Greenhouse environment big data collection and regulation method based on edge-cloud cooperation

Abstract

The present application relates to the field of intelligent agricultural technology, and more particularly to a greenhouse environment big data acquisition and regulation method based on edge-cloud cooperation. The present application collects environment, crop physiology, soil moisture and energy consumption parameters by deploying a multi-modal heterogeneous sensor network, performs spatio-temporal alignment, anomaly detection, PCA dimension reduction and improved D-S evidence theory fusion by the edge gateway, generates a low-dimensional comprehensive feature vector, and makes dynamic offloading decisions according to task timeliness, network bandwidth and edge load; real-time emergency tasks call lightweight models on the edge to achieve millisecond-level inference and regulation; short-period tasks are uploaded to the cloud, and a hybrid digital twin model is used to predict environmental trends and generate pre-regulation strategies; long-period tasks aggregate encrypted gradients from multiple greenhouses through federated learning to achieve cross-greenhouse knowledge sharing. The present application realizes high-dimensional data acquisition, elastic cloud-edge cooperation, active prediction and regulation, and cross-greenhouse privacy cooperation, significantly improving the intelligent level of greenhouse environment regulation.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method for collecting and controlling greenhouse environmental big data based on edge-cloud collaboration. Background Technology

[0002] Greenhouse environmental control is a key link in achieving high yield and quality in facility agriculture. Traditional greenhouse management mainly relies on manual experience or automated control systems based on fixed thresholds. Typically, only basic environmental parameters such as temperature, humidity, and light are collected. Simple data aggregation or weight allocation is performed at the edge, and then all data processing and model training are completed in the cloud before control commands are sent to the execution equipment. This "cloud-heavy, edge-light" architecture leads to a single data acquisition dimension and a lack of edge-end quality enhancement. Most existing systems ignore multimodal data such as crop physiological state (e.g., chlorophyll fluorescence, canopy structure), soil moisture (water content, EC value, pH value), and energy consumption, resulting in information blind spots in regulatory decisions. At the same time, raw sensor data has problems such as noise, outliers, and time asynchrony. Uploading all of it to the cloud would cause network congestion and storage burden, while the edge lacks effective data cleaning, dimensionality reduction, and fusion capabilities. Furthermore, existing systems generally adopt a passive "sensing-response" control mode, that is, the regulatory action is only triggered when environmental parameters deviate from the set threshold, lacking the ability to predict future environmental change trends and crop growth needs, resulting in regulatory lag and affecting crop yield and quality.

[0003] Therefore, to address the above problems, this invention proposes a method and system for big data collection and control of greenhouse environment based on edge-cloud collaboration. Through multimodal sensor networks, edge data quality enhancement, dynamic task offloading, digital twin prediction, and federated learning for cross-greenhouse collaboration, a closed-loop intelligent control paradigm of edge-cloud collaboration is formed. Summary of the Invention

[0004] In view of the above problems, this disclosure provides a method for greenhouse environment big data collection and control based on edge-cloud collaboration to overcome or at least partially solve the above problems, aiming to achieve multi-dimensional, high-quality data collection; flexible and efficient cloud-edge collaborative computing; proactive predictive control; privacy-preserving cross-greenhouse knowledge sharing; and multi-objective collaborative optimization of energy consumption, yield, and quality.

[0005] The objective of this invention can be achieved through the following technical solutions: The first aspect of the technical solution of the present invention provides a method for big data collection and control of greenhouse environment based on edge-cloud collaboration, comprising the following steps: S1 deploys a multimodal heterogeneous sensor network in the greenhouse to collect environmental parameters, crop physiological parameters, soil moisture parameters and energy consumption parameters. The edge gateway performs time and spatial alignment on the raw multimodal data to obtain a structured data field after temporal and spatial alignment. The edge gateway described in S2 performs anomaly detection, dimensionality reduction, and multi-source feature fusion on the structured data field to generate a low-dimensional comprehensive feature vector. Based on the timeliness level of the control task, network bandwidth status, and edge computing load, it makes dynamic task offloading decisions to allocate different types of computing tasks to be executed locally at the edge or uploaded to the cloud for execution. S3 For real-time emergency tasks assigned to local execution at the edge, the edge gateway calls a lightweight prediction model to perform real-time inference on the low-dimensional comprehensive feature vector, generates emergency control instructions, and drives the actuator to act. For short-cycle prediction tasks assigned to the cloud, the edge gateway uploads the low-dimensional integrated feature vector to the cloud. The cloud digital twin engine predicts future environmental change trends and crop growth status based on a hybrid model of physical mechanisms and data-driven approaches. When an environmental stress risk is predicted, a pre-regulation strategy is generated and sent to the edge gateway, which then triggers its execution at a specified time. For long-cycle training tasks assigned to the cloud, S5 performs deep model training based on historical feature data and control feedback data uploaded by multiple greenhouses. It also aggregates the encrypted gradients of each greenhouse through a federated learning framework, updates the global shared model, and then distributes the updated global model to the edge gateways of each greenhouse for local personalized fine-tuning. The edge gateway described in S6 collects environmental response data, crop physiological feedback data, and energy consumption feedback data after the execution of the control action, forming a closed-loop feedback data stream. The cloud verifies the prediction error of the digital twin model based on the closed-loop feedback data and corrects the model parameters online.

[0006] Furthermore, in S1, the edge gateway performs time and spatial alignment on the raw multimodal data, including: the time alignment specifically involves using the edge gateway as the master clock of the IEEE 1588 precision time protocol, with each sensor node acting as a slave clock. Link delay is measured and clock offset is compensated by exchanging Sync, Delay_Req, and Delay_Resp messages, ensuring the corrected sensor local time satisfies a synchronization error ≤ 1 microsecond; the spatial alignment specifically involves: the greenhouse plane is gridded according to management units, with grid sizes ranging from 1m×1m to 5m×5m. Each sensor node is assigned a spatial coordinate label (x, y, z). For grid units with deployed sensors, sensor readings are directly aggregated to the corresponding grid. For grid units without deployed sensors, inverse distance weighted interpolation or Kriging interpolation is used to estimate the environmental parameter values ​​at the grid center point, thereby transforming the raw discrete sampling points into a structured spatial data field containing K grid units and C environmental parameters per grid.

