Greenhouse remote monitoring system for smart agriculture

Abstract

The invention relates to the technical field of environment regulation and control, and discloses a greenhouse remote monitoring system for intelligent agriculture, which comprises an environment sensing and executing unit, a field intelligent control gateway and a cloud management platform, a crop growth stage recognition module in the field intelligent control gateway adopts a Bayesian adaptive multi-target convolutional neural network model to realize high-precision dynamic recognition of a crop phenological period, and an environment disturbance prediction module performs prospective probability prediction on environment disturbances such as temperature, humidity and illumination based on a Fourier attention-state space model. The collaborative optimization decision module executes adaptive dual-objective optimization according to the crop growth stage, and realizes dynamic balance between the crop growth suitability and the system energy consumption, so as to construct a sensing, prediction, decision, execution and feedback coupled full-flow closed-loop intelligent regulation and control mechanism; and the accuracy, the real-time performance and the energy efficiency level of agricultural greenhouse environment management are improved.

Description

Technical Field

[0001] This invention relates to the field of environmental control technology, and in particular to a remote monitoring system for greenhouses used in smart agriculture. Background Technology

[0002] Existing agricultural greenhouse monitoring systems mostly employ a sensor network and centralized processing architecture with a single communication method. Communication delays can easily lead to sluggish control responses and an inability to cope with rapid fluctuations in environmental parameters. Their control strategies are often based on fixed thresholds, failing to consider the stage-specific dynamic needs of crop growth and the coupling relationships between multiple environmental factors, resulting in extensive control and high energy consumption. Furthermore, the systems lack the ability to predict and proactively mitigate abnormal disturbances, relying on manual intervention and exhibiting insufficient intelligence. In addition, the lack of collaborative decision-making among the various on-site actuators leads to frequent action conflicts, further reducing resource utilization efficiency and overall control effectiveness.

[0003] Therefore, it is necessary to design a greenhouse remote monitoring system that integrates edge intelligent decision-making and multi-actuator collaborative scheduling to achieve environmental feedforward-feedback collaborative control and active disturbance suppression based on crop growth models, thereby improving the accuracy, real-time performance and energy efficiency of greenhouse production management. Summary of the Invention

[0004] This invention proposes a remote monitoring system for greenhouses in smart agriculture. To address the problems of slow response, coarse strategies, poor anti-disturbance ability, and inefficient fixed recognition models in existing systems, a collaborative architecture of edge intelligent decision-making and cloud model optimization is constructed to realize the real-time generation and precise execution of control commands.

[0005] This invention provides a remote monitoring system for greenhouses in smart agriculture. The system includes: an environmental sensing and execution unit, a field intelligent control gateway, a cloud management platform, and an actuator controller.

[0006] The environmental sensing and execution unit is deployed inside the greenhouse and in the crop canopy, integrating multiple sensors and control devices to acquire multi-source sensor signals.

[0007] The on-site intelligent control gateway is electrically connected to the environmental sensing and execution unit. It is used to process multi-source sensor signals, acquire real-time environmental status data, and generate the optimal control command set. The gateway is equipped with multiple industrial interfaces and connects to the actuator controller via a wireless local area network. It integrates an environmental perception and data fusion module, a crop growth stage identification module, an environmental disturbance prediction module, a collaborative optimization decision-making module, and a multi-actuator scheduling module.

[0008] Environmental perception and data fusion module: Establish a unified spatiotemporal benchmark, perform time synchronization, spatial interpolation and confidence weighted fusion processing on multi-source sensor signals, and generate a highly reliable global environmental status map;

[0009] Crop growth stage identification module: This module has a built-in preset crop phenological feature library. Based on the time series data in the global environmental status map, it takes crop canopy images at regular intervals through industrial cameras, analyzes the crop canopy images taken at regular intervals using a Bayesian adaptive multi-objective CNN model, and matches them with the crop phenological feature library to dynamically identify the current growth stage of the crop and output the crop growth stage identifier and its confidence level.

[0010] Environmental disturbance prediction module: It takes the global environmental status map, crop growth stage identification and short-term weather forecast data in the cloud management platform as input, and uses the Fourier attention-state space prediction model to make probabilistic predictions of possible abnormal environmental disturbances in the next 1-24 hours, and outputs a probabilistic environmental disturbance prediction report.

[0011] Collaborative optimization decision-making module: Based on the global environmental status map, crop growth stage identification, and probabilistic environmental disturbance prediction report, and calling the crop growth model parameters issued by the cloud management platform, it calculates the corresponding theoretical regulation amount based on the adaptive dual-objective optimization algorithm of growth stage, with crop growth suitability and total system regulation energy consumption as joint optimization objectives.

