a data packet in a specific format for broadcast through the industrial Ethernet; wherein the data packet format is as follows:

[0102] [header flag 0xA5][device ID][instruction code][timestamp][CRC check][tail flag 0xAA];

[0103] After receiving the instruction, the execution device replies with an acknowledgement signal within 200 milliseconds, and if the central controller 200 does not receive the acknowledgement within the timeout period, the instruction retransmission mechanism is automatically triggered to ensure reliable delivery and execution of the control instruction.

[0104] The technical solution significantly improves the reliability, synchronization and safety of system execution through multi-level cooperation of compilation, verification, synchronization and feedback monitoring. The beneficial effects are as follows: on the one hand, the use of industrial-level communication protocol and CRC check mechanism effectively prevents instruction transmission errors or interference, ensuring the accuracy of control; on the other hand, the strict 50 millisecond level synchronization delivery mechanism combined with double buffering feedback monitoring significantly reduces the risk of time difference and state inconsistency in multi-device execution, ensuring the real-time and uniformity of greenhouse environment regulation, thereby creating a highly stable and reliable microenvironment for crop growth.

[0105] As a preferred embodiment of the present application, with reference to Figure 1 , Figure 2 and Figure 8 , the edge intelligent gateway module 500 is further included, which is connected to the environment acquisition module 100 and the regulation and execution module 300, and includes: a lightweight anomaly detection unit 510, configured to perform real-time anomaly detection on multi-dimensional environment data based on a lightweight neural network model, and generate an emergency signal when an anomaly is identified; an emergency regulation unit 520 connected to the lightweight anomaly detection unit 510, configured to start a local emergency regulation protocol after receiving the emergency signal, drive the execution device to intervene, and upload an event diagnosis report to the central controller 200.

[0106] Specifically, in the embodiment of the present application, the module is deployed in the greenhouse on-site control cabinet, directly connecting the environment acquisition module 100 and the regulation and control execution module 300, and constitutes an independent edge intelligent closed-loop control node. Its core function is to run the optimized MobileNetV2 neural network model (the model volume is only 8 MB, deployed on the Jetson Nano embedded platform) through the lightweight anomaly detection unit 510 to perform millisecond-level real-time analysis on the incoming multi-dimensional environment data, so as to realize instantaneous perception of abnormal conditions. Once abnormal patterns such as a local area temperature drop of more than 5℃ within 10 minutes (indicating that there may be cold wind penetration), or a humidity difference between adjacent monitoring points continuously exceeding 20%RH are identified, the lightweight anomaly detection unit 510 will immediately generate an emergency signal.

[0107] Correspondingly, after receiving the signal, the emergency regulation and control unit 520 immediately starts the preset local emergency regulation and control protocol. This unit can drive the execution mechanism to respond within 50 milliseconds, for example, quickly closing the ventilation window of the abnormal area and simultaneously starting the standby heater to the preset power (such as 50%), so as to effectively suppress the spread of environmental deterioration and realize rapid physical intervention of the scene.

[0108] At the same time, the module does not operate in isolation, but generates a structured event diagnosis report at the same time of intervention. The report not only contains key abnormal data segments, but also provides preliminary diagnosis suggestions (such as "east area temperature anomaly, code E102" or "sensor may be blocked"), and uploads the report to the central controller and the cloud platform at the same time, so as to realize the whole link, second-level closed-loop response from edge perception, rapid blocking to cloud warning and tracing.

[0109] As can be seen, the edge intelligent gateway module 500 greatly improves the response speed of the system to local sudden abnormalities, shortens the traditional second-level or even minute-level response relying on the central controller 200 to millisecond-level, effectively avoiding the spread of abnormal environment to cause instantaneous damage to crops; in addition, the independent processing capability on the edge side reduces the computing load of the central controller 200, so that it can focus more on global optimization tasks, and at the same time, even in the case of temporary network interruption, the local can still perform key protection actions, greatly enhancing the reliability and robustness of the whole system; in addition, the accurate event report generated by it provides clear fault location and diagnosis basis for the operation and maintenance personnel, simplifies the subsequent maintenance process, and realizes the unification of intelligent early warning, rapid intervention and efficient operation and maintenance.

