Tobacco planting greenhouse microclimate real-time monitoring and regulation method based on cloud platform

Abstract

The invention relates to the technical field of agricultural internet of things and intelligent regulation and control, in particular to a tobacco planting greenhouse microclimate real-time monitoring and regulation and control method based on a cloud platform, which adopts a'cloud-side-end 'three-level architecture, deploys a multi-source sensor network layout in a tobacco planting greenhouse, is connected with terminal equipment and is used for monitoring and controlling the microclimate of the tobacco planting greenhouse. The multi-source sensor network is a three-level sensor network, the data acquisition and processing module adopts LoRaWAN and 4G / 5G dual-mode redundancy transmission, deploys an LSTM prediction model at an edge computing node, inputs time sequence sensor data, adopts a micro-service architecture, adopts a gRPC protocol for inter-service communication, and constructs a virtual tobacco planting greenhouse model through a Unity 3D engine; according to the method, environmental parameters are collected in real time through a multi-source sensor network, a regulation and control strategy is dynamically optimized through a machine learning algorithm, and precise control over the tobacco growth environment is achieved.

Description

Technical Field

[0001] The present invention relates to the field of agricultural Internet of Things and intelligent control technology, and in particular to a cloud platform-based real-time monitoring and control method for the microclimate of a tobacco planting greenhouse. Background Art

[0002] Currently, tobacco greenhouses primarily rely on manual experience or single-point sensors for environmental monitoring, which presents three major technical bottlenecks. First, existing systems often utilize wired sensor networks, which are costly and difficult to dynamically monitor throughout the greenhouse's three-dimensional microclimate parameters (such as temperature gradients, humidity stratification, and CO2 distribution). Second, environmental control methods are limited, and common temperature and humidity control devices lack multivariable coupling models, leading to frequent energy conflicts between heating / ventilation / shading actuators. Third, there is a lack of intelligent decision-making mechanisms based on the crop growth cycle, making existing PID control algorithms difficult to adapt to the diverse diurnal temperature differences and accumulated light requirements of tobacco during its different growth stages (seedling, root extension, and maturity). Statistics show that traditional systems can experience response delays of over 30 minutes in extreme weather conditions, increasing daily energy consumption by 17%-23%.

[0003] While existing agricultural IoT systems have attempted cloud platform integration, they suffer from significant shortcomings in tobacco cultivation scenarios. First, mainstream systems employ fixed threshold alarm mechanisms that fail to account for the nonlinear coupling between light intensity and CO2 concentration. For example, when light intensity exceeds 20,000 lux, failing to simultaneously raise CO2 levels to the 800-1200 ppm threshold will result in photoinhibition. Second, data transmission protocols are poorly standardized, making millisecond-level time alignment difficult for structured and unstructured data generated by heterogeneous sensors (such as thermal imagers, multispectral cameras, and weather stations). Furthermore, cloud-based algorithms generally lack edge computing collaboration, and even with 5G backhaul bandwidth utilization as high as 68%, they still cannot meet the requirements for leaf-scale (<5 cm) microclimate modeling. Experiments have shown that existing systems can reconstruct 3D environmental fields in multi-span greenhouses with errors exceeding ±2.3°C / ±15% RH, severely impacting the precise control of the diurnal temperature difference (8-12°C) required for high-quality tobacco leaf production. Summary of the Invention

[0004] To overcome the significant errors in existing 3D environmental reconstruction techniques, this paper aims to provide a cloud-based method for real-time monitoring and control of the microclimate in tobacco greenhouses. This method combines multiple advanced technologies, using a multi-source sensor network to collect environmental parameters in real time and a machine learning algorithm to dynamically optimize control strategies, achieving precise control of the tobacco growing environment.

[0005] The present invention is implemented through the following technical solution: a method for real-time monitoring and control of the microclimate in tobacco greenhouses based on a cloud platform, characterized in that the method adopts a three-level "cloud-edge-end" architecture and mainly includes the following steps:

[0006] S1) deploying a multi-source sensor network layout in a tobacco greenhouse and connecting it to terminal devices. The multi-source sensor network is a three-level sensor network, including a top-level evenly distributed meteorological sensor array, a middle-level soil sensor network, and a bottom-level wide-angle camera network. The main architecture of the three-level sensor network is a plurality of end-to-end connected architecture lines; the meteorological sensor collects light intensity I t and rainfall R t , the wide-angle camera collects images and generates point cloud data P t ;

[0007] S2) The data acquisition and processing module uses LoRaWAN and 4G / 5G dual-mode redundant transmission. The LoRaWAN adopts Class A working mode. The SX1276 chip supports AES-128 encryption and adaptive data rate. The 4G / 5G is used as a backup channel and seamless switching is achieved through a heartbeat packet mechanism.