[0012] Multi-actuator scheduling module: Taking the theoretical control quantity as the demand input, and based on a hybrid scheduling rule base that combines the urgency of disturbances, crop sensitivity and resource constraints, it adopts a hybrid scheduling strategy that combines the earliest disturbance deadline and the minimum resource slack to perform conflict detection, resolution and sequencing planning on the action sequence of the actuator controller, generate the optimal control instruction set, and send it to each actuator controller.

[0013] The cloud management platform, deployed on a remote server, communicates bidirectionally with the on-site intelligent control gateway via a cellular network. The platform integrates third-party meteorological data service interfaces to obtain short-term weather forecast data, receives global environmental status maps and crop growth stage identifiers, and iteratively optimizes the staged crop growth model and updates the global strategy library based on historical big data. The optimized data is then transmitted back to the on-site intelligent control gateway, while providing users with a visual monitoring interface, early warning information, and data analysis reports.

[0014] Furthermore, the crop growth stage identification module employs a Bayesian adaptive multi-objective CNN model to process crop canopy images and global environmental state maps, dynamically identifying and outputting crop growth stage identifiers. This process specifically includes the following steps:

[0015] Step S1: Fuse the time-series data of the global environmental state map with the crop canopy image to construct a multimodal tensor;

[0016] Step S2: Use a Bayesian adaptive multi-objective CNN model to extract the global deep semantic feature vector and local phenotypic feature vector of the multimodal tensor, and use adaptive ROI technology to extract the local phenological change feature vector in the local phenotypic feature vector;

[0017] The construction method of a Bayesian adaptive multi-objective CNN model includes the following steps:

[0018] Step T1: Select a high-precision CNN model with structural alignment and a high-efficiency CNN model as the models to be fused;

[0019] Step T2: For the network layer sequence in the model to be fused, extract the parameter difference representation of the network layer sequence, and based on the parameter difference representation, model the model block problem of the network layer sequence as a one-dimensional ordered block optimization problem. Introduce the parameter difference homogeneity constraint within the module and the information balance constraint between modules. Use dynamic programming algorithm to perform global optimal clustering partitioning of the network layer sequence and output the functional module partitioning scheme.

[0020] Step T3: Preset functional module fusion weight vectors. During the forward propagation process, for each functional module in the functional module partitioning scheme, calculate the feature map output of the model to be fused at the corresponding position aligned with the functional module structure, and perform weighted fusion according to the functional module fusion weight vectors to obtain the fused output of the corresponding functional module, and construct the fused convolutional neural network model; perform forward inference on the fused convolutional neural network model to obtain the model performance evaluation results corresponding to the functional module fusion weight vectors;

[0021] Step T4: Construct a Gaussian process surrogate model based on the fusion weight vector of the functional modules and the corresponding model performance evaluation results;

[0022] Step T5: Based on the Gaussian process surrogate model, the expected hypervolume improved acquisition function is used to sample and update the fusion weight vector of the functional modules, and candidate fusion weight configurations are generated iteratively. The model performance corresponding to the candidate fusion weight configurations is evaluated based on the fusion convolutional neural network model until the preset convergence condition is met, and a Pareto front is formed.

[0023] Step T6: Select the fusion weight configuration from the Pareto front and load the fusion weight configuration into the convolutional neural network model to obtain the Bayesian adaptive multi-objective CNN model.

[0024] Step S3: Design a dynamic feature weighting mechanism based on Pareto perception to assign differentiated weights to the local feature vectors of phenological changes, and obtain the weighted local feature vectors of phenological changes;

[0025] Step S4: Aggregate the weighted local feature vector of phenological changes with the global deep semantic feature vector into an image feature vector, encode the time series data into an environmental feature vector, and fuse the image feature vector and the environmental feature vector to form the current fused feature;

[0026] Step S5: Pre-cache historical fusion feature sequences and introduce them into the Bi-GRU network. Input the current fusion features and historical fusion feature sequences into the Bi-GRU network together, and output the preliminary phenological classification results and their preliminary confidence scores after time smoothing.

[0027] Step S6: Construct a standard template for the crop phenological period feature library, calculate the cosine similarity between the image feature vector and the standard template for the crop phenological period feature library, calibrate the preliminary confidence level to determine the final confidence level, and output the crop growth stage identifier.

[0028] Furthermore, the process of predicting environmental disturbances using the Fourier attention-state space prediction model specifically includes the following steps:

[0029] Step B1: Perform a fast Fourier transform on the time series data in the global environmental state map, and use the Fourier attention mechanism to extract the high-frequency residual components of the time series data;

[0030] Step B2: Construct a linear state-space model, the core of which is the system state transition rule;

[0031] Step B3: Use a random perturbation injection layer to couple high-frequency residual components and short-term weather forecast data to a linear state-space model to obtain external information. Through the perturbation coupling matrix, the external information is transformed into a correction term for the system state transition rule.