[0110] As a preferred embodiment of the present application, refer to Figure 1 , Figure 2 and Figure 9The application also comprises an augmented reality interactive terminal 600 connected to the regulation and execution module 300, which comprises a digital twin visualization unit 610 for dynamically rendering environmental parameter cloud maps and virtual device states based on the three-dimensional model of the greenhouse; a gesture recognition and interaction unit 620 connected to the digital twin visualization unit 610 for capturing and recognizing user pre-defined gesture actions, generating artificial correction instructions for controlling the virtual execution device according to pre-set mapping rules; wherein when the deviation of the artificial correction instructions from the optimized regulation strategy exceeds a safety threshold, the augmented reality interactive terminal triggers a multi-expert decision voting mechanism and generates a strategy evaluation report.

[0111] Specifically, in the embodiment of the application, the augmented reality interactive terminal 600 fuses and renders the real-time collected environmental data in the high-precision three-dimensional model of the greenhouse constructed by the Unity3D engine through its digital twin visualization unit 610, generates intuitive environmental parameter cloud maps and virtual device states, and projects them in the form of holographic images into the user's field of view through Hololens2 glasses, realizing immersive monitoring of the overall running state of the greenhouse.

[0112] The user can interact with the virtual environment through natural actions by means of the gesture recognition and interaction unit 620. The system pre-defines intuitive gesture mapping rules, for example, a horizontal sliding action can be used to adjust the opening degree of the virtual sunshade, and is set to correspond to a 5% opening degree change for every 10 centimeters of sliding, and the user can confirm and finally issue the control instruction through a double-click gesture.

[0113] A core beneficial effect of the terminal is embodied in its intelligent safety coordination mechanism. When the deviation between the artificial correction instructions issued by the operator and the AI optimized strategy generated by the central controller exceeds the pre-set safety threshold (for example, the difference between the manually set temperature target and the system recommended value is greater than 3 degrees Celsius), the system will not immediately execute or simply reject the instructions, but will automatically trigger the built-in multi-expert decision voting mechanism, which will call the recommended values of the crop growth model, the suggestions of the embedded planting expert knowledge base and the optimal regulation scheme under similar historical situations in parallel for comprehensive comparison and collaborative decision-making, and automatically generate a detailed strategy evaluation report. This report not only gives decision suggestions, but also provides specific risk warnings, such as "increasing the temperature setting will increase the probability of powdery mildew by 23%", thereby deeply integrating the local experience of the operator with the global optimization of the system and the expert knowledge.

[0114] This design greatly improves the intelligence level of human-computer interaction and the reliability of decision-making. On the one hand, it reduces the operation threshold of complex systems through augmented reality technology, realizing intuitive and convenient visual control. On the other hand, by introducing a safety threshold and a multi-expert voting mechanism, it effectively avoids the systematic risks caused by individual experience misjudgment, combines human subjective initiative with objective rational analysis of data, respects the operator's intention while ensuring the scientificity of control decision-making and the safety of the crop growing environment.

[0115] The following is a specific embodiment:

[0116] Example 5: AR interaction operation process

[0117] When the administrator views the holographic greenhouse model through Hololens, he can interact through specific gestures: pushing the virtual temperature up with the palm (every 10 cm gesture displacement corresponds to an increase of 1℃), and confirming the instruction to the execution device with a fist gesture.

[0118] When the deviation between the manually set temperature and the AI recommended value exceeds 3℃, the system automatically starts the conflict resolution mechanism: first query the knowledge base to get the temperature upper limit of the current tomato flowering period (26℃), then retrieve the optimal temperature control scheme of the same period in the past three years (for example 25.5℃), and finally generate a quantitative risk assessment report (such as "setting 28℃ will cause a 18% decrease in fruit setting rate"), providing data support for the administrator's decision-making.

[0119] As a preferred embodiment of the present application, with reference to Figure 2 , Figure 10 and Figure 11 , it also includes a fault diagnosis and self-recovery module 700 connected to the environment acquisition module 100 and the control execution module 300, which includes: a device health monitoring unit 710 for monitoring the working state parameters of the sensor group and the execution device; an intelligent compensation and redundancy switching unit 720 connected to the device health monitoring unit 710, for automatically starting a data compensation strategy based on multi-source information fusion or a device switching strategy based on hardware redundancy when detecting sensor data anomalies or execution device failures; an operation and maintenance unit 730 connected to the intelligent compensation and redundancy switching unit 720, for automatically generating a device maintenance work order according to the fault device information and pushing it to a remote management terminal.

