Seed seedling artificial climate chamber control system and control method thereof

Abstract

The invention relates to the technical field of phytotron control systems, in particular to a seed seedling phytotron control system and a control method thereof, which integrate multi-dimensional environment perception, intelligent analysis and decision, precise environment regulation and control, a central controller and a data storage and analysis platform. A seedling raising environment is monitored in real time through temperature and humidity, illumination, CO2 concentration, soil moisture and an image sensor, an intelligent analysis decision module preprocesses data, establishes a growth model, generates a control strategy and optimizes a target by using an advanced algorithm, and an accurate environment regulation and control module executes the strategy and regulates temperature, humidity, illumination and the like. The central controller coordinates operation of the modules and realizes man-machine interaction, and the data storage and analysis platform is responsible for storing data and supporting cloud analysis so as to realize efficient seedling culture management and significantly improve the seed emergence rate from 75% of a traditional method to 92%, and greatly reduce seed waste.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial climate chamber control systems, and more specifically, to a control system and a control method for an artificial climate chamber for seedling raising. Background Art

[0002] In recent years, with the progress of technology, some researchers have begun to attempt to apply artificial climate chamber technology to seedling raising. For example, Ren Anhua et al. proposed an intelligent artificial climate chamber for wheat phenotyping research in "Evaluation of an intelligent artificial climate chamber for high-throughput crop phenotyping in wheat" (Plant Methods, 2022, 18:77). Although this research has made certain progress in environmental control and data collection, it mainly focuses on wheat and is not fully applicable to fiber crops.

[0003] Currently, the field of fiber crop seedling raising still faces the following technical problems:

[0004] 1. Lack of a precise environmental control system for specific physiological characteristics, making it difficult to provide optimal growth conditions throughout the seedling raising process.

[0005] 2. The existing systems have insufficient intelligence and cannot achieve adaptive adjustment and optimization of environmental parameters.

[0006] 3. Low resource utilization efficiency, especially in the use of water, fertilizers, and energy, where there is still a large room for optimization.

[0007] 4. Unstable seedling raising quality and poor seedling age consistency, affecting subsequent field management and yield.

[0008] 5. Lack of a comprehensive seedling condition monitoring and early warning mechanism, making it difficult to detect and solve problems during the seedling raising process in a timely manner.

[0009] In response to these problems, the present invention proposes a control system and a control method for an artificial climate chamber for seedling raising. Summary of the Invention

[0010] The purpose of the present invention is to provide a control system and a control method for an artificial climate chamber for seedling raising, which can achieve dynamic and precise control of the seedling raising environment through multi-dimensional environmental perception, intelligent analysis and decision-making, and precise environmental regulation, significantly improve the seed emergence rate and seedling quality, and at the same time optimize the resource utilization efficiency.

[0011] The present invention provides a control system for a seedling artificial climate chamber, comprising: a multi-dimensional environment perception module, an intelligent analysis and decision-making module, a precise environment regulation module, a central controller, and a data storage and analysis platform; wherein,

[0012] The multi-dimensional environment perception module includes a temperature and humidity sensor array, a light sensor, a CO2 concentration sensor, a soil moisture sensor, and an image acquisition system, and is used for collecting multi-dimensional parameter data of the seedling environment in real time;

[0013] The intelligent analysis and decision-making module is connected to the multi-dimensional environment perception module, and includes a multi-modal data fusion unit, a bidirectional recurrent neural network unit, a deep reinforcement learning unit, a plant digital twin model unit, and a multi-objective optimization unit. Among them, the multi-modal data fusion unit is used for uniformly preprocessing heterogeneous environment data; the bidirectional recurrent neural network unit is used for establishing a spatio-temporal growth state model of seeds; the deep reinforcement learning unit is used for generating an environment control strategy; the plant digital twin model unit is used for simulating and predicting the influence of different environmental parameter combinations on plant growth; the multi-objective optimization unit is used for making a balanced decision between maximizing yield and minimizing resource consumption;

[0014] The precise environment regulation module is connected to the intelligent analysis and decision-making module, and includes a temperature regulation system, a humidity regulation system, a light control system, a CO2 supplementation system, and a nutrient solution irrigation system, and is used for executing the environment control strategy;

[0015] The central controller is used for coordinating the work of each module and performing human-computer interaction;

[0016] The data storage and analysis platform is used for storing environmental parameters and growth data, and supports data cloud analysis.

[0017] Preferably, the multi-modal data fusion unit adopts a data fusion algorithm combining Kalman filtering and deep belief network to fuse the numerical data from the temperature and humidity sensor array, the light sensor, the CO2 concentration sensor, and the soil moisture sensor, as well as the image data from the image acquisition system, solves the problems of different sampling frequencies and noise interference of different modal data, and generates a standardized environmental state vector.

[0018] Preferably, the bidirectional recurrent neural network unit constructs a bidirectional long short-term memory network containing 128 hidden units, has a temporal-spatial double-recursive structure, can simultaneously capture the short-term fluctuations and long-term trends of plant growth, the temporal modeling range is from 1 hour to 7 days, the spatial modeling resolution is 5 cubic centimeters, and the prediction accuracy reaches 87.5%.

[0019] Preferably, the deep reinforcement learning unit constructs an environmental regulation decision-making system based on an improved deep Q-network, constructs a multi-dimensional state space including temperature, humidity, light, and CO2 concentration, innovatively introduces plant physiological indicators as direct inputs to the reward function, and uses an experience replay pool technology to store and optimize past decision-making experiences to achieve continuous optimization.