[0032] Step B4: Dynamically adjust the contribution ratio of the deterministic evolution of the linear state-space model and the random perturbation injection layer through a gating switching mechanism;

[0033] Step B5: During the training phase, a joint loss function is used to jointly optimize the attention head weights, linear state space model parameters, and perturbation coupling matrix of the Fourier mechanism to train the Fourier attention-state space prediction model.

[0034] Step B6: In the inference phase, the trained Fourier attention-state space prediction model is subjected to multi-step forward iteration and Monte Carlo sampling to generate a probabilistic environmental disturbance prediction report.

[0035] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0036] This invention integrates a Bayesian adaptive multi-objective CNN model into an on-site intelligent control gateway, achieving high-precision and high-efficiency dynamic identification of crop growth stages. It also introduces a Pareto-aware dynamic feature weighting mechanism, which autonomously optimizes feature extraction paths and fusion strategies based on image content complexity and the real-time computing power of the edge gateway during inference. Under the same computing resource constraints, this solution significantly improves the discriminative ability and recognition confidence of canopy image phenological feature extraction, laying a reliable perceptual foundation for subsequent precise environmental control based on growth stages, and achieving the optimal balance between edge-side accuracy and inference speed.

[0037] This invention achieves precise and coordinated regulation of crop growth requirements and system energy consumption by constructing an adaptive dual-objective optimization algorithm based on growth stages. The mechanism uses real-time identified crop growth stage markers and probabilistic environmental disturbance predictions as dynamic constraints, calls a cloud-optimized staged crop growth model, and performs Pareto optimization with the joint objectives of maximizing comprehensive environmental suitability and minimizing the system's estimated energy consumption. This method upgrades the regulation strategy from static threshold-driven to crop physiological dynamic demand-driven, solving the problem that traditional control ignores the staged physiological changes and multi-factor coupling relationships of crops. While ensuring that crops are in the optimal growth microenvironment, it significantly reduces unnecessary energy consumption caused by blind regulation.

[0038] The multi-actuator scheduling module and its hybrid scheduling strategy of this invention solve the problems of resource competition and action conflict in multi-actuator collaborative regulation, and improve the overall execution efficiency of the system. The module takes the theoretical regulation quantity output by the collaborative optimization decision module as the demand input, and comprehensively considers the urgency of the disturbance event, the sensitivity difference of crops at different growth stages, and actual resource constraints such as electricity. It adopts a strategy that combines the earliest disturbance deadline and the minimum resource slack for intelligent planning. Through this mechanism, the system can always prioritize the scheduling and execution of the most urgent and critical regulation tasks that have the most impact on crop growth under the condition of limited resources. This avoids the cancellation of regulation effects or resource waste caused by conflicting actuator actions, and ensures the reliable, orderly, and efficient execution of complex regulation commands. Attached Figure Description

[0039] Figure 1 This is a system structure block diagram of a greenhouse remote monitoring system for smart agriculture proposed in this invention;

[0040] Figure 2 This is a schematic diagram of the Fourier attention-state space prediction model structure proposed in Embodiment 3 of the present invention. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] Example 1

[0043] according to Figure 1 This invention provides a remote monitoring system for greenhouses in smart agriculture, applicable to multi-span greenhouses, solar greenhouses, and plastic greenhouses. The system includes:

[0044] The environmental sensing and actuation unit, deployed within the greenhouse space and crop canopy, integrates multiple sensors and control devices. Specifically, it includes: a wall-mounted air temperature and humidity sensor (model SHT30, measurement range -40~125℃, humidity 0~100%RH, accuracy ±0.2℃ / ±2% RH), a canopy light sensor (model PAR-100, measurement range 0~2000 μmol / m² / s, accuracy ±5%), an infrared CO2 sensor (model MH-Z19C, measurement range 0~5000 ppm, accuracy ±50 ppm), an insertion-type soil three-parameter sensor (model TDR-300, measuring soil temperature, humidity, and conductivity, temperature accuracy ±0.5℃, humidity accuracy ±1%), and an infrared leaf surface temperature sensor (model MLX90614, measurement range -70~380℃, accuracy ±0.5℃), used for comprehensive acquisition of multi-source sensor signals. Control devices include an electric film rolling machine (stroke 5 m, load 50 kg, response time ≤0.5). s), internal circulation fan (power 40 W, wind speed adjustable from 0.3 to 1.2 m / s), evaporative cooling pad-fan system (evaporative cooling pad thickness 15 cm, fan airflow 8000 m³ / h), LED supplementary lighting array (spectral density 400~700 nm, power 300 W, photosynthetic photon flux density ≥1200 μmol / m 2 / s), a canned CO2 release device (release rate adjustable from 0 to 5 L / min), executes control commands;

