Greenhouse disease propagation monitoring and optimizing system based on physical information neural network
Through a system based on physical information neural network, accurate monitoring and dynamic regulation of greenhouse disease transmission is achieved, and the problems of inaccurate prediction and improper regulation of traditional methods in complex environments are solved, and the accuracy and response speed of disease monitoring are improved.
Patent Information
- Application Number
- CN202510412281.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-07-11
AI Technical Summary
Traditional mechanistic models and purely data-driven machine learning methods are difficult to cope with complex and dynamically changing environments in greenhouse disease monitoring, resulting in inaccurate predictions and improper regulation.
The system based on physical information neural network is adopted to monitor the greenhouse environment through multimodal sensors, combine the edge computing module to perform physical modeling of multi-stage disease propagation, design the physical information neural network architecture, define the loss function, generate dynamic regulation parameter instructions, and optimize the greenhouse environment.
It has achieved the improvement of the accuracy and response speed of disease monitoring, optimized resource allocation, reduced the risk of disease occurrence, and provided technical support for smart agriculture.
Smart Images
Figure CN120295390A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the cross - technical field of smart agriculture and artificial intelligence, and specifically relates to the technical field of greenhouse disease spread monitoring and optimization based on physics - informed neural networks. Background Art
[0002] In greenhouse cultivation, the occurrence and spread of diseases have a significant impact on the healthy growth and yield of crops. To ensure the efficient production of crops, timely monitoring and early warning of disease spread, and formulating scientific regulation strategies are one of the key tasks in modern agriculture. Traditional disease monitoring and early warning technologies usually rely on two main methods: mechanism - model - based methods and pure data - driven machine - learning methods. Although these methods have helped improve the efficiency of greenhouse disease prevention and control to a certain extent, they still have significant limitations and are difficult to cope with the complex and dynamically changing greenhouse environment.
[0003] Traditional mechanism - model - based methods. Traditional mechanism models usually describe the disease spread process by establishing partial differential equations, such as diffusion - convection equations, etc. These models assume that the disease spread follows fixed diffusion and convection laws. However, in practical applications, the disease spread process is often affected by various complex factors, such as temperature, humidity, crop growth stage, etc. The changes in these factors often lead to dynamic changes in the spread laws. In addition, some parameters in traditional models (such as diffusion coefficients, convection velocities, etc.) usually rely on manual setting and lack adaptability, and cannot flexibly cope with the changes in the greenhouse environment (such as temperature and humidity fluctuations).
[0004] Pure data - driven machine - learning methods. Pure data - driven machine - learning methods (such as LSTM, CNN) learn the disease spread laws directly from data and have strong prediction capabilities. However, these methods also have some defects in applications. These models may output unreasonable results, such as negative disease concentration values or prediction results that violate the law of conservation of mass, which may lead to violations of physical laws and thus affect the reliability and practical operability of predictions. And currently, most control systems use fixed - threshold - triggered control instructions without distinguishing the disease spread stages, resulting in over - regulation or under - regulation. Summary of the Invention
[0005] In order to solve the technical problems that although the previous mechanism - model - based methods and pure data - driven machine - learning methods have helped improve the efficiency of greenhouse disease prevention and control to a certain extent, they still have significant limitations and are difficult to cope with the complex and dynamically changing greenhouse environment, the present invention proposes a greenhouse disease spread monitoring and optimization system based on physics - informed neural networks.
[0006] The system includes:
[0007] Multimodal Sensor Module: It is used to monitor greenhouse environmental data in real time, obtain greenhouse crop images in real time, and input the above information into the edge computing module;
[0008] Edge Computing Module: It is used to process the information input by the multimodal sensor module;
[0009] It is used to model the occurrence law of greenhouse diseases and define a multi-stage disease transmission physical model;
[0010] It is used to design a physics-informed neural network architecture and define the loss function of the physics-informed neural network based on the multi-stage disease transmission physical model;
[0011] It is used to obtain the predicted disease severity through the physics-informed neural network
[0012] Decision Module: It is used for the predicted disease severity to carry out severe level classification and generate dynamic adjustment control parameter instructions according to different severe levels;
[0013] Optimization Module: It is used to receive dynamic adjustment control parameter instructions and regulate the greenhouse environment.