[0020] Preferably, the plant digital twin model unit integrates a plant physiology model and a machine learning model, can simulate the effects of more than 1000 combinations of environmental parameters on plant growth in a virtual environment, mathematically models key physiological processes such as photosynthesis, transpiration, and nutrient absorption, with a prediction time up to 7 days and the prediction error controlled within the range of ±8%.

[0021] Preferably, the multi-objective optimization unit uses an improved NSGA-III algorithm to construct a multi-dimensional objective function including yield, energy consumption, water consumption, and fertilizer consumption, conducts Pareto optimal solution search, realizes comprehensive decision-making for maximizing yield, minimizing resource consumption, and minimizing environmental impact, and improves the comprehensive resource utilization efficiency by 38.5%.

[0022] Preferably, the image acquisition system includes a high-definition camera and a multi-spectral camera. The resolution of the high-definition camera is not less than 4K, and a panoramic image is taken once an hour; the multi-spectral camera includes three bands of visible light, near-infrared, and red edge, and multi-spectral images are collected at 10:00, 14:00, and 18:00 every day respectively, for capturing visible changes and non-visible physiological characteristics of plant growth status.

[0023] Preferably, the lighting control system includes a full-spectrum LED lamp and a mobile supplementary lighting vehicle. The full-spectrum LED lamp includes four primary color LEDs of red, blue, green, and far-red, realizes stepless dimming from 0 to 100% through PWM dimming technology, and the dimming accuracy is 1%; the mobile supplementary lighting vehicle can perform directional supplementary lighting on a local area according to the plant growth distribution map provided by the image acquisition system.

[0024] Preferably, the central controller is connected to the multi-dimensional environment perception module, the intelligent analysis and decision-making module, and the precise environment regulation module, and includes an industrial-grade PLC, a human-machine interface, and a remote monitoring module. Among them, the industrial-grade PLC is responsible for the overall coordination and real-time control of the system; the human-machine interface is a touch screen display for parameter setting and status monitoring; the remote monitoring module supports real-time viewing and controlling the operation status of the system through a mobile APP.

[0025] A control method for the seedling raising artificial climate chamber control system according to the above, comprising the following steps:

[0026] 1) Collect multi-dimensional parameter data of the seedling raising environment in real time through the multi-dimensional environment perception module;

[0027] 2) Perform heterogeneous fusion processing on the multi-dimensional parameter data through the multi-modal data fusion unit;

[0028] 3) Establish a spatio-temporal growth state model of seeds through the bidirectional recurrent neural network unit;

[0029] 4) Generate an environmental control strategy through the deep reinforcement learning unit;

[0030] 5) Simulate and predict the growth effects of different environmental parameter combinations through the plant digital twin model unit;

[0031] 6) Make a balance decision between yield and resource consumption through the multi-objective optimization unit, and optimize the environmental control strategy;

[0032] 7) Execute the optimized environmental control strategy through the precise environmental regulation module;

[0033] 8) Update the plant digital twin model and the experience replay pool to achieve continuous optimization of the control system.

[0034] The present invention has the following beneficial effects:

[0035] 1. Significantly improve the seed emergence rate, from 75% of the traditional method to 92%, and greatly reduce seed waste.

[0036] 2. Improve the consistency of seedling age, with the coefficient of variation reduced from 18% to 8%, laying a foundation for subsequent field management.

[0037] 3. Promote root development, with the root-shoot ratio increased from 0.35 to 0.48, enhancing the stress resistance and growth potential of seedlings.

[0038] 4. Increase the chlorophyll content and stem strength, with increases of 28.1% and 50% respectively, laying a foundation for high yield.

[0039] 5. Greatly improve the water use efficiency, from 3.5 g / L to 5.2 g / L, saving water resources.

[0040] 6. Reduce energy consumption, from 12 kWh / m 2 reduced to 10 kWh / m 2 , achieving a win-win situation of economic and ecological benefits. Description of the Drawings

[0041] Figure 1 It is the overall system architecture diagram of the present invention.

[0042] Figure 2This is the architecture diagram of the intelligent analysis and decision-making module of the present invention.

[0043] Figure 3 This is the flow chart of the control method of the present invention.

[0044] Figure 4 This is the work flow chart of the multi-modal data fusion unit of the present invention.

[0045] Figure 5 This is the schematic structural diagram of the bidirectional recurrent neural network unit of the present invention.

[0046] Figure 6 This is the schematic diagram of the decision-making process of the deep reinforcement learning unit of the present invention.

[0047] Figure 7 This is the architecture diagram of the plant digital twin model unit of the present invention.

[0048] Figure 8 This is the schematic diagram of the optimization process of the multi-objective optimization unit of the present invention. Detailed implementation manners

[0049] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not used to limit the present invention.

[0050] Embodiment 1: Overall architecture of the seedling-raising artificial climate chamber control system

[0051] As Figure 1 shown, the seedling-raising artificial climate chamber control system provided by the present invention includes a multi-dimensional environment perception module 1, an intelligent analysis and decision-making module 2, a precise environment regulation module 3, a central controller 4, and a data storage and analysis platform 5.

[0052] The multi-dimensional environment perception module 1 includes a temperature and humidity sensor array, a light sensor, a CO2 concentration sensor, a soil moisture sensor, and an image acquisition system, and is used to collect multi-dimensional parameter data of the seedling-raising environment in real time. Among them, the temperature and humidity sensor array is composed of multiple high-precision temperature and humidity sensors, and is arranged in the seedling-raising area according to a grid of 0.5m×0.5m×0.5m; the light sensor is used to measure the light intensity and spectral distribution; the CO2 concentration sensor is used to monitor the gas composition; the soil moisture sensor is used to monitor the moisture content of the substrate; the image acquisition system includes a high-definition camera and a multi-spectral camera, and is used to collect images of the plant growth state.