[0045] The on-site intelligent control gateway is electrically connected to the environmental sensing and execution units. It processes multi-source sensor signals, acquires real-time environmental status data, and generates the optimal control command set. The gateway adopts an industrial-grade design with an IP65 protection rating, adapting to the high temperature and humidity environment of greenhouses. It integrates an STM32H743ZIT6 main control chip and a Siemens S7-1200 PLC core, equipped with 2 RS-485 bus interfaces (supporting Modbus-RTU protocol), 16 digital input / output (DI / DO) channels, and 8 analog input / output (AI / AO) channels. It connects to the controllers of each actuator via a Zigbee 3.0 wireless LAN (communication distance ≤100 m, transmission rate 250 kbps). The gateway integrates multiple functional modules, including an environmental perception and data fusion module, a crop growth stage identification module, an environmental disturbance prediction module, a collaborative optimization decision-making module, and a multi-actuator scheduling module. All modules achieve data interaction through an Ethernet switch, with a response latency ≤100 ms.

[0046] The cloud-based management platform, deployed on an Alibaba Cloud ECS server (configured with 4 cores, 8GB memory, and 500GB cloud disk), communicates bidirectionally with the on-site intelligent control gateway via a 5G DTU (model USR-G806, supporting SA / NSA dual-mode), with a communication latency of ≤50 ms. The platform integrates third-party meteorological data service interfaces to obtain 72-hour weather forecast data (updated hourly, including temperature, humidity, wind speed, and sunshine duration parameters). It receives global environmental status maps, crop growth stage identifiers, and equipment operation logs uploaded by the on-site intelligent control gateway. Based on the TensorFlow framework, it constructs a staged crop growth model, iteratively optimizes the model based on historical big data, and updates the global strategy library. The optimized crop growth model parameters, strategy rules, and short-term weather forecast data are then transmitted back to the on-site intelligent control gateway. Web-based management functionality is also provided.

[0047] Example 2

[0048] This embodiment is based on Embodiment 1. In this embodiment, the modules in the field intelligent control gateway work together through the following process to achieve real-time generation and precise execution of control commands:

[0049] Environmental perception and data fusion module: Based on the high-precision DS3231 clock (error ±2ppm) inside the on-site intelligent control gateway, it adds precise timestamps (accuracy to milliseconds) to multi-source sensor signals; it uses the Kriging interpolation algorithm to perform gridded reconstruction of spatially sparse point data (grid resolution 0.5 m × 0.5 m) to generate a 10 m × 10 m environmental field grid; it calculates the confidence weight of each sensor data by combining the sensor calibration error model (based on factory calibration data), and fuses multi-source data within the same grid by weighted averaging, finally generating a global environmental state map including temperature field, humidity field, light distribution map, and CO2 concentration distribution, with an update frequency of 5 seconds / time;

[0050] Crop growth stage identification module: A pre-built image feature library of typical phenological stages for four common economic crops—tomato, cucumber, strawberry, and lettuce—constitutes the crop phenological stage feature library, containing standard image features of 12 key growth stages for each crop from sowing to harvest. Crop canopy images are captured every 30 minutes using three wide-angle cameras (1920×1080 resolution, 30 fps) deployed on the greenhouse roof. Feature extraction is performed using a Bayesian adaptive multi-objective CNN model. The core of this model is a segmented feature extraction backbone, consisting of sequentially connected shallow edge blocks (3 convolutional layers, focusing on texture), mid-level semantic blocks (MobileNetV3 improved inverse residual module, focusing on organ morphology), and deep abstract blocks (channel attention module, focusing on global context). Adaptive cross-block attention gating is embedded between blocks, allowing gradients and information to be optimized across blocks during training and dynamically activated based on the output of the previous module during inference. The key channel of the module, the feature weighting mechanism, is implemented by a miniature multilayer perceptron (hypernetwork). It takes the feature variance of the current image and the real-time CPU load rate of the gateway as input, and outputs the soft weights of the contribution to each feature block. It realizes the adaptive operation point selection on the accuracy-efficiency Pareto front. It combines the accumulated temperature data of the past 7 days and performs cosine similarity matching with the crop phenological feature library. It dynamically judges and outputs the most likely crop growth stage identifier (tomato - first inflorescence flowering stage) and its preliminary probability. The confidence threshold is set to 0.7. When it is lower than the threshold, the image re-acquisition mechanism is triggered.