[0014] The beneficial effects of the system described in the present invention are as follows:
[0015] Although monitoring greenhouse disease transmission based on the physics-informed neural network (PINN) has significant advantages, it also faces multiple technical difficulties in the implementation process. First, an efficient physics information embedding method needs to be designed to ensure that the physics-informed neural network can not only capture the biological nonlinear characteristics of disease transmission but also accurately reflect the coupling effects of multiple factors (such as the impact of environmental temperature and humidity on disease transmission), and design appropriate loss functions and training strategies to ensure that the transmission processes at different stages are reasonably modeled. Second, a refined optimization method needs to be developed that can dynamically calculate the ratio of disease concentration to carrying capacity and generate appropriate control strategies at different stages. The system needs to consider the comprehensive impact of different regulation factors (such as temperature, humidity, ventilation, etc.) on disease transmission and precisely adjust control measures according to real-time monitoring data and disease prediction results. Finally, loss functions and training strategies need to be carefully designed so that physical laws can be incorporated into the neural network in a reasonable way, which can not only ensure that the model learns the potential patterns in the data but also ensure that it conforms to biophysical laws.
[0016] The system described in the present invention realizes the accuracy and real-time performance of disease monitoring and control. Under the limitations of traditional methods and the deficiencies of pure data-driven models, it provides a new idea and technical means, which can effectively improve the accuracy and response speed of greenhouse disease transmission monitoring, optimize resource allocation, reduce the risk of disease occurrence, and overall construct an "monitoring - early warning - optimization" integrated solution, providing strong technical support for smart agriculture. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 It is a schematic diagram of the system described in the specific embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0018] The technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.
[0019] Embodiment 1
[0020] This embodiment provides a greenhouse disease transmission monitoring and optimization system based on a physics-informed neural network, as Figure 1 shown in the schematic diagram of the system. The system includes a total of four modules, namely a multi-modal sensor module, an edge computing module, a decision-making module, and an optimization module.
[0021] Multi-modal sensor module: used to monitor greenhouse environment data in real time and obtain greenhouse crop images in real time, and input the above information into the edge computing module.
[0022] The multi-modal sensor network in the multi-modal sensor module is used to monitor greenhouse environment data in real time, including temperature, humidity, light intensity, and CO2 concentration. The multi-modal sensor network includes a temperature sensor, a humidity sensor, a light intensity sensor, a CO2 concentration detector, and other environmental detection devices set according to specific needs and actual situations. Greenhouse crop images are obtained through an RGB imaging device or a hyperspectral imager.
[0023] Edge computing module: used to process the information input by the multi-modal sensor module;
[0024] used to model the occurrence law of greenhouse diseases, define a multi-stage disease transmission physical model; used to design the architecture of the physics-informed neural network and define the loss function of the physics-informed neural network based on the multi-stage disease transmission physical model; used to obtain the predicted disease degree through the physics-informed neural network
[0025] Decision-making module: used for the predicted disease severity to classify the severity level, and generate instructions for dynamically adjusting control parameters according to different severity levels.
[0026] Optimization module: used to receive instructions for dynamically adjusting control parameters and regulate the greenhouse environment. The optimization module is connected to environmental control systems such as the ventilation control system, spray control system, and supplementary lighting control system in the greenhouse, and inputs control parameters into the corresponding systems to achieve the regulation of the greenhouse environment.
[0027] Example 2
[0028] This example further limits Example 1.
[0029] When the edge computing module processes the information input by the multi-modal sensor module, it stores the measured greenhouse environment data, and uses the U-Net model to segment the disease area of the greenhouse crop image and extract the disease severity.
[0030] The edge computing module models the occurrence law of greenhouse diseases, defines a multi-stage disease propagation physical model, and divides the disease propagation process into stages through the model:
[0031] Latent period: The disease severity increases slowly and is significantly affected by the absolute temperature T and humidity H;
[0032] Outbreak period: The disease severity increases exponentially and is affected by the light L and CO2 level influence;
[0033] Decline period: The disease severity tends to be stable or decrease and is affected by control measures.