[0053] The intelligent analysis and decision-making module 2 is connected to the multi-dimensional environment perception module 1 and is the core module of this system. It includes a multi-modal data fusion unit 21, a bidirectional recurrent neural network unit 22, a deep reinforcement learning unit 23, a plant digital twin model unit 24, and a multi-objective optimization unit 25. Through a series of advanced artificial intelligence algorithms, this module deeply analyzes and processes the collected environmental parameter data to generate optimal environmental control strategies.

[0054] The precise environment regulation module 3 is connected to the intelligent analysis and decision-making module 2 and includes a temperature regulation system, a humidity regulation system, a lighting control system, a CO2 supplementation system, and a nutrient solution irrigation system, which are used to precisely regulate the seedling-raising environment according to the environmental control strategies generated by the intelligent analysis and decision-making module 2.

[0055] The central controller 4 is connected to the multi-dimensional environment perception module 1, the intelligent analysis and decision-making module 2, and the precise environment regulation module 3. It includes an industrial-grade PLC, a human-machine interface, and a remote monitoring module, which are used to coordinate the work of each module and conduct human-machine interaction.

[0056] The data storage and analysis platform 5 is connected to the central controller 4 and is used to store environmental parameters and growth data, and supports data cloud analysis to provide data support for the continuous optimization of the system.

[0057] Embodiment 2: Detailed Structure of the Intelligent Analysis and Decision-making Module

[0058] As Figure 2 shown, the intelligent analysis and decision-making module 2 is the core of this system and contains five functional units, which form a progressive intelligent decision-making chain.

[0059] The multi-modal data fusion unit 21 is the foundation of the entire intelligent decision-making chain. It uses a data fusion algorithm that combines Kalman filtering and deep belief networks to preprocess and fuse heterogeneous data from the multi-dimensional environment perception module 1. This unit solves the problems of different sampling frequencies and noise interference of different modal data and generates a standardized environmental state vector. The core mathematical model of the Kalman filtering algorithm is as follows:

[0060] X(k|k-1) = F(k)X(k-1|k-1) + B(k)u(k),

[0061] P(k|k-1) = F(k)P(k-1|k-1)F(k) T + Q(k),

[0062] Among them, X(k|k - 1) represents the state prediction value at time k based on the information at time k - 1, which is an n-dimensional vector containing environmental parameters such as temperature, humidity, light intensity, etc.; X(k - 1|k - 1) represents the state estimation value at time k - 1, F(k) represents the state transition matrix with the dimension of n×n, describing how the state evolves from time k - 1 to time k; B(k) represents the control input matrix with the dimension of n×m; u(k) represents the control input vector at time k with the dimension of m; P(k|k - 1) represents the state prediction error covariance matrix at time k based on the information at time k - 1, with the dimension of n×n; P(k - 1|k - 1) represents the state estimation error covariance matrix at time k - 1, and F(k) T represents the transpose of the matrix F(k); Q(k) represents the process noise covariance matrix with the dimension of n×n, characterizing the uncertainty of the system model.

[0063] For the update of sensor measurement data, the algorithm performs:

[0064] K(k) = P(k|k - 1)H(k) T [H(k)P(k|k - 1)H(k) T + R(k)] -1 ,

[0065] X(k|k) = X(k|k - 1)+K(k)[Z(k)-H(k)X(k|k - 1)],

[0066] P(k|k) = [I - K(k)H(k)]P(k|k - 1),

[0067] Among them, K(k) represents the Kalman gain matrix at time k with the dimension of n×p, which is used to balance the weights of prediction and measurement; H(k) represents the observation matrix with the dimension of p×n, describing how the state variables are mapped to the observation space; H(k) T represents the transpose of the matrix H(k); Z(k) represents the observation vector at time k with the dimension of p, containing the measurement values of each sensor; X(k|k) represents the optimal state estimation value at time k; P(k|k) represents the state estimation error covariance matrix at time k, I represents the identity matrix with the dimension of n×n; R(k) represents the observation noise covariance matrix with the dimension of p×p, characterizing the uncertainty of the measurement process, [H(k)P(k|k - 1)H(k) T + R(k)] -1 represents the inverse matrix of the matrix within the square brackets.

[0068] In this system, the state vector X contains environmental parameters such as temperature, humidity, light intensity, and CO2 concentration, and its dimension is usually 10 - 20; the observation vector Z contains the direct measurement values of each sensor. Through Kalman filtering, the system can effectively fuse multi-source sensor data, filter out random fluctuations and measurement noise, and provide a smooth and accurate environmental state estimate. For example, when the temperature sensor fluctuates, the algorithm can predict a more reasonable temperature value based on historical data and the system model, avoiding overreaction of the control system.

[0069] Through this algorithm, the system can not only effectively remove the noise in the sensor data, but also integrate data from different sources and with different sampling frequencies into an environmental state vector with a unified format and time standard, providing a high-quality data basis for subsequent intelligent decision-making. Experiments show that the data processing accuracy after using Kalman filtering has increased by 43%, significantly enhancing the system's perception ability and decision-making reliability.

[0070] Based on the fused data, the bidirectional recurrent neural network unit 22 constructs a bidirectional long short-term memory network (Bi-LSTM) with 128 hidden units. This network has a dual temporal-spatial recurrent structure and can capture both short-term fluctuations and long-term trends in plant growth simultaneously. The temporal modeling range is from 1 hour to 7 days, the spatial modeling resolution is 5 cubic centimeters, and the prediction accuracy reaches 87.5%. The core of this unit is to construct a spatio-temporal model of the plant growth state, providing a basis for subsequent decision-making.