[0051] The environmental disturbance prediction module takes as input time-series data (last 24 hours) of the current global environmental status map, crop growth stage indicators, and short-term weather forecast data (temperature, humidity, wind speed, sunshine duration) for the next 72 hours obtained from the cloud management platform. It uses a Fourier attention-state space prediction model for rolling prediction (prediction step size 1 hour / time). The output includes the probability, expected start time, and intensity range of preset disturbance events within the next 24 hours. Preset disturbance events include: temperatures below 8℃ for ≥2 hours, and sunshine duration below 200 μmol / m².2 / s and lasting for ≥4 hours, humidity above 95% RH and lasting for ≥3 hours, form a probabilistic environmental disturbance prediction report;

[0052] Collaborative optimization decision-making module: Using the overall environmental status map, crop growth stage identifiers, and probabilistic environmental disturbance prediction reports as dynamic constraints and prediction conditions for the optimization problem, it calls upon the crop growth model distributed from the cloud management platform (this model defines the quantitative relationship between temperature, humidity, light, CO2 concentration, and crop growth suitability at the current growth stage); it constructs a comprehensive environmental suitability index (ESI), with a value range of 0 to 1 (1 being optimal); based on an adaptive bi-objective optimization algorithm for growth stages, it uses the maximization of ESI and the minimization of estimated energy consumption (EC) as joint optimization objectives, and solves the Pareto front using a constrained non-dominated sorting genetic algorithm (NSGA-Ⅲ); it selects the optimal solution according to the user's preset preference for yield preservation (ESI weight 0.7, EC weight 0.3), and outputs the target setpoints for each environmental factor and the required theoretical control amounts (60% power for the wet curtain-fan system and 80% wind speed for the internal circulation fan).

[0053] Multi-actuator scheduling module: Taking the theoretical control quantity as the demand input, and based on the preset conflict rule table (the wet curtain-fan system is prohibited from starting when the electric film roller is open) and resource constraint table (maximum power supply of 5 kW), it adopts a hybrid scheduling strategy that combines the earliest disturbance deadline and the minimum resource slack for action planning; it identifies action conflicts between actuators through a conflict detection algorithm, and uses a priority ranking method, with disturbance urgency > crop sensitivity > energy consumption cost, to resolve conflicts; it generates an executable optimal control instruction set, which is distributed to each actuator controller through a Zigbee network, with instruction execution feedback time ≤300 ms.

[0054] The crop growth stage identification module uses a Bayesian adaptive multi-objective CNN model to process crop canopy images, dynamically identify and output crop growth stage identifiers. The process specifically includes the following steps:

[0055] Step S1: Input crop canopy image and time series data of global environmental status map; embed the time series data into the coding layer, standardize the crop canopy image, copy the time series data after embedding the coding layer in the spatial dimension, and fuse it with the image pixels to form a multimodal tensor;

[0056] Step S2: Using a Bayesian adaptive multi-objective CNN model, extract the global deep semantic feature vector and local phenotypic feature vector of the multimodal tensor, and use adaptive ROI technology to focus on extracting local phenotype feature vectors of climate change from the local phenotypic feature vectors;

[0057] Specifically, the construction method of the Bayesian adaptive multi-objective CNN model includes the following steps:

[0058] Step T1: Selection of Fusion Objects: Select a high-precision CNN model and a high-efficiency CNN model with structural alignment as the models to be fused, wherein the two CNN models have structural heterogeneity in terms of network depth, channel configuration or parameter distribution;

[0059] Step T2: Functional Module Block Construction Step: For the network layer sequences that are structurally aligned with each other in the model to be fused, extract the parameter difference representation of the network layer sequences. Based on the parameter difference representation, model the block problem of the network layer sequence is modeled as a one-dimensional ordered block optimization problem. Introduce the homogeneity constraint of parameter difference within the module and the information balance constraint between modules. Use dynamic programming algorithm to perform global optimal clustering partitioning of the network layer sequences and output the functional module partitioning scheme so that the network layers within the same functional module are consistent in terms of parameter difference magnitude and functional characteristics, and the information contribution between different functional modules is relatively balanced.

[0060] Step T4: Module-level fusion modeling step: A preset functional module fusion weight vector is used. During the forward propagation of the model, for each functional module in the functional module partitioning scheme, the feature maps output by the high-precision CNN model and the high-efficiency CNN model at the corresponding positions aligned with the functional module structure are calculated. These feature maps are then weighted and fused according to the functional module fusion weight vector to obtain the fused output of the corresponding functional module, thereby constructing a fused convolutional neural network model. Based on this, the target dataset is input into the fused convolutional neural network model, forward inference is performed, and statistical analysis is conducted on the model output results to obtain the model performance evaluation results corresponding to the functional module fusion weight vector. Within the same functional module, a set of fusion weights is shared. These fusion weights are used to linearly weight and sum the feature maps output by the high-precision CNN model and the high-efficiency CNN model in that functional module. The fusion weight vector serves as the decision variable to be optimized.