[0007] Furthermore, in S2, the anomaly detection employs a sliding window quartile method, specifically: a sliding window of length W is maintained for each sensor channel; the first quartile Q1, the third quartile Q3, and the interquartile range IQR = Q3 - Q1 are calculated within the window; then, the lower bound of anomaly detection Q1 - 1.5 × IQR and the upper bound Q3 + 1.5 × IQR are calculated; if the current sampled value exceeds this range, it is determined to be an anomaly and replaced with the median within the window; the window length W is set according to the parameter change rate, taking W = 60 for environmental parameters and W = 144 for soil parameters; The dimensionality reduction adopts principal component analysis, specifically: Z-score standardization is performed on the cleaned data matrix, the covariance matrix is ​​calculated and eigenvalue decomposition is performed, the first r principal components are selected such that the cumulative variance contribution rate reaches more than 95%, and the original m-dimensional data is projected to an r-dimensional low-dimensional feature space, where r≤30-50; The multi-source feature fusion adopts an improved Dempster evidence theory, specifically: each principal component is used as an evidence source, and the membership degree of the sample belonging to different crop growth states is output by fuzzy support vector machine. The membership degree is normalized to a basic probability assignment, and the reliability weight of the evidence source is introduced. All evidence sources are fused through the weighted Dempster combination rule to obtain a comprehensive basic probability assignment. The state corresponding to the maximum support proposition is used as the determination result of the crop growth state. If the maximum probability assignment is less than 0.6, it is determined to be an uncertain state.

[0008] Furthermore, the dynamic task unloading decision specifically includes: Greenhouse control tasks are categorized into three levels based on timeliness requirements: real-time emergency tasks, short-cycle prediction tasks, and long-cycle training tasks, with corresponding maximum tolerable latency of 100ms, 10s, and 3600s, respectively. Each task has attributes for computational load, input data volume, output data volume, and timeliness level. A multi-objective optimization model is established to minimize the total system latency. Total energy consumption With maximizing computational accuracy The objective function is: ,in The constraints are: real-time emergency tasks are forced to be executed at the edge; long-cycle training tasks are forced to be uploaded to the cloud when the edge computing power utilization exceeds 80%; in each scheduling cycle, the current uplink network bandwidth and edge computing power utilization are read, and the optimal offloading decision vector is solved. ,in This indicates that the task is executed at the edge. This indicates that the data will be uploaded to the cloud for execution.

[0009] Furthermore, in S4, the cloud-based digital twin engine employs a hybrid modeling approach combining a physical mechanism model and a long short-term memory (LSM) network data-driven model. The physical mechanism model is based on the greenhouse heat balance equation, crop photosynthesis equation, and transpiration equation, using measured environmental parameters as boundary conditions, and solves for the physical prediction of the greenhouse state through numerical integration. The LSM network model takes a sequence of historical low-dimensional comprehensive feature vectors as input, memorizes long-term dependencies through a gating mechanism, and outputs a correction term for the prediction error of the physical mechanism model. The hybrid prediction value is expressed as: Twin model prediction value = Physical mechanism model prediction value + Adaptive hybrid weights × LSM network correction term. The adaptive hybrid weights are dynamically adjusted based on the root mean square error of the physical mechanism model and the LSM network model on the most recent 48 hours of historical data. The weight calculation formula is: , This represents the root mean square error of the physical mechanism model. λ represents the root mean square error of the long short-term memory network model. [0,1]; When the twin model predicts that any environmental parameter exceeds the crop's tolerance range within the next 1-4 hours, a pre-regulation strategy is triggered.

[0010] Furthermore, the execution flow of the federated learning framework described in S5 specifically includes: Each greenhouse edge node is based on a local private dataset Perform local stochastic gradient descent training to obtain the local model update gradient. Among them, the local model parameters are: The training rounds are set to E=5, the batch size to B=32 (meaning 32 samples are randomly selected from the dataset each time to calculate the gradient), and the learning rate to η=0.001. L2 norm clipping of gradients, clipping threshold Then add values ​​with a mean of 0 and a variance of 0. Gaussian noise, where the noise scale is... Differential privacy budget =1.0, failure probability δ=10-5, gradient after encryption Uploaded to the cloud, among which It is the identity matrix; The cloud-based federated aggregation module aggregates all encrypted gradients using a data volume weighting method and updates the globally shared model parameters. ; This indicates that the parameters are from the previous round of the global model. The cloud distributes the updated global model to each greenhouse edge gateway. After receiving the global model, each edge gateway combines the local crop variety code, growth stage label and geographical location features, and uses the global model as the initialization parameter to perform local fine-tuning to obtain a personalized model, which is used to replace the old model in the edge real-time inference module.

[0011] Furthermore, the verification and online correction of the prediction error of the digital twin model described in S6 specifically includes: The actual environmental parameter sequence collected after the control action is executed is compared with the predicted value of the twin model at the corresponding time, and the root mean square error (RMSE) and mean absolute error (MAE) of each parameter are calculated. Online model correction is triggered when the predicted temperature RMSE is greater than 1.5℃, the predicted humidity RMSE is greater than 5%RH, the relative error of light intensity is greater than 10%, or the relative error of CO2 concentration is greater than 15%.

[0012] Furthermore, the first level of correction involves adjusting the parameters of the physical mechanism model. Using multiple sets of steady-state operating data, the heat transfer coefficient is calculated by back-calculating the measured heating power, internal and external temperature difference, and enclosure area. With the updated Replace the old values ​​in the mechanistic model; The second level of correction is data-driven incremental fine-tuning of the model. Feature-test paired data from the most recent n hours are collected as the incremental training set. The parameters of the first two layers of the long short-term memory network model are frozen, and only the top layer and the output layer are fine-tuned to obtain the updated correction term output, where n is 168.