[0034] Through stage-by-stage modeling, the model can more accurately reflect the dynamic changes of disease propagation and adapt to different environmental conditions and control strategies.
[0035] The multi-stage disease propagation physical model is specifically:
[0036]
[0037] Among them, C represents the disease severity; t represents time; represents the partial derivative of the disease severity C with respect to time t; D(T,H) represents the diffusion coefficient, and the diffusion coefficient depends on temperature and humidity, Among them, D0 represents the reference diffusion coefficient, E a represents the activation energy, R represents the ideal gas constant, T represents the absolute temperature, H represents the humidity, and H0 is a constant; represents the Laplace operation on C; v(t) represents the air flow velocity; represents the gradient operation on C; L represents the light intensity, Represents the CO2 concentration.
[0038] For Modeling is carried out in three stages, considering the influence of different environmental factors, which are respectively:
[0039] Latent period:
[0040] Outbreak period:
[0041] Decline period:
[0042] Among them, α1, α2 and α3 are the coefficients of different stages of the disease; T opt Represents the optimal temperature; H opt Represents the optimal humidity; Represents the maximum carrying capacity of the greenhouse for the disease degree, which depends on the light intensity and CO2 concentration; Q spray Represents the spray volume, Q max Represents the maximum spray volume.
[0043] The embedding of these physical mechanisms in the multi-stage disease propagation physical model enables the model to accurately describe the dynamic process of disease propagation, with high scientificity and reliability.
[0044] max Among them, K represents the maximum possible value of the disease degree, L0 represents the half-saturation constant of the light intensity, and C0 represents the half-saturation constant of the CO2 concentration.
[0045] The edge computing module designs the architecture of the physics-informed neural network and defines the loss function of the physics-informed neural network based on the multi-stage disease propagation physical model; the architecture of the physics-informed neural network includes an input layer, a hidden layer, and an output layer, specifically:
[0046] Input layer: Input the normalized spatial coordinates (x, y), time t, and environmental parameters
[0047] Hidden layer: Adopt a 5-layer residual network, with 256 nodes in each layer, and the Swish activation function is used as the activation function;
[0048] Output layer: Output the predicted disease degree And can also output The partial derivative with respect to time For The result of the gradient operation For The result of the Laplace operation
[0049] The loss function of the physical information neural network is specifically as follows:
[0050]
[0051] Among them, λ1, λ2, λ3, and λ4 are the coefficients of each term in the loss function respectively;
[0052] represents the data loss term, where N represents the number of data points, represents the disease degree predicted by the physical information neural network for the data point with the normalized spatial coordinates of (x i , y i , y i ) at time t, and C obs (x i , t i ) represents the actual observed disease degree of the data point with the normalized spatial coordinates of (x i , y i , y i ) at time t.
[0053] represents the physical information neural network residual term, where M represents the number of sample points used to calculate the loss, represents the predicted disease degree 's partial derivative with respect to time t j , D(T j , H j ) represents the diffusion coefficient when the physical information neural network reaches the j-th round, represents the Laplace operation on ; v j represents the air flow velocity when the physical information neural network reaches the j-th round; represents the gradient operation on ; T j represents the absolute temperature when the physical information neural network reaches the j-th round, H j represents the humidity when the physical information neural network reaches the j-th round, L j represents the light intensity when the physical information neural network reaches the j-th round, represents the CO2 concentration when the physical information neural network reaches the j-th round.
[0054] is modeled in three stages, which are respectively:
[0055] Incubation period:
[0056] Outbreak period: Outbreak period:
[0057] Decline period:
[0058] Among them, Among them, L 0,j represents the light intensity half-saturation constant when the physics-informed neural network progresses to the j-th round, and C 0,j represents the CO2 concentration half-saturation constant when the physics-informed neural network progresses to the j-th round. Ensure that the predicted disease propagation process of the model conforms to the kinetic law of the multi-stage disease propagation physical model.
[0059] represents the boundary condition term, Among them, P represents the number of points for boundary condition inspection, represents the boundary when conducting boundary condition inspection, represents the disease severity predicted by the physics-informed neural network at time t k for the boundary point with the normalized spatial coordinates (x k , y k ). Ensure that the prediction result of the model under the boundary condition conforms to the physical constraint.