[0008] S3) deploying the LSTM prediction model on the edge computing node, and t Extract key feature vegetation height mean m z , vegetation height standard deviation n z , the number of obstacles N obstacle , the input is time series sensor data X t =[T t ,H t ,I t ,R t ] and point cloud data P t The structured vector point cloud data F obtained after feature extraction t =[m z ,n z ,N obstacle ], where T t is the temperature, H t For temperature, the LTSM model adopts a double-layer stacking structure, and uses the LAMB optimizer and sliding window cross-validation during training;

[0009] S4) A microservice architecture is adopted, wherein the microservices include data access service, storage service, computing service and display service. The gRPC protocol is used for communication between services. A virtual tobacco greenhouse model is constructed using the Unity 3D engine, and CAD design drawings are imported to generate a precise structure, which includes span, height, and column position. The association and update in the data access service are realized through the MQTT protocol. Sensor data is mapped to the virtual model area ID, and the mapping error is less than 2 cm.

[0010] Optionally, the data acquisition and processing module includes sensor data acquisition, data preprocessing and data transmission and storage; the sensor data acquisition includes a 5-minute acquisition interval for meteorological sensors and a 10-minute interval for soil sensors, and the camera supports timed shooting and event-triggered shooting modes; the data preprocessing dynamic feature selection subunit uses a sliding window dynamic time warping algorithm to perform time domain alignment on multi-source sensor data, and the alignment error is less than 0.5 seconds; a feature vector for the LSTM model is generated, wherein meteorological data uses the mean within the window, soil data uses differential encoding, and point cloud features use principal component analysis to retain the first k dimensions, and the k value is dynamically adjusted according to the current prediction error: when the LSTM model validation set error is exceeded, k is increased to min(k+2,15), otherwise it remains unchanged; the data transmission compression, E`=E(1-ρ), where ρ is the compression rate, the receiving end uses the same algorithm for restoration and compression, supports multiple sensors mounted on the same bus, adopts a mixed precision storage strategy and control strategy, the original data retains the original precision, and the preprocessed data uses FP16 format.

[0011] Optionally, the control strategy includes a fuzzy logic decision tree, an actuator and energy consumption monitoring; the input parameter InputVector in the fuzzy logic decision tree is the predicted temperature deviation ΔT=T t pred -T t target , humidity deviation The set of light intensity I and CO2 concentration C also includes the key features extracted from the greenhouse model point cloud data, including vegetation height distribution H and obstacle position distribution O, that is, InputVector = [ΔT, ΔH, I, C, H, O]. The priority P of each parameter is calculated based on the input parameters. priority =FuzzyLogicTree(ΔT,ΔH,I,C,H,O), the output is G={A1,A2,...,A n =SortByWeight(P priority ,t), the P priorityIt is a priority queue with an update frequency of t seconds and supports multi-parameter coupled control. The actuator control unit uses the CANopen protocol to send control commands, supports electronic gear ratio and PDO / SDO communication, and is equipped with a watchdog timer with a control accuracy of ±0.1%. Predictive feedforward compensation is introduced to generate a composite control variable, which is the environmental change amount predicted by LSTM for the next three steps. A dual-mode control strategy is configured: when the prediction confidence is high, the feedforward dominant mode is adopted, otherwise it switches to the feedback dominant mode.

[0012] Optionally, in step S4), the sensor data and the virtual model area ID are mapped to form a feedback loop, a credibility index is defined, and the 3D coordinate information in the sensor network layout is used. v =f match (x i ,y i ,z i ; model), model represents the geometric structure of the Unity 3D virtual model. The position of each sensor node is matched with the corresponding area ID in the virtual model, and the matching accuracy reaches the centimeter level.

[0013] Optionally, the data preprocessing uses an isolation forest anomaly detection algorithm, and also includes marking the detected abnormal data and repairing it, E={e i |e i ∈p i (x i ,y i ,z i ),score(e i )>θ}, where θ is the set threshold.