[0013] It also includes a step for generating environmental setpoints based on time-segmented multi-objective optimization: Step 1: The cloud-based multi-objective optimization module uses environmental setpoint vectors. Establish an energy consumption objective function as the decision variable. The sum of power required to maintain the set environment for each control device, and the output objective function. Quality objective function ,in and Predicted by a digital twin model; among which, The target temperature (unit: °C) represents the set air temperature that the greenhouse environment needs to maintain. The target relative humidity (unit: %RH) represents the required relative humidity setting for the greenhouse environment. Indicates the target illumination intensity (unit: μ). , which is the photosynthetically active radiation flux density, represents the set value of light intensity that needs to be achieved at the crop canopy; The target carbon dioxide concentration (unit: ppm) represents the set value of carbon dioxide concentration that needs to be maintained inside the greenhouse. Step 2: Use the NSGA-II non-dominated sorting genetic algorithm to solve the Pareto front solution set. Set the population size to 100, the number of generations to 200, the crossover probability to 0.9, the mutation probability to 0.1, and output no more than 20 non-dominated solutions. Step 3: Set different target weights according to the crop growth stage: the target weight for daytime photosynthesis is (energy consumption 0.3, yield 0.5, quality 0.2), the target weight for nighttime respiration is (energy consumption 0.5, yield 0.3, quality 0.2), and the target weight for seasonal transition is (energy consumption 0.2, yield 0.3, quality 0.5). Step 4: Select a set value from the Pareto frontier solution set that matches the weight of the current time period, and send it to the edge gateway as the target parameter of the pre-regulation strategy for execution.

[0014] A second aspect of the technical solution of the present invention provides a greenhouse environment big data acquisition and control system based on edge-cloud collaboration, comprising: The sensing layer consists of a multimodal heterogeneous sensor network deployed in the greenhouse. The sensor network includes environmental parameter sensors, crop physiological sensors, soil moisture sensors, and energy consumption sensors, which are used to collect data on air temperature, humidity, light intensity, CO2 concentration, chlorophyll fluorescence parameters, canopy images, micro-changes in stem diameter, soil moisture content, electrical conductivity, pH value, and the real-time power and cumulative energy consumption of each control device. The edge layer includes one or more edge gateways. Each edge gateway integrates a time alignment module, a spatial alignment module, an anomaly detection module, a principal component analysis dimensionality reduction module, a DS evidence theory fusion module, a task offloading decision module, a lightweight model real-time inference module, and a local control execution module. The edge gateway communicates with the sensors in the perception layer via LoRa, ZigBee, WiFi, or RS485 protocols, and communicates with the cloud layer via HTTP / HTTPS or MQTT over TLS. The cloud layer, deployed on public, private, or hybrid cloud platforms, includes a big data storage and analysis module, a deep model training module, a federated aggregation module, a digital twin engine, and a multi-objective optimization module. It is used to store historical data from multiple greenhouses, perform incremental learning training, aggregate encrypted gradients to update the global model, run hybrid twin models for state prediction, and solve for Pareto optimal environment settings. The execution layer consists of heaters, cooling equipment, supplemental lighting, shading equipment, irrigation systems, fertilization systems, ventilation equipment, and roller shutter equipment. The execution layer receives control commands from the edge gateway through relays, frequency converters, and servo drivers, and executes corresponding physical actions.

[0015] A third aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the greenhouse environment big data collection and control method based on edge-cloud collaboration as described in the first aspect.

[0016] The technical solution proposed in this application can bring the following beneficial effects: 1. The data dimensions of this invention are comprehensive and of high quality. Through four types of heterogeneous sensor networks and spatiotemporal alignment at the edge, anomaly detection, PCA dimensionality reduction, and improved DS evidence theory fusion, the data availability is improved by about 40% to 60%, and the cloud processing burden is reduced by more than 50%.

[0017] 2. This invention features flexible and efficient cloud-edge collaboration. The dynamic task offloading mechanism enables millisecond-level response at the edge for real-time tasks and high-precision calculation in the cloud for non-urgent tasks. It reduces response latency by 60% to 80% and bandwidth usage by approximately 70%.

[0018] 3. This invention enables cross-greenhouse collaboration and ensures privacy and security. It achieves model sharing while protecting data privacy through a federated learning framework, improving training efficiency by 3 to 5 times and generalization accuracy by 15% to 25%.

[0019] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, specific embodiments of this disclosure are described below. Attached Figure Description

[0020] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this disclosure. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a flowchart illustrating the steps of the greenhouse environment big data collection and control method based on edge-cloud collaboration provided in the embodiments of this specification. Figure 2 This is a schematic diagram of the structure of the greenhouse environment big data acquisition and control system based on edge-cloud collaboration provided in the embodiments of this specification. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. The technical solutions provided by various embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0022] like Figure 1 The diagram shows the steps of a method for collecting and controlling greenhouse environmental big data based on edge-cloud collaboration. The first aspect of this invention provides a method for collecting and controlling greenhouse environmental big data based on edge-cloud collaboration. Includes the following steps: S1 deploys a multimodal heterogeneous sensor network in the greenhouse to collect environmental parameters, crop physiological parameters, soil moisture parameters and energy consumption parameters. The edge gateway performs time and spatial alignment on the raw multimodal data to obtain a structured data field after temporal and spatial alignment. Among them, environmental parameter sensors collect environmental parameters. These sensors are deployed at three heights—0.5m, 1.5m, and 3.0m above the ground—as well as at an external weather station in the greenhouse, collecting data on air temperature (°C), relative humidity (%RH), and light intensity (%). Data includes photosynthetically active radiation (PAR), carbon dioxide concentration (ppm), wind speed (m / s), etc. Crop physiological sensors collect crop physiological parameters. These sensors are deployed above and inside the crop canopy to collect chlorophyll fluorescence parameters (including maximum photochemical efficiency and actual photochemical quantum yield). ), canopy RGB-D images, micro-changes in stem diameter, etc. Soil moisture sensors collect soil moisture parameters and are deployed in the form of soil profiles in the main root distribution layers (0-20cm, 20-40cm, and 40-60cm). The sensors collect soil temperature (°C), volumetric water content (%), electrical conductivity EC value (mS / cm), pH value, etc. Energy consumption sensors collect energy consumption parameters. These sensors are installed in the power supply circuits of various control devices, collecting data on heater power (P_heat), refrigeration equipment power (P_cool), supplemental lighting power (P_light), ventilation fan power (P_vent), irrigation pump power (P_irr), as well as cumulative electricity consumption (kWh) and cumulative water consumption (m³). 3 ), cumulative carbon dioxide replenishment (kg).