[0060] represents the stage consistency term, Among them, Q represents the number of data points for evaluating stage consistency, represents the disease severity predicted by the physics-informed neural network at time t q for the data point with the normalized spatial coordinates (x q , y q ), and C stage (x q , t q ) represents the expected threshold of the disease development stage for the data point with the normalized spatial coordinates (x q , y q , y q ) at time t Ensure that the prediction results of the model in different stages conform to the stage characteristics of disease propagation.
[0061] The loss function of the physics-informed neural network enables the output of the model to have a clear physical meaning, avoiding the "black box" problem that may occur in pure data-driven models; in the case of sparse data or high noise, physical constraints can help the model maintain reasonable prediction ability and reduce the dependence on a large amount of high-quality data; physical constraints enable the model to still maintain stable prediction performance in the face of noisy data or outliers.
[0062] Example 3
[0063] This example further limits Example 1. For the predicted disease severity to conduct a severe level classification specifically as follows:
[0064] Mild:
[0065] Moderate:
[0066] Severe:
[0067] Generate a dynamic adjustment control parameter instruction according to different severe levels, specifically as follows:
[0068] Mild: Based on the current basic ventilation volume, continue to increase the ventilation volume Q fan ,
[0069] Moderate: Start the spray system, and the spray volume is Q spray ,
[0070] Severe: Adjust the spray volume of the spray system to Q spray , and adjust the light supplement amount of the light supplement system to Q light ,
[0071] Wherein, represents the maximum value of the disease severity predicted by the physics-informed neural network,
[0072] k1 is a proportionality coefficient used to adjust the ventilation volume according to the disease severity; k2 is a proportionality coefficient used to adjust the spray volume according to the severity of the disease; k3 is a proportionality coefficient used to adjust the intensity of light supplement according to the maximum predicted value of the disease severity; for the determination of k1, k2 and k3, these coefficients are set through experimental adjustment or statistical analysis based on historical operation data to ensure that they can effectively reflect the actual needs of corresponding measures in different disease stages. The setting of these coefficients should consider the flexibility of operation and the minimum interference to the plant growth environment, while effectively controlling the development of the disease.
[0073] Example 4
[0074] This example further limits Example 1. The predicted disease severity is obtained through a physics-informed neural network, and the greenhouse is monitored and controlled for strawberry gray mold disease.
[0075] I. Training stage.
[0076] 1. Data preprocessing:
[0077] (1) Data collection:
[0078] Collect greenhouse environmental data (absolute temperature T, humidity H, light intensity L, CO2 concentration ) using a multi-modal sensor network;
[0079] Obtain strawberry gray mold disease images through an RGB imaging device or a multi-spectral imager, segment the disease area using a U-Net model, and extract the disease severity C obs .
[0080] (2) Data normalization:
[0081] Normalize the input data so that it is distributed in the interval [0,1]: where T min represents the lowest absolute temperature, T max represents the highest absolute temperature, H min represents the lowest humidity, H max represents the highest humidity.
[0082] Perform a logarithmic transformation on the disease severity C obs to reduce data skewness:
[0083] where C min represents the value when the disease severity is the mildest, C max represents the value when the disease severity is the most severe.
[0084] (3) Dataset division:
[0085] Divide the data into a training set, a validation set, and a test set in chronological order.
[0086] 2. Training of the physics-informed neural network model:
[0087] (1) Network initialization:
[0088] Initialize the weights θ and biases b of the physics-informed neural network; set the learning rate η = 10 -3 , and the optimizer Adam.
[0089] (2) Forward propagation:
[0090] Input the normalized spatial coordinates (x,y), time t, and environmental parameters
[0091] Calculate the predicted disease severity value and its partial derivatives
[0092] (3) Loss calculation:
[0093] Calculate data loss
[0094] Calculate the residual loss of the physics-informed neural network
[0095] Calculate the boundary condition loss
[0096] Calculate the stage consistency loss
[0097] Total loss:
[0098] (4) Backpropagation and optimization:
[0099] Calculate the gradient of the loss function with respect to the network parameters θ
[0100] Update the parameters using the Adam optimizer:
[0101] (5) Dynamic weight adjustment:
[0102] Statistically calculate the gradient magnitudes of each loss term every 1000 iterations, and dynamically adjust the weights λ1, λ2, λ3, and λ4:
[0103]
[0104] II. Inference stage
[0105] 1. Disease severity prediction:
[0106] (1) Input data preparation:
[0107] Collect environmental data and spatial coordinates in real time; normalize the input data.