[0061] Step T5: Multi-objective surrogate model construction steps: Based on the functional module fusion weight vector and the corresponding model performance evaluation results, construct Gaussian process surrogate models for each optimization objective to approximately represent the mapping relationship between the functional module-level fusion weight vector and the corresponding model performance index;

[0062] Step T6: Bayesian Multi-Objective Optimization Step: Based on the Gaussian process surrogate model, the expected hypervolume improvement acquisition function is used to sample and update the functional module fusion weight vector to drive the Bayesian multi-objective optimization process; in the functional module-level fusion weight space, candidate fusion weight configurations are iteratively generated, and the model performance corresponding to each candidate fusion weight configuration is evaluated based on the fusion convolutional neural network model until the preset convergence condition is met, obtaining a set of Pareto optimal fusion weight configurations and their corresponding model performance evaluation results, thereby forming the Pareto front between model accuracy and inference efficiency; the expected hypervolume improvement acquisition function is the q-expected hypervolume improvement acquisition function;

[0063] Step T7: Model Configuration Selection and Output Step: Based on the constraints on model accuracy and inference efficiency in the actual deployment scenario, select the fusion weight configuration that satisfies the target preference from the Pareto front, and load the fusion weight configuration into the convolutional neural network model to construct the fusion convolutional neural network model under the corresponding trade-off point, thereby obtaining the Bayesian adaptive multi-objective convolutional neural network model as the final output model.

[0064] Step S3: Design a Pareto-aware dynamic feature weighting mechanism. This mechanism aims to balance the discriminative contribution of features with computational overhead. Specifically, it is implemented as a learnable dynamic feature weight generator, which receives the local feature vector of phenological changes from step S2 as input and assigns an adaptive weight to each feature component. The weight generation process is guided by the following optimization objective:

[0065] ;

[0066] in, It is an index variable; The total number of features; The first one extracted from crop canopy images Local feature vectors of phenological changes; It is to be sought, corresponding to Adaptive fusion weights; It is a discrimination scoring function used to quantify the first... eigenvectors Contribution to the discriminative power of identifying the state of faulty crops; To calculate the budget in real time; It is the total number of features; It is the first eigenvectors The computational cost constant; These are limiting conditions; It is to maximize the sum of the weighted discriminative power of features; Total computational cost; The variable is the feature weight vector to be learned, which is the variable that maximizes the objective function.

[0067] Step S4: The global depth feature vector is obtained by global average pooling, concatenated with the weighted local feature vector of phenological changes, and then projected through a fully connected layer to obtain the image feature vector. At the same time, the time series data vector is encoded into the environmental feature vector through another fully connected network. The weighted local feature vector of phenological changes is fused with the time series data vector to form the current fused feature.

[0068] Step S5: Introduce the Bi-GRU network and input the pre-cached historical fusion feature sequence into the Bi-GRU network R. The Bi-GRU network captures temporal dependencies from both forward and backward directions and outputs temporal enhancement features. Input the temporal enhancement features into the classifier to obtain preliminary phenological classification results and preliminary confidence scores.

[0069] Step S6: Construct a pre-built crop phenological feature library, which contains standard feature prototypes for each stage. Calculate the cosine similarity between the image feature vector and the prototype corresponding to the stage with the highest probability in the preliminary phenological classification results. Finally, generate a high-reliability confidence score through a calibration function. The module finally outputs the crop growth stage identifier.

[0070] Example 3

[0071] according to Figure 2 This embodiment is based on Embodiment 2. In this embodiment, the process of predicting environmental disturbances using the Fourier attention-state space prediction model specifically includes the following steps:

[0072] Step B1: Perform a Fast Fourier Transform on the historical time-series window data (i.e., time-series data) of key parameters of temperature and humidity in the global environmental state map to obtain its amplitude spectrum and phase spectrum; use the Fourier attention mechanism to design a learnable Fourier attention weight vector, the dimension of which is the same as the number of frequency components of the Fourier transform, and use softmax normalization to make the model focus on the frequency band most important to the prediction target, and finally extract the high-frequency residual components of the time-series data;

[0073] Step B2: Construct a linear Gaussian state-space model, the core of which is the system state transition rule;

[0074] Step B3: Using a random perturbation injection layer, the high-frequency residual components and short-term weather forecast data (i.e., external weather forecast vector, dimension 4, containing predicted values ​​of temperature, humidity, wind speed, and sunshine hours for the next 24 hours) are concatenated and coupled to the linear state-space model to obtain external information. Through the perturbation coupling matrix, the external information is processed through a fully connected network (input dimension 28, output dimension 2) to generate a perturbation vector. The perturbation vector is transformed into a correction term for the system state transition rule and added as an input to the state equation.