[0023] Each sensor connects to the edge gateway via LoRa, ZigBee, WiFi, or RS485 protocols; the edge gateway broadcasts a data acquisition command in each scheduling cycle, for example, a scheduling cycle of 1 minute; then the sensor responds and sends back the raw data frame.

[0024] Due to crystal oscillator drift within each sensor, the actual physical time of data collected at the same "nominal time" may deviate by several seconds. Therefore, time alignment is required first. Specifically, the edge gateway is used as the master clock of the IEEE 1588 precision time protocol, and each sensor node is used as a slave clock. The link delay is measured and the clock offset is compensated by exchanging Sync, Delay_Req, and Delay_Resp messages. The corrected local time of the sensor meets the synchronization error ≤ 1 microsecond. All sensor data is re-marked with a unified global timestamp and enters the time-aligned data pool.

[0025] The greenhouse environment exhibits spatial heterogeneity (e.g., higher temperatures on the south side compared to the north, and stronger light intensity in the upper layers compared to the lower layers), making it impossible to simply treat all sensor readings as equal. Therefore, spatial alignment is necessary. Specifically, the greenhouse plane is gridded into management units, with grid sizes ranging from 1m×1m to 5m×5m. Each sensor node is assigned a spatial coordinate label (x, y, z). For grid units with deployed sensors, sensor readings are directly aggregated to the corresponding grid. For grid units without deployed sensors, inverse distance weighted interpolation or Kriging interpolation is used to estimate the environmental parameter values ​​at the grid's center point. This transforms the original discrete sampling points into a structured spatial data field containing K grid units and C environmental parameters per grid. For example, when K=100 and C=6 (temperature, humidity, light intensity, carbon dioxide, leaf area index, soil moisture content), the data dimensionality is reduced from hundreds to 600 dimensions, while still providing clear spatial location information.

[0026] The edge gateway described in S2 performs anomaly detection, dimensionality reduction, and multi-source feature fusion on the structured data field to generate a low-dimensional comprehensive feature vector. Based on the timeliness level of the control task, network bandwidth status, and edge computing load, it makes dynamic task offloading decisions to allocate different types of computing tasks to be executed locally at the edge or uploaded to the cloud for execution. This step is executed at the edge gateway and includes anomaly detection, PCA dimensionality reduction, DS evidence theory fusion, and dynamic task offloading decision. This is one of the key aspects that distinguishes this invention from existing technologies: the edge is no longer a simple S2. The anomaly detection adopts the sliding window quartile method, specifically: a sliding window of length W is maintained for each sensor channel, and the first quartile Q1, the third quartile Q3, and the interquartile range IQR = Q3 - Q1 of the data within the window are calculated. Then, the lower bound of anomaly detection Q1 - 1.5 × IQR and the upper bound Q3 + 1.5 × IQR are calculated. If the current sampled value exceeds this range, it is determined to be an anomaly and replaced with the median within the window. The window length W is set according to the parameter change rate; for environmental parameters, W = 60, and for soil parameters, W = 144. The dimensionality reduction method uses principal component analysis (PCA), which involves: standardizing the cleaned data matrix using Z-score, calculating the covariance matrix and performing eigenvalue decomposition, selecting the first r principal components such that the cumulative variance contribution rate reaches more than 95%, and projecting the original m-dimensional data into an r-dimensional low-dimensional feature space, where r ≤ 30-50. The multi-source feature fusion employs an improved DS evidence theory. Specifically, while principal component analysis (PCA) reduces the dimensionality of the feature vectors, eliminating redundancy, each principal component is an abstract mathematical quantity lacking physical interpretability. To form a comprehensive evaluation of the greenhouse environment and crop status (e.g., "normal," "mild stress," "moderate stress," "severe stress"), this invention uses an improved DS evidence theory for multi-source feature fusion. Each principal component is treated as an evidence source. Fuzzy support vector machine (FSVM) is used to output the membership degree of the sample to different crop growth states. The membership degree is normalized to a basic probability assignment, and the reliability weight of the evidence source is introduced. All evidence sources are fused through the weighted Dempster combination rule to obtain a comprehensive basic probability assignment. The state corresponding to the maximum support proposition is used as the determination result of the crop growth state. If the maximum probability assignment is less than 0.6, it is determined to be an uncertain state. Further explanation: Let's define the recognition framework. Corresponding to four growth states. Each principal component component As a source of evidence, each source of evidence is first classified using a fuzzy support vector machine (FSVM) to obtain the fuzzy membership degree of the sample belonging to each category. Then, the membership degrees are normalized to basic probability assignments (BPA): ,in This indicates uncertainty regarding the source of evidence.

[0027] Introducing a reliability weight for evidence sources (Determined by FSVM classification accuracy), all evidence sources are fused using a weighted Dempster combination rule: ;in The weighted conflict coefficient is used; the largest supported proposition in the fused comprehensive BPA is the determination result of the crop state. If the largest BPA value is lower than the confidence threshold of 0.6, it is determined to be an "uncertain state", and the sampling frequency needs to be increased or the observation window extended before making a decision.