[0120] Specifically, in the embodiments of the present application, the fault diagnosis and self-recovery module 700 collects the state parameters of the voltage, current, response time, etc. of the sensors and the execution equipment in real time through the equipment health monitoring unit 710, and performs dynamic evaluation based on the preset health model; when the intelligent compensation and redundancy switching unit 720 identifies a fault such as abnormal light sensor data or irrigation valve response timeout, it immediately starts the multi-source data fusion algorithm to calculate the missing parameters or switches to the redundant hardware link to ensure system control continuity; at the same time, the operation and maintenance management unit 730 automatically generates a work order containing the fault location, equipment model and maintenance suggestions, and pushes it to the mobile terminal of the management personnel in real time, thereby realizing the full-process automatic closed-loop management from fault perception, intelligent fault tolerance to maintenance response, and significantly improving the reliability and operation and maintenance efficiency of the system.

[0121] The following is a specific embodiment:

[0122] Embodiment 6: Fault diagnosis and recovery

[0123] 1) When the light sensor fails, automatically start the data compensation process: the system automatically calls the fisheye camera to shoot the sky image, uses image segmentation technology to accurately quantify the cloud coverage, and converts it into a quantitative index of 0 to 100%; combined with historical light data, it dynamically estimates the current photosynthetically active radiation value according to the established "light intensity = sunny baseline value x (1- cloud coverage x 0.8)" model, realizing continuous and reliable perception of light parameters.

[0124] 2) If it is detected that the irrigation valve is stuck, and the water flow sensor response delay is more than 5 seconds (normal threshold < 2 seconds), a quick switching mechanism based on hardware redundancy is immediately triggered: the control relay switches the waterway to the standby pipeline within 100 milliseconds to ensure uninterrupted irrigation operation; at the same time, the operation and maintenance management unit 730 automatically records the fault device ID and detailed information, generates a detailed maintenance work order containing the specific model of the replacement spare parts needed, and pushes an alarm SMS to the management personnel's mobile APP end in real time, such as "B area 3 valve fault, standby line enabled", thereby realizing the full-process intelligent closed-loop processing from fault detection, automatic switching to operation and maintenance notification.

[0125] In addition, for other types of equipment faults, the system also has a perfect intelligent diagnosis and self-recovery mechanism. For example: when the system detects that the indoor carbon dioxide sensor has permanent drift or abnormal reading, the intelligent compensation and redundancy switching unit 720 will immediately start a soft compensation strategy based on multi-source data fusion. This strategy dynamically calculates a more reliable indoor carbon dioxide concentration reference value by analyzing the current ventilation rate, crop photosynthesis intensity (which can be estimated according to light intensity and growth stage), and outdoor carbon dioxide concentration in real time, to maintain the continuous and stable operation of the gas fertilizer regulation function.

[0126] If the key execution equipment such as the circulating fan appears to be stuck or overloaded, the intelligent compensation and redundancy switching unit 720 will complete the diagnosis within seconds and automatically enable the standby fan, while marking the original fan as a fault state. At the same time, the operation and maintenance unit 730 will generate a maintenance work order simultaneously, which not only contains the equipment number and fault type, but also intelligently recommends possible spare parts (such as bearings or capacitors) according to historical maintenance records, and automatically pushes the information to the handheld terminal of the maintenance personnel, significantly shortening the troubleshooting and repair time, and improving the availability and maintenance efficiency of the system.

[0127] As can be seen, this technical solution greatly improves the availability and robustness of the system through real-time health monitoring, intelligent fault compensation and closed-loop management of redundancy switching, effectively avoiding overall environmental imbalance or production interruption caused by local equipment failure; at the same time, the automated operation and maintenance process greatly reduces the dependence on manual inspection, improves the maintenance efficiency, and realizes the intelligentization and precision of fault perception to repair response, providing a solid guarantee for the long-term stable and unmanned operation of the greenhouse.