[0014] Optionally, the data transmission adopts Manchester coded time division multiplexing technology, and also includes real-time compression of the transmission data, and the transmission data is compressed for key parameters such as temperature and humidity. Compress data such as light and CO2 mid =a||x t ||2,a∈(0,1), for point cloud feature vectors, principal component analysis is performed and then compressed. The receiving end decompresses the data.

[0015] Optionally, the fuzzy logic decision tree in the control strategy also includes a multi-parameter coupling control rule, which is formulated according to different combinations of predicted temperature deviation, humidity deviation, light intensity and CO2 concentration. k , R k :(ΔT∈A)∧(ΔH∈B)∧(I∈C), where the trigger weight of each rule is w k , w k=f(ΔT,ΔH,I,C,m,O).

[0016] Optionally, the virtual model region ID mapping further includes region division and optimization: according to tobacco growth stage, environmental requirements or physical layout.

[0017] Optionally, the actuator is controlled by CANopen protocol, and different electronic gear ratios r are configured according to the actuator type and control requirements. g .

[0018] Optionally, the abnormal pattern recognition also includes an early warning mechanism, which is bound to the confidence of the prediction model and defines three levels of response: (1) When and , only log is recorded; (2) When or , redundant sensor verification is started; (3) When and , the full system safety mode is triggered; when an abnormal environmental pattern is identified, the early warning mechanism is automatically triggered and the control strategy is adjusted.

[0019] The beneficial effects of the present invention are as follows: First, through the integrated application of a cloud platform, a multi-source sensor network, and an LSTM prediction model, the present invention significantly improves the efficiency and accuracy of microclimate control in tobacco greenhouses. The cloud platform enables real-time data collection, transmission, processing, and prediction. Combined with the LSTM model's accurate prediction of future environmental trends, this enables control strategies to respond to environmental changes in advance, providing more stable and suitable environmental conditions for tobacco growth. This precise control not only helps improve the yield and quality of tobacco crops, but also effectively reduces production costs and energy consumption by optimizing resource utilization, such as precise control of irrigation, ventilation, and lighting.

[0020] The data-driven control method of the present invention not only enhances the flexibility and real-time performance of environmental control, but also detects and resolves potential environmental problems in advance through anomaly detection and early warning mechanisms, thereby further improving the safety and sustainability of tobacco cultivation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below with reference to the accompanying drawings and examples.

[0022] Figure 1 This is a cloud-edge architecture design diagram of a cloud-based platform-based tobacco greenhouse microclimate real-time monitoring and control method of the present invention;

[0023] Figure 2 This is a schematic diagram of a layered soil sensor for a cloud platform-based tobacco greenhouse microclimate real-time monitoring and control method of the present invention;

[0024] Figure 3 This is a data collection schematic diagram of a cloud platform-based tobacco greenhouse microclimate real-time monitoring and control method of the present invention. DETAILED DESCRIPTION

[0025] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. In addition, features described with respect to some examples may also be combined in other examples.

[0026] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in one or more embodiments of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprising" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, but do not exclude other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0027] To overcome the significant errors in existing 3D environmental reconstruction techniques, this paper aims to provide a cloud-based method for real-time monitoring and control of the microclimate in tobacco greenhouses. This method combines multiple advanced technologies, using a multi-source sensor network to collect environmental parameters in real time and a machine learning algorithm to dynamically optimize control strategies, achieving precise control of the tobacco growing environment.

[0028] The present invention aims to address technical bottlenecks such as blind spots in monitoring, delayed decision-making, and imbalanced energy efficiency in the field of tobacco planting environment regulation, and proposes a method for intelligent microclimate regulation based on a "cloud-edge-end" collaborative architecture. Its core inventive concept presents a three-dimensional technological evolution: vertical deepening from global perception to precise decision-making, horizontal expansion from a single data source to multimodal fusion, and dynamic optimization from open-loop control to virtual-reality closed-loop. The present invention provides a real-time monitoring and regulation method for the microclimate of tobacco planting greenhouses based on a cloud platform, characterized in that the method adopts a three-level "cloud-edge-end" architecture, and mainly includes the following steps:

[0029] S1) deploying a multi-source sensor network layout in a tobacco greenhouse and connecting it to terminal devices. The multi-source sensor network is a three-level sensor network, including a top-level evenly distributed meteorological sensor array, a middle-level soil sensor network, and a bottom-level wide-angle camera network. The main architecture of the three-level sensor network is a plurality of end-to-end connected architecture lines; the meteorological sensor collects light intensity I t and rainfall R t , the wide-angle camera collects images and generates point cloud data P t ;Using a layered sensor network to achieve full coverage monitoring of environmental parameters, point cloud data provides the basis for three-dimensional spatial analysis;

[0030] S2) Dual-mode transmission ensures data transmission reliability. LoRaWAN's low power consumption is suitable for dense sensor nodes, and 4G / 5G ensures the real-time delivery of critical data. The data acquisition and processing module uses LoRaWAN and 4G / 5G dual-mode redundant transmission. The LoRaWAN adopts Class A operating mode. The SX1276 chip supports AES-128 encryption and adaptive data rate. The 4G / 5G is used as a backup channel, and seamless switching is achieved through a heartbeat packet mechanism.

[0031] S3) deploying the LSTM prediction model on the edge computing node, and t Extract key feature vegetation height mean m z , vegetation height standard deviation n z , the number of obstacles N obstacle , the input is time series sensor data X t =[T t ,H t ,I t ,R t ] and point cloud data P t The structured vector point cloud data F obtained after feature extraction t =[m z ,n z ,N obstacle ], where T t is the temperature, H t For temperature, the LTSM model adopts a double-layer stacking structure, and uses the LAMB optimizer and sliding window cross-validation during training;

[0032] S4) A microservice architecture is adopted, wherein the microservices include data access service, storage service, computing service and display service. The gRPC protocol is used for communication between services. A virtual tobacco greenhouse model is constructed using the Unity 3D engine, and CAD design drawings are imported to generate a precise structure, which includes span, height, and column position. The association and update in the data access service are realized through the MQTT protocol. Sensor data is mapped to the virtual model area ID, and the mapping error is less than 2 cm.

[0033] Because environmental parameters within tobacco greenhouses exhibit significant temporal heterogeneity, different types of sensors have varying response speeds and sensitivities to environmental changes. A unified sampling cycle would lead to resource waste or information loss, so differentiated acquisition strategies need to be set based on the characteristics of the physical process. Preferably, the data acquisition and processing module includes sensor data acquisition, data preprocessing, and data transmission and storage. The sensor data acquisition includes a 5-minute acquisition interval for meteorological sensors and a 10-minute interval for soil sensors. The camera supports both timed and event-triggered capture modes. Differentiated acquisition intervals balance data accuracy with storage costs, while the event-triggered mode accurately captures sudden environmental changes and reduces ineffective data transmission. This technical solution addresses the issues of "single acquisition frequency, severe data redundancy, and delayed response to sudden events" in traditional monitoring systems, improving the system's ability to perceive sudden environmental changes. Meteorological parameters (such as temperature and humidity) change rapidly, so a high-frequency acquisition of 5 minutes is used; soil parameters are relatively stable, so a low-frequency acquisition of 10 minutes is used. In addition to timed capture, the camera also triggers capture by setting thresholds (such as sudden changes in temperature and humidity), achieving event-driven acquisition.

[0034] Multi-source sensor data suffers from temporal misalignment. Directly inputting it into deep learning models like LSTM introduces noise and affects training effectiveness, thus requiring temporal alignment and structured feature extraction. This technical solution addresses the difficulties of multi-source heterogeneous data fusion, inconsistent model input, and feature redundancy, improving the accuracy and stability of the prediction model. The data preprocessing dynamic feature selection subunit employs a sliding window dynamic time warping algorithm to temporally align multi-source sensor data, achieving an alignment error of less than 0.5 seconds. It generates feature vectors for the LSTM model, using the window mean for meteorological data, differential encoding for soil data, and principal component analysis for point cloud features, retaining the first k dimensions. The k value is dynamically adjusted based on the current prediction error: when the LSTM model validation set has an error, k is increased to min(k+2,15); otherwise, it remains unchanged. A sliding window DTW algorithm is used to align data at different sampling rates. Different feature engineering methods are designed for different data types (meteorological data averaging, soil data differential encoding, and point cloud PCA dimensionality reduction), and the PCA dimension k is dynamically adjusted through a prediction error feedback mechanism. This achieves high-precision time domain synchronization and feature modeling, effectively supports subsequent prediction and control decisions, and improves the intelligence level of the system.