[0028] Furthermore, the dynamic task unloading decision specifically includes: Greenhouse control tasks are divided into three levels according to timeliness requirements: real-time emergency tasks, short-cycle prediction tasks, and long-cycle training tasks, with corresponding maximum tolerable latency of 100ms, 10s, and 3600s, respectively. in, Real-time emergency tasks (L=1): Temperature over-limit alarm, equipment fault detection, emergency sunshade / ventilation, etc., with a maximum tolerable latency of ≤100ms, and are forced to be executed at the edge. Short-cycle forecasting tasks (L=2): forecasting environmental trends in the next 1-4 hours and identifying crop growth stages, with a maximum tolerable latency of ≤10s and flexible unloading capability; Long-cycle training tasks (L=3): global model training, federated aggregation, digital twin parameter optimization, maximum tolerable latency ≤3600s, preferably executed in the cloud; Each task has a computational cost. Input data volume Output data volume Including timeliness level attributes; establishing a multi-objective optimization model to minimize the total system delay. Total energy consumption With maximizing computational accuracy The objective function is: ,in ,in The value is 0.4. The value is 03. The value is 0.3; in, , For uplink bandwidth, For downlink bandwidth, For cloud computing power, For edge computing power; , For transmission power, Power consumption for edge gateways; , and The expected accuracy for execution in the cloud and at the edge, respectively.

[0029] The constraints are as follows: real-time emergency tasks are forced to be executed at the edge; long-cycle training tasks are forced to be uploaded to the cloud when the edge computing power utilization exceeds 80%; in each scheduling cycle, the current uplink network bandwidth and edge computing power utilization are read, and the optimal offloading decision vector is solved. ,in This indicates that the task is executed at the edge. This indicates that the data will be uploaded to the cloud for execution.

[0030] S3 For real-time emergency tasks assigned to local execution at the edge, the edge gateway calls a lightweight prediction model to perform real-time inference on the low-dimensional comprehensive feature vector, generates emergency control instructions, and drives the actuator to act. For tasks identified as real-time emergency tasks (L=1) in S2 and assigned to be executed at the edge, the edge gateway calls the lightweight prediction model for real-time inference. The lightweight model is compressed from the complex model in the cloud through techniques such as knowledge distillation, quantization and pruning. The number of parameters is usually 1 / 10 to 1 / 100 of the cloud model, and the inference latency can be controlled within 100ms.

[0031] The edge real-time inference module comprises three sub-functions: Environmental anomaly detection: The lightweight model receives the current low-dimensional feature vector y and outputs the anomaly probability of each environmental parameter deviating from the appropriate range; if the probability exceeds the threshold (e.g., 0.85), it is judged as an "urgent anomaly"; Short-term trend prediction: Based on the feature sequence of the most recent 5 time steps, a lightweight LSTM or GRU model is used to predict the rate of change of environmental parameters in the next 10-30 minutes; even if the current parameters are not exceeded, if the trend prediction shows that they are about to exceed the limit, intervention can be carried out in advance. Emergency control decision-making: When anomaly detection or trend prediction determines that intervention is necessary, a target setpoint vector is generated. The local control execution module uses a PID control algorithm to convert the deviation between the setpoint and the measured value into a power percentage command for the equipment.

[0032] For short-cycle prediction tasks assigned to the cloud, the edge gateway uploads the low-dimensional integrated feature vector to the cloud. The cloud digital twin engine predicts future environmental change trends and crop growth status based on a hybrid model of physical mechanisms and data-driven approaches. When an environmental stress risk is predicted, a pre-regulation strategy is generated and sent to the edge gateway, which then triggers its execution at a specified time. For short-cycle prediction tasks assigned to the cloud, the edge gateway performs lossy compression (quantization encoding or differential encoding) on ​​the low-dimensional comprehensive feature vector generated by S2 and then uploads it to the cloud; the cloud digital twin engine adopts a hybrid modeling method of physical mechanism and data-driven approach to predict the future state of the greenhouse.

[0033] In S4, the cloud-based digital twin engine employs a hybrid modeling approach combining a physical mechanism model and a long short-term memory (LSM) network data-driven model. The physical mechanism model is based on the greenhouse heat balance equation, crop photosynthesis equation, and transpiration equation, using measured environmental parameters as boundary conditions. It solves for the physical prediction of the greenhouse state through numerical integration. The LSM network model takes a sequence of historical low-dimensional comprehensive feature vectors as input, memorizes long-term dependencies through a gating mechanism, and outputs a correction term for the prediction error of the physical mechanism model. The hybrid prediction value is expressed as: Twin model prediction value = Physical mechanism model prediction value + Adaptive hybrid weights × LSM network correction term. The adaptive hybrid weights are dynamically adjusted based on the root mean square error of the physical mechanism model and the LSM network model on the most recent 48 hours of historical data. The weight calculation formula is as follows: , This represents the root mean square error of the physical mechanism model. λ represents the root mean square error of the long short-term memory network model, where λ∈[0,1]; when the twin model predicts that any environmental parameter exceeds the crop's tolerance range within the next 1-4 hours, a pre-regulation strategy is triggered.

[0034] The twin model generates an environmental state sequence for the next 1-4 hours using a rolling forecasting method. If any environmental parameter is predicted to exceed the crop's tolerance range (e.g., temperature exceeding 35°C), a multi-objective optimization module is triggered; this module uses the environmental setpoint... Using energy consumption as the decision variable and aiming at minimizing energy consumption, maximizing output, and optimizing quality, the NSGA-II algorithm is used to solve the Pareto front solution set. From the solution set, the setpoints that match the weights of the current time period are selected to generate a pre-regulation strategy, including the target setpoint, execution schedule, and device priority. The strategy is then distributed to the edge gateway via HTTPS / MQTT.