[0075] Step B4: Through a gating switching mechanism, a gating signal is designed, which is calculated from the trace of the uncertainty covariance matrix of the current state estimate. This dynamically adjusts the deterministic evolution of the linear state-space model, corresponding to the contribution ratio of the stationary evolution process and the random perturbation injection layer. The final state update integrates the deterministic and random parts. When the uncertainty is large, the model relies more on the external perturbation information of the random perturbation injection layer, and vice versa.

[0076] Step B5: During the training phase, a joint loss function is used to jointly optimize the attention head weights, linear state space model parameters, and perturbation coupling matrix of the Fourier mechanism to train the Fourier attention-state space prediction model. The construction process of the joint loss function is as follows: introduce the negative log-likelihood loss term and the spectrum reconstruction loss term, and weight the two to construct the joint loss function.

[0077] Joint loss function formula:

[0078] ;

[0079] in, At any moment Actual environmental observations; At any moment The hidden state vector inside the model, maintained by the state-space model, is an estimate of the real but unobservable state of the system. It is a conditional probability, representing the probability given the current hidden state vector of the model. Under the condition of, the observed true value The probability of; It is the total length of the time steps of the training sequence, that is, the number of historical observation data points used for one training session; For conditional probability Take the natural logarithm; Indicates from arrive Calculate it at every point in time, and then add up all the results; Indicates because of probability The logarithm between 0 and 1 If it's a negative number or zero, add a minus sign, - It becomes a non-negative value. When the prediction is completely accurate (probability is 1), this loss is 0. The more inaccurate the prediction, the greater this loss. The negative log-likelihood loss directly measures the degree of agreement between the distribution of environmental parameter values ​​predicted by the model and the distribution of actual observed values. It is the core loss for training state-space models.

[0080] ;

[0081] in, It is a sequence of future environmental parameters predicted by the model; The actual environmental parameter sequence corresponding to the prediction period is a sequence that is related to... Vectors of the same length; It is the Fast Fourier Transform operation, which transforms a sequence in the time domain to the frequency domain to obtain its spectrum; It is the L2 norm (Euclidean norm), which calculates the Euclidean distance between two vectors and measures the magnitude of their difference; For the spectral reconstruction loss, the spectral characteristics of the predicted sequence are constrained to be consistent with those of the real sequence;

[0082] ;

[0083] in, These are weighting coefficients. Model training requires minimizing the joint loss function (total loss) to measure the difference between the model's overall predictive performance and the real data.

[0084] The Adam optimizer is used to jointly optimize the attention head weights, state transition moments, coupling matrix, observation matrix, and noise parameters of the Fourier mechanism.

[0085] Step B6: In the inference phase, the trained Fourier attention-state space prediction model undergoes multi-step forward iteration (starting from the current state, iteratively executing steps B3-B4) and Monte Carlo sampling (i.e., for each step, sampling is performed from the distribution to generate multiple future trajectories); the proportion of all trajectories that meet the preset perturbation condition (temperature < threshold) is counted as the probability of occurrence, and the distribution of perturbation intensity is calculated, finally generating a probabilistic environmental perturbation prediction report.

[0086] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A remote monitoring system for greenhouses in smart agriculture, characterized in that, The system includes: an environmental sensing and actuation unit, a field intelligent control gateway, a cloud management platform, and an actuator controller; Environmental sensing and execution unit acquires multi-source sensor signals; The field intelligent control gateway connects to the actuator controller via a wireless local area network to generate the optimal control command set. The field intelligent control gateway integrates an environmental perception and data fusion module, a crop growth stage identification module, an environmental disturbance prediction module, a collaborative optimization decision-making module, and a multi-actuator scheduling module. The environmental perception and data fusion module processes multi-source sensor signals to generate a global environmental status map. The crop growth stage identification module has a built-in preset crop phenological feature library. It captures crop canopy images through an industrial camera and uses a Bayesian adaptive multi-objective CNN model to process the crop canopy images and the global environmental status map, dynamically identifying and outputting crop growth stage identifiers. The environmental disturbance prediction module takes the global environmental status map, crop growth stage indicators, and short-term weather forecast data from the cloud management platform as inputs, uses the Fourier attention-state space prediction model to make probabilistic predictions, and outputs a probabilistic environmental disturbance prediction report. The collaborative optimization decision-making module, based on the overall environmental status map, crop growth stage identifiers, and probabilistic environmental disturbance prediction reports, and calling the crop growth model of the cloud management platform, uses an adaptive bi-objective optimization algorithm to calculate and output theoretical control quantities. The multi-actuator scheduling module plans and outputs the optimal control instruction set based on theoretical control values, and then sends it to the actuator controller.