[0035] In agricultural environments, edge devices often face the problems of limited storage space and bandwidth, and resource utilization must be optimized while ensuring data availability. This technical solution solves the problems of "redundant data transmission, high storage pressure, and waste of communication resources" in traditional systems, and achieves lightweight deployment. The data transmission compression, E`=E(1-ρ), where ρ is the compression ratio, uses the same algorithm for restoration and compression at the receiving end, supports multiple sensors mounted on the same bus, adopts mixed precision storage strategy and control strategy, retains the original precision of raw data, and uses FP16 format for preprocessed data. Mixed precision storage reduces storage overhead, lossless compression of key parameters ensures control accuracy, and bus reuse reduces hardware deployment costs.

[0036] The environment in tobacco greenhouses is affected by multiple factors. Single variable control is difficult to meet the needs of crop growth. An intelligent decision-making mechanism that can make comprehensive judgments and dynamic adjustments is needed. As an optimal strategy, the control strategy includes a fuzzy logic decision tree, an executive mechanism, and energy consumption monitoring. The input parameter InputVector in the fuzzy logic decision tree is the predicted temperature deviation ΔT = T t pred -T t target , humidity deviation A fuzzy logic decision tree combines light intensity I and CO2 concentration C to achieve multi-parameter coupled control. It dynamically adjusts priorities based on spatial characteristics, improving the refinement of environmental control. This technical solution addresses the problems of "parameter isolation, slow response, and extensive control" in traditional greenhouse control systems, achieving coordinated multi-parameter regulation. Fuzzy logic is used to establish a rule base, determine the current environmental state based on the combination of input parameters, output the priority of each actuator, and dynamically adjust the control sequence based on spatial distribution characteristics. This improves the intelligence and response speed of regulation, achieving more refined environmental control and improving tobacco yield and quality.

[0037] Environmental control not only needs to consider physical parameters, but also needs to combine the actual growth status of crops and spatial layout to achieve personalized and regionalized precise control. Therefore, it also includes the key features extracted from the greenhouse model point cloud data, including vegetation height distribution H and obstacle position distribution O, that is, InputVector = [ΔT, ΔH, I, C, H, O], and the priority P of each parameter is calculated based on the input parameters. priority =FuzzyLogicTree(ΔT,ΔH,I,C,H,O), the output is G={A1,A2,...,A n =SortByWeight(P priority ,t), the P priority The system creates a priority queue with an update frequency of t seconds, supporting multi-parameter coupled control. Spatial features such as vegetation height and obstacle distribution are extracted from point cloud data and combined with environmental parameters to generate a priority queue. This queue is updated regularly to adapt to environmental changes. This implements a multi-parameter control mechanism driven by spatial perception, improving output per unit area and resource utilization.

[0038] Agricultural actuators are diverse, requiring unified communication standards and high-precision control interfaces while balancing remote management and system stability. The actuator control unit uses the CANopen protocol to send control commands, supports electronic gear ratios and PDO / SDO communication, and features a watchdog timer with a control accuracy of ±0.1%. Predictive feedforward compensation is introduced to generate a composite control variable, representing the environmental change predicted three steps into the future by an LSTM. A dual-mode control strategy is implemented: feedforward dominates when prediction confidence is high, switching to feedback dominate otherwise. Energy consumption monitoring optimizes system efficiency, and the MQTT protocol enables lightweight data exchange, supporting cross-platform device access and remote management. The CANopen protocol supports high-speed, reliable communication; the electronic gear ratio adapts to different actuator transmission characteristics; a watchdog function prevents system crashes; feedforward and feedback dual-mode control improves responsiveness; and MQTT is used for remote communication. This creates a standardized, modular, and high-precision actuator control system that supports remote monitoring and automated operation and maintenance.