[0035] The specific steps are as follows: Step 1: The cloud-based multi-objective optimization module uses environmental setpoint vectors. Establish an energy consumption objective function as the decision variable. The sum of power required to maintain the set environment for each control device, and the output objective function. Quality objective function ,in and Predicted by a digital twin model; among which, The target temperature (unit: °C) represents the set air temperature that the greenhouse environment needs to maintain. The target relative humidity (unit: %RH) represents the required relative humidity setting for the greenhouse environment. Indicates the target illumination intensity (unit: μ). , which is the photosynthetically active radiation flux density, represents the set value of light intensity that needs to be achieved at the crop canopy; The target carbon dioxide concentration (unit: ppm) represents the set value of carbon dioxide concentration that needs to be maintained inside the greenhouse. Step 2: Use the NSGA-II non-dominated sorting genetic algorithm to solve the Pareto front solution set. Set the population size to 100, the number of generations to 200, the crossover probability to 0.9, the mutation probability to 0.1, and output no more than 20 non-dominated solutions. Step 3: Set different target weights according to the crop growth stage: the target weight for daytime photosynthesis is (energy consumption 0.3, yield 0.5, quality 0.2), the target weight for nighttime respiration is (energy consumption 0.5, yield 0.3, quality 0.2), and the target weight for seasonal transition is (energy consumption 0.2, yield 0.3, quality 0.5). Step 4: Select a set value from the Pareto frontier solution set that matches the weight of the current time period, and send it to the edge gateway as the target parameter of the pre-regulation strategy for execution.

[0036] For long-cycle training tasks assigned to the cloud, S5 performs deep model training based on historical feature data and control feedback data uploaded by multiple greenhouses. It also aggregates the encrypted gradients of each greenhouse through a federated learning framework, updates the global shared model, and then distributes the updated global model to the edge gateways of each greenhouse for local personalized fine-tuning. The execution process of the federated learning framework specifically includes: Each greenhouse edge node is based on a local private dataset Perform local stochastic gradient descent training to obtain the local model update gradient. Among them, the local model parameters are: The training rounds are set to E=5, the batch size to B=32 (meaning 32 samples are randomly selected from the dataset each time to calculate the gradient), and the learning rate to η=0.001. L2 norm clipping of gradients, clipping threshold Then add values ​​with a mean of 0 and a variance of 0. Gaussian noise, where the noise scale is... Differential privacy budget =1.0, failure probability δ=10-5, gradient after encryption Uploaded to the cloud, among which It is the identity matrix; The cloud-based federated aggregation module aggregates all encrypted gradients using a data volume weighting method and updates the globally shared model parameters. ; This indicates that the parameters are from the previous round of the global model. The cloud distributes the updated global model to each greenhouse edge gateway. After receiving the global model, each edge gateway combines the local crop variety code, growth stage label and geographical location features, and uses the global model as the initialization parameter to perform local fine-tuning to obtain a personalized model, which is used to replace the old model in the edge real-time inference module.

[0037] The edge gateway described in S6 collects environmental response data, crop physiological feedback data, and energy consumption feedback data after the execution of the control action, forming a closed-loop feedback data stream. The cloud verifies the prediction error of the digital twin model based on the closed-loop feedback data and corrects the model parameters online. The specific steps for verifying and correcting the prediction errors of the digital twin model online include: The actual environmental parameter sequence collected after the control action is executed is compared with the predicted value of the twin model at the corresponding time, and the root mean square error (RMSE) and mean absolute error (MAE) of each parameter are calculated. Online model correction is triggered when the predicted temperature RMSE is greater than 1.5℃, the predicted humidity RMSE is greater than 5%RH, the relative error of light intensity is greater than 10%, or the relative error of CO2 concentration is greater than 15%.

[0038] Furthermore, the first level of correction involves adjusting the parameters of the physical mechanism model. Using multiple sets of steady-state operating data, the heat transfer coefficient is calculated by back-calculating the measured heating power, internal and external temperature difference, and enclosure area. With the updated Replace the old values ​​in the mechanistic model; The second level of correction is data-driven incremental fine-tuning of the model. Feature-test paired data from the most recent n hours are collected as the incremental training set. The parameters of the first two layers of the long short-term memory network model are frozen, and only the top layer and the output layer are fine-tuned to obtain the updated correction term output, where n is 168.

[0039] This embodiment takes the prevention of midday high temperature stress in summer as an example. The complete workflow is as follows: At 10:00 AM, the sensors in the sensing layer collected data on temperature (28℃), humidity (65%), light intensity (80,000 lux), and crop physiological parameters. (Slightly lower than the normal value of 0.78).

[0040] The edge gateway performs time and spatial alignment. The structured data field is processed by anomaly detection, PCA dimensionality reduction, and DS fusion to output a low-dimensional feature vector, and the crop status is determined to be "mild high temperature stress".

[0041] The task unloading decision module determines that "Environmental trend prediction for the next 4 hours" is a short-cycle task (L=2), the current network bandwidth is sufficient (10Mbps), and the edge computing power utilization rate is 45%, so it decides to upload it to the cloud for execution.

[0042] Based on real-time data updates, the cloud-based digital twin engine predicts that the temperature will reach 38℃ at 14:00 in the afternoon (exceeding the tomato's tolerance limit of 35℃), triggering the risk of high-temperature stress.

[0043] The multi-objective optimization module generates a pre-control strategy based on the daytime photosynthesis option weight (yield 0.5): start the wet curtain cooling + sunshade curtain deployment + top ventilation at 12:30 in advance, with the goal of controlling the temperature below 32℃ at 14:00.

[0044] The strategy is sent to the edge gateway, and the edge network executes PID control precisely at 12:30 to drive the execution layer devices.

[0045] At 14:00, the actual temperature was 33℃, and the crop's physiological parameters returned to normal. Feedback data was uploaded to the cloud, and the twin model verification showed a prediction error of 1.0℃ (below the correction threshold), requiring no correction.