[0071] The deep reinforcement learning unit 23 is the intelligent decision-making core of the system, and constructs an environmental regulation decision-making system based on the improved deep Q-network (DQN). This unit first constructs a multi-dimensional state space including temperature, humidity, light, and CO2 concentration; secondly, defines an action space including the operation instructions of each regulation device; thirdly, innovatively introduces plant physiological indicators as the direct input of the reward function; then, uses the experience replay pool technology to store and optimize past decision-making experiences. The core update formula of Q learning is:

[0072]

[0073] Among them, Q(s,a) represents the value function of taking action a in state s; α is the learning rate; r is the immediate reward; γ is the discount factor; s ′ is the next state; represents the maximum action value of the next state. By continuously interacting with the environment and learning, this unit can generate an optimal environmental control strategy, and the control accuracy has been improved by 65%.

[0074] The plant digital twin model unit 24 integrates the plant physiology model and the machine learning model to simulate the effects of more than 1000 combinations of environmental parameters on plant growth in a virtual environment. This unit mathematically models key physiological processes such as photosynthesis, transpiration, and nutrient absorption, constructs a digital twin of plant growth, enabling the system to predict the effects of different combinations of environmental parameters before actual regulation. The prediction duration is up to 7 days, and the prediction error is controlled within the range of ±8%.

[0075] The multi-objective optimization unit 25 uses the improved NSGA-III algorithm to construct a multi-dimensional objective function including yield, energy consumption, water consumption, and fertilizer consumption, and conducts Pareto optimal solution search. This unit realizes the comprehensive decision-making of maximizing yield, minimizing resource consumption, and minimizing environmental impact, improving the comprehensive resource utilization efficiency by 38.5%. This unit is the final link of the entire intelligent decision-making chain, responsible for selecting the optimal solution among various possible environmental control strategies.

[0076] These five units form a complete intelligent decision-making chain, progressing step by step from multi-source data fusion to multi-objective optimization decision-making. Each step is based on the previous step and provides higher-level functions, ultimately achieving highly intelligent environmental control decisions.

[0077] Example 3: Implementation of the multi-dimensional environmental perception module

[0078] This example details the composition and working principle of the multi-dimensional environmental perception module 1.

[0079] The temperature and humidity sensor array includes multiple high-precision temperature and humidity sensors distributed at different positions in the seedling raising area. These sensors are arranged in a 0.5m×0.5m×0.5m grid to form a three-dimensional perception network. The measurement accuracy of each sensor is ±0.1℃ and ±1%RH, ensuring the accurate monitoring of environmental parameters. The central controller 4 obtains the data of all sensors every 10 seconds through the RS485 bus, forming a temperature and humidity data stream with high time resolution.

[0080] The light sensor uses a quantum type photosynthetically active radiation (PAR) sensor, with a measurement range of 0 - 3000 μmol·m -2 ·s -1 , and the accuracy is ±5 μmol·m -2 ·s -1 . At the same time, a spectral analyzer is equipped to monitor the spectral distribution in the 400 - 700nm band, with a sampling interval of 1 minute.

[0081] The CO2 concentration sensor uses non-dispersive infrared (NDIR) technology, with a measurement range of 0 - 5000 ppm, an accuracy of ±30 ppm, and a response time of less than 30 seconds. The sensors are distributed at different heights in the seedling-raising area to monitor the vertical distribution of CO2 concentration.

[0082] The soil moisture sensor uses a capacitive sensor, with a measurement range of 0 - 100% volumetric water content and an accuracy of ±2%. The sensor is inserted into the seedling-raising substrate to a depth of 2 - 5 cm, and is arranged at a density of one per 1 m 2 in each seedling-raising area.

[0083] The image acquisition system includes a high-definition camera and a multispectral camera. The resolution of the high-definition camera is not less than 4K, and a panoramic image is taken once an hour to monitor the overall growth status of the plants. The multispectral camera includes three bands: visible light, near-infrared, and red edge. Multispectral images are acquired at 10:00, 14:00, and 18:00 every day to analyze the physiological state of the plants. The high-definition camera is installed on the top of the seedling-raising room to cover the entire seedling-raising area; the multispectral camera is installed on a movable robotic arm, which can take close-up pictures of the plants to obtain more detailed image data.

[0084] Data collected by all sensors is transmitted in real time to the data preprocessing unit 21 through the central controller 4 for fusion processing. At the same time, all original data is also saved in the data storage and analysis platform 5 to provide a basis for subsequent analysis and system optimization.

[0085] Preferably, the arrangement of the sensors takes into account the air flow channels and light distribution, avoiding data deviation caused by internal environmental inhomogeneity. The sensors are calibrated regularly to ensure the accuracy and consistency of the data.

[0086] Example 4: Implementation of the multimodal data fusion unit

[0087] This example details the working principle and implementation method of the multimodal data fusion unit 21.

[0088] The multimodal data fusion unit 21 adopts a hybrid fusion architecture combining Kalman filtering and deep belief network to process heterogeneous data from the multi-dimensional environmental perception module 1. The working process of this unit is as Figure 4 shown, including five steps: data preprocessing, feature extraction, temporal alignment, anomaly detection, and multimodal fusion.

[0089] In the data preprocessing stage, first, noise reduction and normalization processing are performed on various types of sensor data. For numerical data such as temperature, humidity, light, and CO2 concentration, median filtering is used to remove outliers, and min-max normalization is performed; for image data, resolution unification, brightness correction, and geometric correction are performed to ensure data quality.