[0128] As a preferred embodiment of the present application, with reference to Figure 1 and Figure 11 , the system adopts a three-layer architecture, including a perception layer (environment acquisition module 100), a decision layer (central controller 200) and an execution layer (control execution module 300), forming a full-closed-loop control system of "perception-decision-execution-evolution", which realizes a technical breakthrough in greenhouse environment adaptive management. Specifically, in the perception layer, the distributed multi-source sensor network fuses air, soil and crop physiological parameters, and combines Kalman filter to dynamically eliminate environmental noise, providing a high-reliability data foundation for decision-making; in the decision layer, the spatio-temporal joint AI model uses a three-dimensional convolution network to analyze the spatial distribution characteristics of the greenhouse, couples with a gate recurrent unit to capture the environmental time evolution law, and generates a predictive control strategy through an attention mechanism, fundamentally overcoming the response lag defect of traditional threshold triggering; at the same time, through a multi-objective Pareto optimization engine, the device energy consumption, crop demand and carbon emission factor are cooperatively weighed to dynamically reconstruct the initial strategy to achieve optimal resource allocation; in the execution layer, the industrial Internet of Things instruction compiling unit converts the strategy parameters into device executable code, combined with the edge computing node to realize millisecond-level emergency response, ensuring precise synchronization of control actions; in the evolution layer, the online transfer learning mechanism based on incremental data continuously absorbs crop physiological feedback and climate characteristics, enabling the system to have cross-species and cross-season adaptive ability; supplemented by augmented reality human-computer interaction and multi-expert decision voting mechanism, a visual operation channel and risk warning support are provided for manual intervention, ultimately forming a smart greenhouse control paradigm with precise and stable environment control, efficient and intensive resource utilization, and long-term self-evolution.

[0129] The above merely describes preferred embodiments of the present application, and is not intended to limit the implementation and protection scope of the present application. Those skilled in the art should be able to understand that any equivalent substitutions and obvious changes made according to the present application description and drawings should be included in the protection scope of the present application.

Claims

1. An artificial intelligence-based adaptive control system for greenhouse environment, characterized in that, include: The environmental acquisition module is used to acquire multi-dimensional environmental data inside the greenhouse in real time through distributed multi-source sensors; The central controller, connected to the environmental acquisition module, is used to receive the multi-dimensional environmental data and generate optimized control strategies. The control and execution module is connected to the central controller and is used to receive the optimized control strategy and drive the execution device to perform environmental adjustment. The central controller includes: A multi-source data fusion unit is used to filter and fuse the multi-dimensional environmental data to obtain an environmental feature matrix. The AI ​​decision-making unit, connected to the multi-source data fusion unit, is used to extract and analyze the spatiotemporal change characteristics of the environmental feature matrix through a spatiotemporal joint AI model, and generate an initial predictive control strategy. A dynamic optimization engine, connected to the AI ​​decision-making unit, is used to reconstruct and optimize the initial predictive control strategy through a multi-objective optimization algorithm to generate an optimized environmental control strategy. The optimization objectives of the multi-objective optimization algorithm include equipment energy consumption, crop growth requirement satisfaction, and carbon emission intensity. The spatiotemporal joint AI model in the AI ​​decision-making unit is constructed through a model training module, which is used to perform the following operations: Construct a cross-seasonal training dataset covering multiple complete crop growth cycles. The cross-seasonal training dataset includes greenhouse environment time-series data, outdoor weather station data, crop hyperspectral images, and yield and quality record data. Design a multi-task loss function that simultaneously optimizes environmental parameter stability index, crop accumulated temperature satisfaction rate, and water resource utilization efficiency. The spatiotemporal joint AI model is trained using a distributed training framework, and the model convergence is completed within a preset iteration cycle. The trained spatiotemporal joint AI model was validated using a validation set.

2. The adaptive control system for greenhouse environment based on artificial intelligence according to claim 1, characterized in that, The multi-source data fusion unit includes: An outlier filtering subunit is used to detect and remove outliers in the multidimensional environmental data in real time based on preset dynamic threshold rules. A Kalman filter subunit, connected to the outlier filter subunit, is used to fuse redundant data from the multi-source sensors and eliminate environmental noise using an improved Kalman filter algorithm. The timing alignment subunit, connected to the Kalman filter subunit, is used to synchronize the timestamps of the multi-source sensors to the central controller clock and generate the environmental feature matrix.