[0046] That night, the edge gateway archives and uploads the day's data. At the weekend's federated aggregation cycle, the cloud aggregates the gradient updates from each greenhouse to the global model, which is then distributed to each greenhouse for fine-tuning in the early hours of the following day.

[0047] A second aspect of the technical solution of the present invention provides a greenhouse environment big data acquisition and control system 200 based on edge-cloud collaboration, comprising: The sensing layer 201 consists of a multimodal heterogeneous sensor network deployed in the greenhouse. The sensor network includes environmental parameter sensors, crop physiological sensors, soil moisture sensors, and energy consumption sensors, which are used to collect data on air temperature, humidity, light intensity, CO2 concentration, chlorophyll fluorescence parameters, canopy images, micro-changes in stem diameter, soil moisture content, electrical conductivity, pH value, and the real-time power and cumulative energy consumption of each control device. Edge layer 202 includes one or more edge gateways. Each edge gateway integrates a time alignment module, a spatial alignment module, an anomaly detection module, a principal component analysis dimensionality reduction module, a DS evidence theory fusion module, a task offloading decision module, a lightweight model real-time inference module, and a local control execution module. The edge gateway communicates with the sensors of the perception layer via LoRa, ZigBee, WiFi, or RS485 protocols, and communicates with the cloud layer via HTTP / HTTPS or MQTT over TLS. Layer 203 in the cloud is deployed on public cloud, private cloud or hybrid cloud platforms. It includes a big data storage and analysis module, a deep model training module, a federated aggregation module, a digital twin engine and a multi-objective optimization module. It is used to store historical data in multiple greenhouses, perform incremental learning training, aggregate encrypted gradients to update the global model, run hybrid twin models for state prediction, and solve for Pareto optimal environment settings. The execution layer 204 consists of a heater, a cooling device, a supplementary light, a shading device, an irrigation system, a fertilization system, a ventilation device, and a roller shutter device. The execution layer receives control commands from the edge gateway through relays, frequency converters, and servo drivers, and executes corresponding physical actions.

[0048] A third aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed by one or more processors, cause the processors to perform the greenhouse environment big data collection and control method based on edge-cloud collaboration as described in the first aspect.

[0049] This embodiment can divide the method into functional modules based on the above method example. For example, each function can be assigned to a separate module, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware. It should be noted that the module division in this embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0050] When dividing each function into modules according to its corresponding functions, it should be noted that all relevant content of each step involved in the above method embodiment can be referenced from the functional description of the corresponding functional module, and will not be repeated here.

[0051] This embodiment also provides a computer-readable storage medium (including but not limited to disk storage, CD-ROM, optical storage, etc.) storing computer program code. When the computer program code is run on a computer, the computer executes the above-mentioned related method steps to realize the greenhouse environment big data collection and control method based on edge-cloud collaboration provided in the above embodiment.

[0052] This embodiment also provides a computer program product. When the computer program product is run on a computer, it causes the computer to perform the aforementioned steps to realize the greenhouse environment big data collection and control method based on edge-cloud collaboration provided in the above embodiment. The beneficial effects of the above embodiments can be found in the corresponding methods described above, and will not be repeated here.

[0053] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0054] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. In the description of this disclosure, it should be understood that if terms such as "upper," "lower," "front," "rear," "left," and "right" are used to indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, they are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the indicated position or element must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this disclosure.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0056] The above are merely embodiments of this disclosure and are not intended to limit the scope of this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of the claims of this disclosure.

Claims

1. A method for big data collection and control of greenhouse environment based on edge-cloud collaboration, characterized in that, Includes the following steps: S1 deploys a multimodal heterogeneous sensor network in the greenhouse to collect environmental parameters, crop physiological parameters, soil moisture parameters and energy consumption parameters. The edge gateway performs time and spatial alignment on the raw multimodal data to obtain a structured data field after temporal and spatial alignment. The edge gateway described in S2 performs anomaly detection, dimensionality reduction, and multi-source feature fusion on the structured data field to generate a low-dimensional comprehensive feature vector. Based on the timeliness level of the control task, network bandwidth status, and edge computing load, it makes dynamic task offloading decisions to allocate different types of computing tasks to be executed locally at the edge or uploaded to the cloud for execution. S3 For real-time emergency tasks assigned to local execution at the edge, the edge gateway calls a lightweight prediction model to perform real-time inference on the low-dimensional comprehensive feature vector, generates emergency control instructions, and drives the actuator to act. For short-cycle prediction tasks assigned to the cloud, the edge gateway uploads the low-dimensional integrated feature vector to the cloud. The cloud digital twin engine predicts future environmental change trends and crop growth status based on a hybrid model of physical mechanisms and data-driven approaches. When an environmental stress risk is predicted, a pre-regulation strategy is generated and sent to the edge gateway, which then triggers its execution at a specified time. For long-cycle training tasks assigned to the cloud, S5 performs deep model training based on historical feature data and control feedback data uploaded by multiple greenhouses. It also aggregates the encrypted gradients of each greenhouse through a federated learning framework, updates the global shared model, and then distributes the updated global model to the edge gateways of each greenhouse for local personalized fine-tuning. The edge gateway described in S6 collects environmental response data, crop physiological feedback data, and energy consumption feedback data after the execution of the control action, forming a closed-loop feedback data stream. The cloud verifies the prediction error of the digital twin model based on the closed-loop feedback data and corrects the model parameters online.

2. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 1, characterized in that, In S1, the edge gateway performs time and spatial alignment on the raw multimodal data, including: the time alignment specifically involves using the edge gateway as the master clock of the precision time protocol and each sensor node as a slave clock, measuring link delay and compensating for clock offset by exchanging messages, and ensuring that the corrected local sensor time meets the synchronization error ≤ 1 microsecond; the spatial alignment specifically involves: the greenhouse plane is gridded according to management units, each sensor node is assigned a spatial coordinate label, and for grid units with deployed sensors, sensor readings are directly aggregated to the corresponding grid.

3. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 1, characterized in that, In S2, the anomaly detection adopts the sliding window quartile method, which is as follows: a sliding window of length W is maintained for each sensor channel, the first quartile, the third quartile, and the interquartile range of the data in the window are calculated, and then the lower and upper bounds of the anomaly detection are calculated; if the current sampled value exceeds the interval, it is determined to be an anomaly value and replaced with the median in the window. The dimensionality reduction adopts principal component analysis, specifically: Z-score standardization is performed on the cleaned data matrix, the covariance matrix is ​​calculated and eigenvalue decomposition is performed, the first r principal components are selected so that the cumulative variance contribution rate reaches more than 95%, and the original m-dimensional data is projected to the r-dimensional low-dimensional feature space. The multi-source feature fusion adopts an improved Dempster evidence theory, specifically: each principal component is used as an evidence source, and the membership degree of the sample belonging to different crop growth states is output by fuzzy support vector machine. The membership degree is normalized to a basic probability assignment, and the reliability weight of the evidence source is introduced. All evidence sources are fused through the weighted Dempster combination rule to obtain a comprehensive basic probability assignment. The state corresponding to the maximum support proposition is used as the determination result of the crop growth state. If the maximum probability assignment is less than 0.6, it is determined to be an uncertain state.

4. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 3, characterized in that, The dynamic task unloading decision specifically includes: Greenhouse control tasks are categorized into three levels based on timeliness requirements: real-time emergency tasks, short-cycle prediction tasks, and long-cycle training tasks. Each task has attributes for computational load, input data volume, output data volume, and timeliness level. A multi-objective optimization model is established to minimize the total system latency. Total energy consumption With maximizing computational accuracy The objective function is: ,in .

5. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 1, characterized in that, In S4, the cloud-based digital twin engine employs a hybrid modeling approach combining a physical mechanism model and a long short-term memory (LSTM) network data-driven model. The physical mechanism model is based on the greenhouse heat balance equation, crop photosynthesis equation, and transpiration equation, using measured environmental parameters as boundary conditions, and solves for the physical prediction of the greenhouse state through numerical integration. The LTM network model takes a sequence of historical low-dimensional comprehensive feature vectors as input, memorizes long-term dependencies through a gating mechanism, and outputs a correction term for the prediction error of the physical mechanism model. The hybrid prediction value is expressed as: Twin model prediction value = Physical mechanism model prediction value + Adaptive hybrid weights × LTM network correction term.

6. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 1, characterized in that, The execution flow of the federated learning framework described in S5 specifically includes: Each greenhouse edge node is based on a local private dataset Perform local stochastic gradient descent training to obtain the local model update gradient. ; L2 norm clipping of gradients, clipping threshold Then add values ​​with a mean of 0 and a variance of 0. Gaussian noise, where the noise scale is... Differential privacy budget =1.0, failure probability δ=10-5, gradient after encryption Uploaded to the cloud, among which It is the identity matrix; The cloud-based federated aggregation module aggregates all encrypted gradients using a data volume weighting method and updates the globally shared model parameters. ; The cloud distributes the updated global model to each greenhouse edge gateway. After receiving the global model, each edge gateway combines the local crop variety code, growth stage label and geographical location features, and uses the global model as the initialization parameter to perform local fine-tuning to obtain a personalized model, which is used to replace the old model in the edge real-time inference module.

7. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 1, characterized in that, The verification and online correction of the prediction error of the digital twin model described in S6 specifically includes: The actual environmental parameter sequence collected after the control action is executed is compared with the predicted value of the twin model at the corresponding time, and the root mean square error (RMSE) and mean absolute error (MAE) of each parameter are calculated. Online model correction is triggered when the predicted temperature RMSE is greater than 1.5℃, the predicted humidity RMSE is greater than 5%RH, the relative error of light intensity is greater than 10%, or the relative error of CO2 concentration is greater than 15%.

8. The method for big data collection and control of greenhouse environment based on edge-cloud collaboration according to claim 7, characterized in that, The first level of correction involves adjusting the parameters of the physical mechanism model. This is achieved by using multiple sets of steady-state operating data and back-calculating the heat transfer coefficient based on measured heating power, internal and external temperature difference, and the area of ​​the building envelope. With the updated Replace the old values ​​in the mechanistic model; The second level of correction is data-driven incremental fine-tuning of the model. Feature-test paired data from the most recent n hours are collected as the incremental training set. The parameters of the first two layers of the long short-term memory network model are frozen, and only the top layer and the output layer are fine-tuned to obtain the updated correction term output.

9. A greenhouse environment big data acquisition and control system based on edge-cloud collaboration, characterized in that... The system includes: The sensing layer consists of a multimodal heterogeneous sensor network deployed in the greenhouse, including environmental parameter sensors, crop physiological sensors, soil moisture sensors, and energy consumption sensors. The edge layer includes one or more edge gateways. Each edge gateway integrates a time alignment module, a spatial alignment module, an anomaly detection module, a principal component analysis dimensionality reduction module, a DS evidence theory fusion module, a task offloading decision module, a lightweight model real-time inference module, and a local control and execution module. The cloud layer, deployed on public, private, or hybrid cloud platforms, includes a big data storage and analysis module, a deep model training module, a federated aggregation module, a digital twin engine, and a multi-objective optimization module. It is used to store historical data from multiple greenhouses, perform incremental learning training, aggregate encrypted gradients to update the global model, run hybrid twin models for state prediction, and solve for Pareto optimal environment settings. The execution layer receives control commands from the edge gateway through relays, frequency converters, and servo drives, and executes the corresponding physical actions.

10. A computer-readable storage medium, characterized in that, It stores instructions that, when executed by one or more processors, cause the processors to perform the greenhouse environment big data collection and control method based on edge-cloud collaboration as described in any one of claims 1-8.

Images