[0090] In the feature extraction stage, key features are extracted from data of different modalities. For numerical sensor data, statistical features such as mean, variance, and rate of change are mainly extracted; for image data, high-dimensional feature vectors including color, texture, and morphological features are extracted through a pre-trained convolutional neural network (such as ResNet-50). This step converts the original high-dimensional heterogeneous data into feature vectors of a unified dimension.

[0091] In the time series alignment stage, the problem of inconsistent sampling frequencies of different sensors is solved. The system uses the dynamic time warping (DTW) algorithm for time series alignment. The basic principle is to minimize the distance between different time series and achieve sequence alignment through adaptive time stretching. For sensor data with high-frequency sampling (such as temperature and humidity every 10 seconds), time window averaging is performed; for data with low-frequency sampling (such as multispectral images 3 times a day), interpolation methods are used to generate estimated values at intermediate times.

[0092] In the anomaly detection stage, a method combining the local outlier factor (LOF) and autoencoder is used to identify and mark outliers in the data stream. For the marked outliers, the system does not directly discard them, but reduces their weights in the fusion process to ensure data continuity.

[0093] Multimodal fusion is the core step of this unit, using a method combining Kalman filtering and deep belief network. Kalman filtering mainly processes continuously changing numerical environmental parameters such as temperature, humidity, CO2 concentration, etc., and continuously optimizes the state estimation through two steps of prediction and update. The deep belief network mainly processes high-dimensional feature data such as image features and constructs multi-layer feature representations. The fusion of the two methods is achieved through an attention mechanism, and the system dynamically adjusts the weights of different data sources according to the current environmental state and plant growth stage.

[0094] The fused data forms a standardized environmental state vector with a dimension of 128, containing comprehensive information of all environmental parameters. At the same time, the system retains the original features of each data source for subsequent analysis and decision-making. The data fusion accuracy is evaluated by comparing with an independent reference sensor, and the average improvement is 43%.

[0095] This hybrid fusion architecture overcomes the limitations of traditional data fusion methods and can effectively process multi-modal, heterogeneous, and environmental data with different sampling frequencies, providing a high-quality data basis for subsequent intelligent analysis and decision-making.

[0096] Example 5: Implementation of a bidirectional recurrent neural network unit

[0097] This example details the structure and working mechanism of the bidirectional recurrent neural network unit 22, which is responsible for constructing a spatio-temporal growth state model of seeds.

[0098] As shown Figure 5 in Figure Figure 5 , the bidirectional recurrent neural network unit 22 is based on the bidirectional long short-term memory network (Bi-LSTM) architecture and has a temporal-spatial double-recursive structure. This network can simultaneously process time series information and spatial distribution information, capturing the short-term fluctuations and long-term trends of plant growth, as well as the spatial distribution characteristics.

[0099] The network input includes three parts: the environmental state vector output by the multi-modal data fusion unit 21, the plant growth image features collected by the image acquisition system 15, and the historical growth data. The temporal modeling range is from 1 hour to 7 days, and the spatial modeling resolution is 5 cubic centimeters.

[0100] The network structure includes an input layer, a feature extraction layer, a temporal processing layer, a spatial processing layer, and an output layer. The input layer receives a 128-dimensional environmental state vector and a 2048-dimensional image feature vector; the feature extraction layer consists of fully connected layers for dimensionality reduction and key feature extraction; the temporal processing layer consists of forward and backward LSTM units, with 128 hidden units in each direction, for capturing time series patterns; the spatial processing layer uses a 3D convolutional structure for processing spatial distribution information; the output layer generates a comprehensive representation of the plant growth state.

[0101] The core advantage of this unit is its ability to simultaneously model the growth states in both the time and space dimensions, providing a more comprehensive characterization of the plant growth process. By using the forward LSTM to capture the influence of historical information on the current state and the backward LSTM to capture the constraints of future possible states on the current state, the bidirectional structure enables the model to fully utilize context information and improve prediction accuracy.

[0102] The model is trained using the mini-batch gradient descent method with a batch size of 64. The initial learning rate is set to 0.001, and a learning rate decay strategy is used. To prevent overfitting, Dropout (with a ratio of 0.3) and L2 regularization (with a coefficient of 0.0001) are introduced. The network is trained on 5000 sets of seedling raising historical data, including growth records under different varieties and environmental conditions. The validation set shows that the prediction accuracy of this model reaches 87.5%, significantly outperforming traditional single-temporal or spatial models.

[0103] This bidirectional recursive structure enables the system to accurately capture the dynamic change laws of plant growth, providing a precise growth state reference for the generation of subsequent environmental control strategies and effectively solving the problem that traditional systems cannot accurately represent the complex growth states of plants.

[0104] Example 6: Implementation of the Deep Reinforcement Learning Unit

[0105] This example details the structure and working mechanism of the deep reinforcement learning unit 23, which is responsible for generating environmental control strategies.

[0106] The deep reinforcement learning unit 23 constructs an environmental regulation decision-making system based on the improved deep Q-network (DQN). As Figure 6 shown, it includes five parts: state space construction, action space definition, reward function design, network structure building, and policy optimization.

[0107] In the state space construction part, the environmental state is represented as a multi-dimensional vector, including environmental parameters such as temperature, humidity, light, CO2 concentration, etc., and the plant growth state representation generated by the bidirectional recurrent neural network unit 22. The dimension of the state vector is 192, including 128-dimensional environmental state and 64-dimensional plant growth state.