[0108] (2) Forward propagation calculation:
[0109] Input the normalized data into the trained physics-informed neural network model to calculate the disease severity prediction value Calculate the partial derivatives through automatic differentiation
[0110] (3) Visualization of the disease propagation process:
[0111] Map the prediction results to a heatmap and display the disease propagation trend in real time.
[0112] 2. Regulation Strategy Generation:
[0113] (1) Disease Severity Grading:
[0114] Based on the prediction level Classify the disease severity levels:
[0115] Mild:
[0116] Moderate:
[0117] Severe:
[0118] (2) Regulation Instruction Generation:
[0119] Mild: Continuously increase the ventilation volume Q based on the current basic ventilation volume fan ,
[0120] Moderate: Activate the spray system with a spray volume of Q spray ,
[0121] Severe: Adjust the spray volume of the spray system to Q spray , and adjust the light supplement volume of the light supplement system to Q light ,
[0122] III. Formula Calculation Example
[0123] Taking the calculation of the residual term of the physics-informed neural network in the disease environment as an example:
[0124] (1) Input environmental parameters and the spatial coordinates (x j , y j );
[0125] (2) Calculate the diffusion coefficient D(T j , H j ):
[0126]
[0127] (3) Calculate the source term
[0128] If it is the incubation period:
[0129] If it is the outbreak period:
[0130] (4) Calculate the residual of the physics-informed neural network in the disease environment:
[0131] The system described in the present invention constructs a multi-stage physical model of disease spread by integrating the disease occurrence law, physical mechanism model and multi-source sensor data. Compared with the traditional PDE numerical solution method, the prediction accuracy of the physical information neural network model is significantly improved, and the mean square error of prediction is reduced by 28%. The regulation strategy based on the disease severity grading reduces the amount of pesticide used by 15%, reduces the agricultural production cost, and guarantees the quality of agricultural products.
[0132] Limitations of the traditional PDE model: The traditional PDE model usually assumes that the disease spread follows fixed diffusion and convection laws. However, in practical applications, the disease spread process is often affected by various complex factors, such as temperature, humidity, crop growth stage, etc. The changes in these factors lead to dynamic changes in the spread law, and it is difficult for the traditional model to respond flexibly.
[0133] Deficiencies of the pure data-driven model: Although pure data-driven machine learning methods (such as LSTM, CNN) have strong prediction capabilities, they may output unreasonable results (such as negative disease concentration values or prediction results that violate the law of conservation of mass), resulting in violations of physical laws and affecting the reliability and practical operability of the prediction.
[0134] In contrast, the multi-stage physical model of disease spread can not only accurately describe the dynamic process of disease spread, but also flexibly respond to the changes in environmental factors, with higher scientificity and reliability.
Claims
1. A greenhouse disease spread monitoring and optimization system based on a physics-informed neural network, characterized in that, The system includes: A multi-modal sensor module: used to monitor greenhouse environment data in real time and obtain greenhouse crop images in real time, and input the above information into the edge computing module; An edge computing module: used to process the information input by the multi-modal sensor module; Used to model the occurrence law of greenhouse diseases and define a multi-stage disease propagation physical model; Used to design a physics-informed neural network architecture and define the loss function of the physics-informed neural network based on the multi-stage disease propagation physical model; For obtaining a predicted disease severity through a physics-informed neural network Decision-making module: used for the predicted disease severity to conduct a severity level classification and generate a dynamically adjusted control parameter instruction according to different severity levels; An optimization module: used to receive instructions for dynamically adjusting control parameters and regulate the greenhouse environment.
2. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 1, characterized in that The greenhouse environment data includes temperature, humidity, light intensity, and CO2 concentration.
3. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 1, wherein The greenhouse crop images are obtained by an RGB imaging device or a multispectral imager.
4. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 1, characterized in that, The specific form of the multi-stage disease propagation physical model is: Among them, C represents the disease severity; t represents time; represents the partial derivative of the disease severity C with respect to time t; D(T, H) represents the diffusion coefficient, where D0 represents the reference diffusion coefficient, E a represents the activation energy, R represents the ideal gas constant, T represents the absolute temperature, H represents the humidity, and H0 is a constant; represents the Laplace operation on C; v(t) represents the air flow velocity; represents the gradient operation on C; L represents the light intensity, represents the CO2 concentration; Pairwise Modeling is carried out in three stages, namely: Incubation period: Outbreak period: Decline stage: Among them, α1, α2, and α3 are coefficients at different stages of the disease; T opt represents the optimal temperature; H opt represents the optimal humidity; represents the maximum bearing capacity of the greenhouse for the disease severity, Q spray represents the spraying volume, Q max represents the maximum spraying volume.
5. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 4, characterized in that, Among them, K max represents the maximum possible value of the disease severity, L0 represents the half-saturation constant of the light intensity, and C0 represents the half-saturation constant of the CO2 concentration.
6. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 5, characterized in that The physics-informed neural network architecture includes an input layer, a hidden layer, and an output layer, specifically: Input layer: Input normalized spatial coordinates (x, y), time t, and environmental parameters Hidden layer: Adopt a 5-layer residual network, with 256 nodes in each layer, and the activation function adopts the Swish activation function; Output layer: Output the predicted disease severity 7. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 6, characterized in that The specific form of the loss function of the physics-informed neural network is: Among them, λ1, λ2, λ3, and λ4 are the coefficients of each term in the loss function respectively; Represents the data loss term, where N represents the number of data points, represents the disease severity predicted by the physics-informed neural network for the data point with normalized spatial coordinates (x i , y i , y i ) at time t, and C obs (x i , t i ) represents the actual observed disease severity of the data point with normalized spatial coordinates (x i , y i , y i ) at time t; Represents the residual term of the physics-informed neural network, where M represents the number of sample points used to calculate the loss, represents the predicted disease severity at time t j partial derivative of, D(T j , H j ) represents the diffusion coefficient of the physics-informed neural network at the j-th iteration, represents the Laplace operation on ; v j represents the air flow velocity of the physics-informed neural network at the j-th iteration; represents the gradient operation on ; T j represents the absolute temperature of the physics-informed neural network at the j-th iteration, H j represents the humidity of the physics-informed neural network at the j-th iteration, L j represents the light intensity of the physics-informed neural network at the j-th iteration, represents the CO2 concentration of the physics-informed neural network at the j-th iteration; For Modeling is carried out in three stages, namely: Incubation period: Outbreak period: Decline period: Among them, where L 0,j represents the light intensity half-saturation constant at the j-th round of the physics-informed neural network, and C 0,j represents the CO2 concentration half-saturation constant at the j-th round of the physics-informed neural network; Indicates the boundary condition item, where P represents the number of points for boundary condition checking, represents the boundary when performing boundary condition checking, represents the physical information neural network at time t k when the predicted disease severity of the boundary point with normalized spatial coordinates (x k , y k ) is Indicates the phase consistency term, where Q represents the number of data points for evaluating phase consistency, Indicates the disease severity predicted by the physics-informed neural network at time t q for a data point with normalized spatial coordinates (x q , y q ), and C stage (x q , t q ) represents the expected threshold of the disease development phase for a data point with normalized spatial coordinates (x q , y q , y q ) at time t.
8. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 7, characterized in that For the predicted disease severity The specific severity levels are as follows: Mild: Moderate: Severe:
9. The greenhouse disease spread monitoring and optimization system of the physical information neural network according to claim 8, characterized in that, The instruction for dynamically adjusting the control parameters generated according to different severity levels is specifically: Mild: Continue to increase the ventilation volume Q based on the current basic ventilation volume fan , Medium: Activate the spray system with a spray volume of Q spray , Severe: Adjust the spraying volume of the spraying system to Q spray and adjust the supplementary light volume of the supplementary light system to Q light , Among them, represents the maximum value of the disease degree predicted by the physical information neural network, and k1, k2, and k3 are all proportionality coefficients.