2. The greenhouse remote monitoring system for smart agriculture according to claim 1, characterized in that, The construction method of a Bayesian adaptive multi-objective CNN model includes the following steps: Step T1: Select a high-precision CNN model with structural alignment and a high-efficiency CNN model as the models to be fused; Step T2: For the network layer sequence in the model to be fused, extract the parameter difference representation of the network layer sequence, and based on the parameter difference representation, model the model block problem of the network layer sequence as a one-dimensional ordered block optimization problem. Introduce the parameter difference homogeneity constraint within the module and the information balance constraint between modules. Use dynamic programming algorithm to perform global optimal clustering partitioning of the network layer sequence and output the functional module partitioning scheme. Step T3: Preset functional module fusion weight vectors. During the forward propagation process, for each functional module in the functional module partitioning scheme, calculate the feature map output of the model to be fused at the corresponding position aligned with the functional module structure, and perform weighted fusion according to the functional module fusion weight vectors to obtain the fused output of the corresponding functional module, and construct the fused convolutional neural network model; perform forward inference on the fused convolutional neural network model to obtain the model performance evaluation results corresponding to the functional module fusion weight vectors; Step T4: Construct a Gaussian process surrogate model based on the fusion weight vector of the functional modules and the corresponding model performance evaluation results; Step T5: Based on the Gaussian process surrogate model, the expected hypervolume improved acquisition function is used to sample and update the fusion weight vector of the functional modules, and candidate fusion weight configurations are generated iteratively. The model performance corresponding to the candidate fusion weight configurations is evaluated based on the fusion convolutional neural network model until the preset convergence condition is met, and a Pareto front is formed. Step T6: Select the fusion weight configuration from the Pareto front and load the fusion weight configuration into the convolutional neural network model to obtain the Bayesian adaptive multi-objective CNN model.

3. The greenhouse remote monitoring system for smart agriculture according to claim 2, characterized in that, The process of using a Bayesian adaptive multi-objective CNN model to process crop canopy images, dynamically identify and output crop growth stage indicators, specifically includes the following steps: Step S1: Fuse the time-series data of the global environmental state map with the crop canopy image to construct a multimodal tensor; Step S2: Use the Yeesian adaptive multi-objective CNN model to extract the global deep semantic feature vector and local phenotypic feature vector of the multimodal tensor, and use the adaptive ROI technique to extract the local phenological change feature vector in the local phenotypic feature vector; Step S3: Design a dynamic feature weighting mechanism based on Pareto perception to assign differentiated weights to the local feature vectors of phenological changes, and obtain the weighted local feature vectors of phenological changes; Step S4: Aggregate the weighted local feature vector of phenological changes with the global deep semantic feature vector into an image feature vector, encode the time series data into an environmental feature vector, and fuse the image feature vector and the environmental feature vector to form the current fused feature; Step S5: Pre-cache historical fusion feature sequences and introduce them into the Bi-GRU network. Input the current fusion features and historical fusion feature sequences into the Bi-GRU network together, and output the preliminary phenological classification results and their preliminary confidence scores after time smoothing. Step S6: Construct a standard template for the crop phenological period feature library, calculate the cosine similarity between the image feature vector and the standard template for the crop phenological period feature library, calibrate the preliminary confidence level to determine the final confidence level, and output the crop growth stage identifier.

4. The greenhouse remote monitoring system for smart agriculture according to claim 1, characterized in that, The process of using a Fourier attention-state space prediction model to perform probabilistic predictions and outputting a probabilistic environmental disturbance prediction report includes the following steps: Step B1: Perform a fast Fourier transform on the time series data in the global environmental state map, and use the Fourier attention mechanism to extract the high-frequency residual components of the time series data; Step B2: Construct a linear state-space model, the core of which is the system state transition rule; Step B3: Use a random perturbation injection layer to couple high-frequency residual components and short-term weather forecast data to a linear state-space model to obtain external information. Through the perturbation coupling matrix, the external information is transformed into a correction term for the system state transition rule. Step B4: Dynamically adjust the contribution ratio of the deterministic evolution of the linear state-space model and the random perturbation injection layer through a gating switching mechanism; Step B5: During the training phase, a joint loss function is used to jointly optimize the attention head weights, linear state space model parameters, and perturbation coupling matrix of the Fourier mechanism to train the Fourier attention-state space prediction model. Step B6: In the inference phase, the trained Fourier attention-state space prediction model is subjected to multi-step forward iteration and Monte Carlo sampling to generate a probabilistic environmental disturbance prediction report.

5. A greenhouse remote monitoring system for smart agriculture according to claim 4, characterized in that, The process of constructing the joint loss function is as follows: introduce the negative log-likelihood loss term and the spectrum reconstruction loss term, and then weight the two to construct the joint loss function.

Images