[0108] In the action space definition part, the operation instructions of each regulation device are discretized into a finite action set. The actions of the temperature regulation system include heating, cooling, and maintaining, with a step size of 0.5 °C; the actions of the humidity regulation system include humidifying, dehumidifying, and maintaining, with a step size of 2% RH; the actions of the lighting control system include enhancing, weakening, adjusting the spectral ratio, etc., with a total of 15 combinations; the actions of the CO2 supplementation system include increasing, decreasing, and maintaining, with a step size of 50 ppm; the actions of the nutrient solution irrigation system include adjusting the irrigation amount and concentration, with a total of 9 combinations.

[0109] The design of the reward function is one of the innovations of this unit. It innovatively introduces plant physiological indicators as the direct input of the reward function and constructs a multi-objective comprehensive reward function:

[0110] R = w1·R growth + w2·R resource + w3·R stability + w4·R transition ,

[0111] where R growth represents the plant growth reward, which is related to indicators such as growth rate, consistency, and health status; R resource represents the resource utilization reward, which is related to resource consumption such as energy consumption and water consumption; R stability represents the environmental stability reward, which measures the degree of fluctuation of environmental parameters; R transition represents the state transition reward, which is related to the smoothness of control actions; w1 to w4 are weight coefficients, which are dynamically adjusted according to the growth stage and optimization objectives. This multi-objective reward design enables the system to find a balance between promoting plant growth and saving resources.

[0112] In the network structure construction part, a dual-network architecture is adopted, including a policy network and a target network. The two network structures are the same and both consist of 4 fully connected layers. The number of neurons in the hidden layers is 512, 256, and 128 respectively. The activation function is ReLU, and the number of neurons in the output layer is equal to the dimension of the action space. The policy network is used to generate actions, and the target network is used to calculate the target Q value to reduce training instability.

[0113] In the policy optimization part, an improved experience replay mechanism is adopted. A priority experience replay pool with a capacity of 10,000 is designed to preferentially select valuable experiences for learning. The ε-greedy strategy is used to select actions. The ε value starts from 1.0 and gradually decays to 0.1, balancing exploration and exploitation. The network parameters are updated using the Adam optimizer with a learning rate of 0.0001. The target network is updated every 100 steps to enhance training stability.

[0114] This unit continuously optimizes the environmental control strategy through continuous interaction with the virtual environment and the real environment. The experimental results show that the control strategy based on deep reinforcement learning improves the control accuracy by 65% compared with the traditional rule-based control and can adapt to the needs of different seed varieties and growth stages.

[0115] Example 7: Implementation of the plant digital twin model unit

[0116] This example details the structure and working mechanism of the plant digital twin model unit 24, which is responsible for predicting the impact of different combinations of environmental parameters on plant growth in the virtual environment.

[0117] The plant digital twin model unit 24 integrates a plant physiology model and a machine learning model to construct a digital twin of seed growth. As Figure 7 shown, this unit includes four parts: physiological process modeling, environmental response modeling, growth and development simulation, and verification and feedback.

[0118] In the physiological process modeling part, mathematical models are established for the key physiological processes of plants, including photosynthesis, respiration, transpiration, and nutrient absorption. The photosynthesis model is based on the modified Farquhar model, which describes the influence of environmental factors such as light intensity, CO2 concentration, and temperature on the photosynthetic rate:

[0119] A n =min(A c ,A j )-R d ,

[0120] where A n is the net photosynthetic rate; A c is the photosynthetic rate limited by Rubisco enzyme; A j is the photosynthetic rate limited by electron transport; Rd is the dark respiration rate. These parameters are regulated by environmental factors and expressed through mathematical functions.

[0121] The respiration model takes into account maintenance respiration and growth respiration, and uses the Q10 temperature coefficient to describe the effect of temperature on the respiration rate. The transpiration model is based on the Penman-Monteith equation and considers factors such as stomatal conductance, vapor pressure deficit, and light. The nutrient uptake model is based on Michaelis-Menten kinetics and describes the uptake kinetics of different concentrations of nutrient elements.

[0122] In the environmental response modeling part, a response relationship model between environmental factors and plant growth is constructed. Machine learning methods (support vector regression and random forest) are used to train the environmental response model. The input is a combination of environmental parameters, and the output is a plant growth index. This model can predict the impact of environmental changes on plant growth, filling the deficiencies of pure physiological models under complex environmental conditions.

[0123] In the growth and development simulation part, the physiological process model and the environmental response model are integrated to simulate the growth and development process of plants under different environmental conditions. This model can simulate the impact of more than 1000 combinations of environmental parameters on plant growth in a virtual environment, with a prediction time of up to 7 days, providing a reference for the formulation of environmental control strategies. The model update frequency is once an hour, which can reflect the changes in plant status in real time.

[0124] In the verification and feedback part, the model is verified and calibrated through actual measurement data. The system collects plant growth data under actual environmental conditions, compares them with the model prediction results, and calculates the prediction error. According to the error situation, the system automatically adjusts the model parameters to improve the prediction accuracy. Currently, the model prediction error is controlled within the range of ±8%, meeting the actual application requirements.

[0125] This plant digital twin method realizes the transformation from empirical control to predictive control. The system can predict the effects of different strategies before implementing environmental regulation, select the optimal solution, avoid the blindness and lag of traditional systems, and truly achieve precise control centered on plants.

[0126] Example 8: Implementation of the multi-objective optimization unit

[0127] This example details the structure and working mechanism of the multi-objective optimization unit 25, which is responsible for making a balanced decision between maximizing yield and minimizing resource consumption.

[0128] The multi-objective optimization unit 25 uses an improved NSGA-III algorithm to construct a multi-dimensional objective function including yield, energy consumption, water consumption, and fertilizer consumption, and conducts a Pareto optimal solution search. As Figure 8As shown, the unit includes four parts: objective function construction, constraint condition setting, optimization algorithm execution, and decision-making selection.