[0039] Preferably, in the step S4), if the virtual model cannot accurately correspond to the real world, it will cause control distortion, and visualization and precise partition control need to be achieved through precise mapping. The sensor data and the virtual model area ID are mapped to form a feedback loop, and a credibility index is defined. The feedback loop realizes dynamic calibration of the virtual and real models. The credibility index quantifies the prediction reliability. The centimeter-level mapping accuracy supports precise control decisions, which solves the problem of disconnection between the virtual model and the real environment and mismatch of the control area. Using the three-dimensional coordinate information in the sensor network layout, ID v =f match (x i ,y i ,z i ; model), model represents the geometric structure of the Unity 3D virtual model. The position of each sensor node is matched with the corresponding area ID in the virtual model with a matching accuracy of centimeters, realizing high-precision virtual-reality mapping and dynamic calibration, supporting personalized control by area and stage, and improving the system visualization and ease of operation.

[0040] As an example, the data preprocessing adopts the isolation forest anomaly detection algorithm, and also includes marking the detected abnormal data, anomaly detection to ensure data quality, automatic repair to reduce manual intervention, setting a threshold mechanism to adapt to the needs of different growth stages; and repair, E = {e i |e i ∈p i (x i ,y i ,z i ),score(e i )>θ}, where θ is the set threshold. As an option, the data transmission adopts Manchester coding time division multiplexing technology, Manchester coding improves anti-interference ability, hierarchical compression optimizes bandwidth utilization, and principal component analysis reduces the point cloud transmission load; it also includes real-time compression of the transmitted data, and the transmission data is compressed for key parameters such as temperature and humidity. Compress data such as light and CO2 mid =a||x t ||2,a∈(0,1), for point cloud feature vectors, principal component analysis is performed and then compressed. Decompression is performed at the receiving end. This solves the problems of traditional systems, such as difficulty identifying abnormal data, frequent manual intervention, and slow response. Outliers are detected using the Isolation Forest algorithm; they are automatically marked and then interpolated or removed. A dynamic threshold mechanism is set to adapt to the different tolerance levels during the seedling, growth, and maturity stages.

[0041] Preferably, the fuzzy logic decision tree in the control strategy also includes a multi-parameter coupling control rule, which is formulated according to different combinations of predicted temperature deviation, humidity deviation, light intensity and CO2 concentration. k , R k :(ΔT∈A)∧(ΔH∈B)∧(I∈C), where the trigger weight of each rule is w k , w k =f(ΔT, ΔH, I, C, m, O). Multi-parameter coupled control rules dynamically adjust control strategies based on actual environmental conditions, improving the precision and responsiveness of greenhouse environmental control. Multiple sets of rules are formulated based on environmental parameter combinations, and the optimal control strategy is dynamically selected through fuzzy logic, enhancing the system's adaptability. This improves control response speed and stability, enhancing the system's environmental adaptability and control precision.

[0042] Preferably, the virtual model region ID mapping also includes region division and optimization, dynamically dividing the virtual model region based on tobacco growth stage, environmental requirements, or physical layout to optimize the spatial resolution of the control strategy. Dynamic region division allows for flexible adjustment of the control strategy based on the actual growth status of the crop and environmental requirements, thereby improving resource utilization efficiency and crop yield.

[0043] As an example, the actuator is controlled by CANopen protocol, and different electronic gear ratios r are configured according to the actuator type and control requirements. g , Optimize the communication protocols for process data objects and service data objects. The CANopen protocol supports efficient communication mechanisms. By configuring different electronic gear ratios and optimizing the communication protocol, the response speed and accuracy of the actuator can be improved.

[0044] Preferably, the abnormal warning mechanism is bound to the confidence level of the prediction model and defines a three-level response: (1) When ≥ ...

[0045] In one embodiment of the present invention, the multi-source sensor network layout adopts a three-dimensional architecture. The evenly distributed meteorological sensor array on the top layer consists of 16 high-precision temperature and humidity sensors, 4 light intensity sensors, and 2 CO2 concentration sensors, forming a 4×4 grid layout (see Figure 1The middle soil sensor network adopts a layered layout scheme, deploying 12 soil temperature and humidity sensors and 4 nutrient sensors at the depth of 0-15cm and 15-30cm respectively to ensure accurate monitoring of the root layer environment (see Figure 2 The underlying wide-angle camera network is equipped with four cameras that support night vision and use a fisheye lens design to achieve comprehensive coverage. This layout uses LoRaWAN and 4G / 5G dual-mode transmission. Meteorological data is uploaded every 5 minutes, and soil data is updated every 10 minutes. The cameras support event-triggered shooting (such as automatic shooting when temperature and humidity suddenly exceed thresholds).