3. The adaptive control system for greenhouse environment based on artificial intelligence according to claim 1, characterized in that, The AI ​​decision-making unit includes: The spatial feature extraction subunit is used to process the spatial distribution features of the environment in the environmental feature matrix through a three-dimensional convolutional neural network. The temporal feature extraction subunit is used to extract the temporal evolution law of the environment from the environmental feature matrix described by the gated loop unit; The feature fusion subunit, which connects the spatial feature extraction subunit and the temporal feature extraction subunit, is used to generate the initial predictive control strategy by weightedly fusing the spatial distribution features of the environment and the temporal evolution pattern of the environment through a multi-head attention mechanism.

4. The artificial intelligence-based adaptive control system for greenhouse environment according to claim 3, characterized in that, The dynamic optimization engine performs the following operations: A multi-objective function is established based on equipment energy consumption model, crop growth demand model and carbon emission intensity factor; The constraint is that the daily cumulative light intensity of crops should not be lower than the light saturation point of crops. The initial predictive regulation strategy is solved using a non-dominated sorting genetic algorithm to generate a Pareto optimal solution set; Based on preset preference weights, the optimal solution is selected from the Pareto optimal solution set to generate the optimization control strategy.

5. The artificial intelligence-based adaptive control system for greenhouse environment according to claim 1, characterized in that, The central controller further includes a model evolution unit, which comprises: The incremental dataset construction subunit is used to periodically collect crop physiological indicators and combine them with historical control strategy execution records to construct the incremental dataset; The model optimization subunit, which connects to the incremental dataset construction subunit, is used to fine-tune and optimize the spatiotemporal joint AI model based on online transfer learning technology using the incremental dataset.

6. The adaptive control system for greenhouse environment based on artificial intelligence according to claim 1, characterized in that, The control execution module includes: The instruction compilation unit is used to compile the strategy parameters in the optimized environment control strategy into executable code for the Industrial Internet of Things. A security encoding and verification unit, connected to the instruction compilation unit, is used to add timestamps and cyclic redundancy check codes to the executable code and encapsulate it into a control instruction package; The instruction synchronization and distribution unit is connected to the security encoding and verification unit and is used to distribute the control instruction packet to each execution device through a timestamp synchronization mechanism. The execution feedback and monitoring unit is connected to the instruction synchronization and distribution unit. It is used to receive and parse the feedback signal of the execution device through a double buffering mechanism, and compare the consistency of the feedback signal with the control instruction package in real time.

7. The adaptive control system for greenhouse environment based on artificial intelligence according to claim 1, characterized in that, It also includes an edge intelligent gateway module, which connects the environment acquisition module and the control execution module, and includes: A lightweight anomaly detection unit is used to perform real-time anomaly detection on the multidimensional environmental data based on a lightweight neural network model, and generate an emergency signal when an anomaly is detected. An emergency control unit, connected to the lightweight anomaly detection unit, is used to activate a local emergency control protocol upon receiving the emergency signal, drive the execution device to intervene, and upload an event diagnostic report to the central controller.

8. The adaptive control system for greenhouse environment based on artificial intelligence according to claim 1, characterized in that, It also includes an augmented reality interactive terminal, which is connected to the control and execution module, and includes: The digital twin visualization unit is used to dynamically render environmental parameter cloud maps and virtual equipment status based on a 3D greenhouse model; The gesture recognition and interaction unit, connected to the digital twin visualization unit, is used to capture and recognize user-defined gesture actions and generate manual correction instructions for controlling the virtual execution device according to preset mapping rules. When the deviation between the manual correction instruction and the optimized control strategy exceeds a safety threshold, the augmented reality interactive terminal triggers a multi-expert decision-making voting mechanism and generates a strategy evaluation report.

9. The adaptive control system for greenhouse environment based on artificial intelligence according to claim 1, characterized in that, It also includes a fault diagnosis and self-recovery module, which is connected to the environmental acquisition module and the control execution module, and includes: The equipment health monitoring unit is used to monitor the operating status parameters of the sensor group and the actuator. The intelligent compensation and redundancy switching unit is connected to the device health monitoring unit and is used to automatically start a data compensation strategy based on multi-source information fusion or a device switching strategy based on hardware redundancy when abnormal sensor data or device failure is detected. The operation and maintenance management unit, connected to the intelligent compensation and redundancy switching unit, is used to automatically generate equipment maintenance work orders based on faulty equipment information and push them to the remote management terminal.

Images