[0129] In the objective function construction part, four optimization objectives are defined: maximizing yield, minimizing energy consumption, minimizing water resource consumption, and minimizing fertilizer consumption. The yield objective function is based on the biomass growth rate predicted by the plant digital twin model; the energy consumption objective function takes into account the energy consumption of equipment such as temperature control, lighting, and CO2 supplementation; the water resource consumption objective function is related to the irrigation volume and evaporation volume; the fertilizer consumption objective function is related to the nutrient solution concentration and irrigation volume. There are conflicting relationships among these four objectives. For example, increasing yield usually requires increasing resource input. Therefore, multi-objective optimization techniques are needed to find a balance point.

[0130] In the constraint condition setting part, a series of constraint conditions are set according to the physiological requirements of plants and the operating limitations of equipment. These include environmental parameter range constraints (such as temperature 18 - 30°C, relative humidity 40 - 95%, CO2 concentration 400 - 1500 ppm, etc.), equipment operation constraints (such as equipment start-stop frequency limitations, power limitations, etc.), and physiological safety constraints (such as avoiding plant stress caused by sudden environmental changes). These constraints ensure that the optimization results are within the practical and feasible range.

[0131] In the optimization algorithm execution part, an improved NSGA-III algorithm is used for multi-objective optimization. This algorithm maintains population diversity through the reference point mechanism and is particularly suitable for dealing with the 4D objective space in this system. The algorithm parameters are set as follows: population size 100, number of iterations 500, crossover probability 0.9, and mutation probability 0.1. To improve the convergence speed, adaptive crossover and mutation operators are introduced, and the parameters are dynamically adjusted according to the population distribution. In addition, a Kriging surrogate model is used to reduce the demand for computing resources and accelerate the optimization process.

[0132] In the decision-making selection part, the final environmental control strategy is selected from the Pareto optimal solution set. The system uses a TOPSIS-based decision-making method. Combining the current growth stage and operation objectives, weights are assigned to each Pareto solution, and the solution with the highest comprehensive score is selected as the final control strategy. For different growth stages, the system sets different objective weight matrices. For example, in the germination stage, more attention is paid to environmental stability; in the rapid growth stage, more attention is paid to yield; and in the hardening-off stage, more attention is paid to plant quality.

[0133] Through this multi-objective optimization method, the system can find the best balance point between yield and resource consumption and achieve sustainable development. Experimental results show that compared with traditional single-objective optimization, multi-objective optimization improves the comprehensive resource utilization efficiency by 38.5% while maintaining a relatively high yield level.

[0134] Example 9: Implementation of the Precise Environment Regulation Module

[0135] This embodiment details the composition and working principle of the precise environment control module 3, which is responsible for implementing the environment control strategy generated by the intelligent analysis and decision-making module 2.

[0136] The precise environment control module 3 includes a temperature regulation system, a humidity regulation system, a lighting control system, a CO2 supplementation system, and a nutrient solution irrigation system. These systems convert control instructions into actual environmental parameter adjustments through actuators.

[0137] The temperature regulation system adopts a zoning control strategy, including floor radiant heating, air heat pumps, and top cooling equipment. The system achieves precise temperature regulation through the PID control algorithm, with a control accuracy of ±0.5°C. To reduce energy consumption, the system preferentially utilizes natural heating and passive heat dissipation and only activates active control equipment when necessary. Based on the temperature distribution data provided by the multi-dimensional environment perception module 1, the system can identify temperature non-uniform areas and perform targeted adjustments to ensure the temperature uniformity of the entire seedling-growing area.

[0138] The humidity regulation system includes an ultrasonic humidifier, a dehumidifier, and ventilation equipment. The system automatically selects the humidification or dehumidification mode based on the target value and current value of the relative humidity. To prevent stress reactions caused by rapid humidity changes to plants, the system sets a limit on the humidity change rate, generally controlled within 5% RH / hour. In addition, the system also considers the interaction between temperature and humidity and synchronously considers humidity changes during temperature regulation to achieve coordinated control of temperature and humidity.

[0139] The lighting control system includes full-spectrum LED lights and mobile supplementary lighting vehicles. The full-spectrum LED lights include four primary color LEDs: red (630 - 660nm), blue (450 - 470nm), green (510 - 530nm), and far-red (730 - 740nm), which can simulate different spectral ratios. The central controller 4 adjusts the power ratio of each primary color LED according to the growth stage of the seeds and the control strategy output by the deep reinforcement learning unit 23. Stepless dimming from 0 - 100% is achieved through PWM dimming technology, with a dimming accuracy of 1%. In addition, the system is also equipped with mobile supplementary lighting vehicles, which can perform directional supplementary lighting on uneven growth areas according to the plant growth distribution map provided by the image acquisition system 15 to ensure the consistency of plant growth.

[0140] The CO2 supplementation system includes a CO2 storage tank, a flow control valve, and a distributed injection pipeline. The system accurately controls the release amount of CO2 based on the feedback from the CO2 concentration sensor 13, maintaining the CO2 concentration in the seedling-raising environment within the target range. To improve the CO2 utilization efficiency, the system increases the CO2 concentration when the light is sufficient to promote photosynthesis, and appropriately reduces the CO2 concentration when the light is insufficient to save resources. In addition, the system also takes into account the impact of air flow on the CO2 distribution, and ensures the uniform distribution of CO2 concentration by optimizing the injection position and ventilation strategy.