[0046] In another embodiment of the present invention, an LSTM prediction model is deeply integrated with the control strategy. The two-layer LSTM model deployed on the edge computing node uses sliding window cross-validation, with an input dimension of 6-hour time series data (72 time steps and 4 feature dimensions). The LAMB optimizer is used for training, and the learning rate is dynamically adjusted between 0.001 and 0.01. The prediction results are transmitted to the control strategy module via the gRPC protocol. A fuzzy logic decision tree generates an actuator priority queue based on the predicted temperature deviation (ΔT), humidity deviation (ΔH), light intensity (L), and CO2 concentration (C). When ΔT>2°C or ΔH>10%, the sprinkler system and ventilation equipment are automatically triggered; when L<2000 lux, the supplemental lighting system is prioritized. The actuators are controlled using the CANopen protocol, and the electronic gear ratio is dynamically adjusted according to the control requirements (for example, the ventilation equipment gear ratio is set to 1:8, with a control accuracy of ±0.1%). The energy consumption monitoring module collects device current data via the Modbus RTU protocol and uploads it to the cloud using the MQTT protocol, enabling energy consumption trend analysis and abnormal warnings.

[0047] The above describes the embodiments of this embodiment, but this embodiment is not limited to the above specific implementation methods. The above specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for real-time monitoring and control of tobacco greenhouse microclimate based on a cloud platform, characterized in that: This approach adopts a three-level "cloud-edge-end" architecture and mainly includes the following steps: S1) deploying a multi-source sensor network layout in a tobacco greenhouse and connecting it to terminal devices. The multi-source sensor network is a three-level sensor network, including a top-level evenly distributed meteorological sensor array, a middle-level soil sensor network, and a bottom-level wide-angle camera network. The main architecture of the three-level sensor network is a plurality of end-to-end connected architecture lines; the meteorological sensor collects light intensity I t and rainfall R t , the wide-angle camera collects images and generates point cloud data P t ; S2) The data acquisition and processing module uses LoRaWAN and 4G / 5G dual-mode redundant transmission. The LoRaWAN adopts Class A working mode. The SX1276 chip supports AES-128 encryption and adaptive data rate. The 4G / 5G is used as a backup channel and seamless switching is achieved through a heartbeat packet mechanism. S3) deploying the LSTM prediction model on the edge computing node, and t Extract key feature vegetation height mean m z , vegetation height standard deviation n z , the number of obstacles N obstacle , the input is time series sensor data X t =[T t ,H t ,I t ,R t ] and point cloud data P t The structured vector point cloud data F obtained after feature extraction t =[m z ,n z ,N obstacle ], where T t is the temperature, H t For temperature, the LTSM model adopts a double-layer stacking structure, and uses the LAMB optimizer and sliding window cross-validation during training; S4) A microservice architecture is adopted, wherein the microservices include data access service, storage service, computing service and display service. The gRPC protocol is used for communication between services. A virtual tobacco greenhouse model is constructed using the Unity 3D engine, and CAD design drawings are imported to generate a precise structure, which includes span, height, and column position. The association and update in the data access service are realized through the MQTT protocol. Sensor data is mapped to the virtual model area ID, and the mapping error is less than 2 cm.

2. A method for real-time monitoring and control of microclimate in tobacco greenhouses based on a cloud platform according to claim 1, characterized in that: The data acquisition and processing module includes sensor data acquisition, data preprocessing, and data transmission and storage. The sensor data acquisition includes a 5-minute acquisition interval for meteorological sensors and a 10-minute interval for soil sensors. The camera supports timed shooting and event-triggered shooting modes. The data preprocessing dynamic feature selection subunit uses a sliding window dynamic time warping algorithm to perform time domain alignment on multi-source sensor data, with an alignment error of less than 0.5 seconds. Generate feature vectors for the LSTM model, where the meteorological data uses the mean within the window, the soil data uses differential encoding, and the point cloud features use principal component analysis to retain the first k dimensions. The k value is dynamically adjusted according to the current prediction error: when the LSTM model verification set error occurs, k is increased to min(k+2,15), otherwise it remains unchanged; the data transmission compression, E`=E(1-ρ), where ρ is the compression rate, the receiving end uses the same algorithm for restoration and compression, supports multiple sensors mounted on the same bus, adopts mixed precision storage strategy and control strategy, the original data retains the original precision, and the preprocessed data uses FP16 format.

3. A method for real-time monitoring and control of tobacco greenhouse microclimate based on a cloud platform according to claim 2, characterized in that: The control strategy includes a fuzzy logic decision tree, an actuator and energy consumption monitoring; the input parameter InputVector in the fuzzy logic decision tree is the predicted temperature deviation ΔT=T t pred -T t target , humidity deviation The set of light intensity I and CO2 concentration C also includes the key features extracted from the greenhouse model point cloud data, including vegetation height distribution H and obstacle position distribution O, that is, InputVector = [ΔT, ΔH, I, C, H, O]. The priority P of each parameter is calculated based on the input parameters. priority =FuzzyLogicTree(ΔT,ΔH,I,C,H,O), the output is G={A1,A2,...,A n =SortByWeight(P priority ,t), the P priority It is a priority queue with an update frequency of t seconds and supports multi-parameter coupled control. The actuator control unit uses the CANopen protocol to send control commands, supports electronic gear ratio and PDO / SDO communication, and is equipped with a watchdog timer with a control accuracy of ±0.1%. Predictive feedforward compensation is introduced to generate a composite control variable, which is the environmental change amount predicted by LSTM for the next three steps. A dual-mode control strategy is configured: when the prediction confidence is high, the feedforward dominant mode is adopted, otherwise it switches to the feedback dominant mode.

4. The method for real-time monitoring and control of microclimate in tobacco greenhouses based on a cloud platform according to claim 2 is characterized in that: In the step S4), the sensor data and the virtual model area ID are mapped to form a feedback loop, and the credibility index is defined. The three-dimensional coordinate information in the sensor network layout is used to determine the ID. v =f match (x i ,y i ,z i ; model), model represents the geometric structure of the Unity 3D virtual model. The position of each sensor node is matched with the corresponding area ID in the virtual model, and the matching accuracy reaches the centimeter level.

5. The method for real-time monitoring and control of microclimate in tobacco greenhouses based on a cloud platform according to claim 2 is characterized in that: The data preprocessing adopts the isolation forest anomaly detection algorithm, and also includes marking the detected abnormal data and repairing it, E={e i |e i ∈p i (x i ,y i ,z i ),score(e i )>θ}, where θ is the set threshold.

6. The method for real-time monitoring and control of microclimate in tobacco greenhouses based on a cloud platform according to claim 2, characterized in that: The data transmission adopts Manchester coded time division multiplexing technology, and also includes real-time compression of the transmission data. The transmission data is compressed for key parameters such as temperature and humidity. Compress data such as light and CO2 mid =a||x t ||2,a∈(0,1), for point cloud feature vectors, principal component analysis is performed and then compressed. The receiving end decompresses the data.

7. The method for real-time monitoring and control of tobacco greenhouse microclimate based on a cloud platform according to claim 3 is characterized in that: The fuzzy logic decision tree in the control strategy also includes multi-parameter coupling control rules. According to different combinations of predicted temperature deviation, humidity deviation, light intensity and CO2 concentration, multi-parameter coupling control rules R are formulated. k , R k :(ΔT∈A)∧(ΔH∈B)∧(I∈C), where the trigger weight of each rule is w k , w k =f(ΔT,ΔH,I,C,m,O).

8. The method for real-time monitoring and control of tobacco greenhouse microclimate based on a cloud platform according to claim 4, characterized in that: The virtual model area ID mapping also includes area division and optimization: according to tobacco growth stage, environmental requirements or physical layout.

9. The method for real-time monitoring and control of tobacco greenhouse microclimate based on a cloud platform according to claim 5, characterized in that: The actuator is controlled by CANopen protocol and also includes configuring different electronic gear ratios r according to the actuator type and control requirements. g .

10. The method for real-time monitoring and control of tobacco greenhouse microclimate based on a cloud platform according to claim 6, characterized in that: The abnormal pattern recognition also includes an early warning mechanism, which is bound to the confidence of the prediction model and defines three levels of response: (1) When and , only log is recorded; (2) When or , redundant sensor verification is started; (3) When and , the whole system safety mode is triggered; when an abnormal environmental pattern is identified, the early warning mechanism is automatically triggered and the control strategy is adjusted.

Images