[0141] The nutrient solution irrigation system adopts a fuzzy control algorithm and automatically determines the irrigation strategy according to soil moisture, environmental factors, and growth stages. The system first obtains input variables such as soil moisture, air temperature, humidity, light intensity, and the growth stage of seeds. Secondly, it performs fuzzy processing on these input variables and converts them into fuzzy sets. Thirdly, it performs reasoning according to the preset fuzzy rule base to obtain the fuzzy outputs of the irrigation amount and nutrient solution concentration. Then, it performs defuzzification processing on the fuzzy outputs to obtain specific irrigation instructions. Finally, it sends the irrigation instructions to the irrigation execution device to achieve precise irrigation. The system adopts drip irrigation technology and conducts quantitative irrigation according to the actual water demand of plants, avoiding problems such as water resource waste and root hypoxia caused by over-irrigation.

[0142] All regulation devices are coordinated and controlled through the central controller 4 to ensure the coordinated adjustment of various environmental factors. To improve the energy utilization efficiency, the system adopts a time-sharing and zoning control strategy, flexibly adjusting the environmental parameters of different regions according to actual needs, and avoiding resource waste caused by unified control of the entire region.

[0143] Example 10: Control Method of the Seedling-Raising Artificial Climate Chamber Control System

[0144] This example details the control method of the seedling-raising artificial climate chamber control system, as Figure 3 shown, this method includes the following steps:

[0145] Step S1: Multidimensional parameter data of the seedling-raising environment are collected in real time through the multidimensional environment perception module 1. Specifically, it includes: the temperature and humidity sensor array collects temperature and humidity data, the light sensor collects light intensity and spectral distribution data, the CO2 concentration sensor 13 collects gas composition data, the soil moisture sensor 14 collects substrate water content data, and the image acquisition system 15 collects images of the plant growth state.

[0146] Step S2: The multi-modal data fusion unit 21 performs heterogeneous fusion processing on the multi-dimensional parameter data. First, noise reduction and normalization processing are performed on various types of sensor data; secondly, the key features of different modal data are extracted; thirdly, time series alignment is performed to solve the problem of inconsistent sampling frequencies of different sensors; then anomaly detection is carried out to identify and mark the outliers in the data stream; finally, multi-modal fusion is carried out by combining Kalman filtering and deep belief network to generate a standardized environmental state vector.

[0147] Step S3: The bidirectional recurrent neural network unit 22 is used to establish a seed spatio-temporal growth state model. This unit is based on the bidirectional long short-term memory network architecture and has a double recurrence structure of time series - space, which can capture both the short-term fluctuations and long-term trends of plant growth, as well as the spatial distribution characteristics, to establish an accurate plant growth state model.

[0148] Step S4: The deep reinforcement learning unit 23 generates an environmental control strategy. This unit constructs an environmental regulation decision-making system based on an improved deep Q network. By constructing a multi-dimensional state space, defining an action space, designing a multi-objective reward function, building a network structure, and optimizing the strategy, an optimal environmental control strategy is generated.

[0149] Step S5: The plant digital twin model unit 24 simulates and predicts the growth effects of different combinations of environmental parameters. This unit integrates a plant physiology model and a machine learning model to simulate the effects of different combinations of environmental parameters on plant growth in a virtual environment, providing a predictive reference for the formulation of environmental control strategies.

[0150] Step S6: The multi-objective optimization unit 25 makes a balance decision between yield and resource consumption to optimize the environmental control strategy. This unit uses an improved NSGA-III algorithm to construct a multi-dimensional objective function including yield, energy consumption, water consumption, and fertilizer consumption, and performs a Pareto optimal solution search to achieve a comprehensive decision-making of maximizing yield, minimizing resource consumption, and minimizing environmental impact.

[0151] Step S7: The precise environmental regulation module 3 executes the optimized environmental control strategy. The temperature regulation system adjusts the temperature, the humidity regulation system adjusts the humidity, the light control system adjusts the light intensity and spectrum, the CO2 supplementation system adjusts the CO2 concentration, and the nutrient solution irrigation system performs precise irrigation to jointly construct the most suitable environment for seed growth.

[0152] Step S8: According to the actual growth effect, update the plant digital twin model and the experience replay pool to achieve continuous optimization of the control system. The system collects plant growth data under actual environmental conditions, compares it with the model prediction results, adjusts the model parameters to improve the prediction accuracy; at the same time, updates the data in the experience replay pool to provide better learning samples for deep reinforcement learning, enabling the system to continuously self-optimize.

[0153] Through the above steps, the present invention realizes the intelligent perception, analysis and decision-making, and precise regulation of the seedling raising environment, forming a complete closed-loop control system. This method can not only improve the seed emergence rate and seedling quality, but also optimize the resource utilization efficiency, achieving the unity of economic benefits and ecological benefits.

[0154] Example 11: Application Example and Effect Verification

[0155] In this example, the technical effect of the present invention is verified through an actual application case.

[0156] In a 300-square-meter artificial climate chamber for seedling raising, the control system of the present invention is installed. To comprehensively evaluate the system performance, a comparative experiment is designed, including four groups of treatments: traditional field seedling raising (T1), ordinary greenhouse seedling raising (T2), conventional artificial climate chamber seedling raising (T3), and the seedling raising with the system of the present invention (T4). Each group has 3 replicates, with a total of 12 experimental plots. The same batch of seeds is selected and sown and initially treated according to a unified standard.

[0157] The seedling raising process lasts for 37 days, divided into a germination stage (0 - 7 days), an initial growth stage (8 - 15 days), a rapid growth stage (16 - 30 days), and an acclimatization stage (31 - 37 days). The system of the present invention automatically adjusts the environmental parameters according to the requirements of different stages. For example, in the germination stage, the temperature is maintained at 24 ± 0.5 °C / 20 ± 0.5 °C (day / night), the relative humidity is 95 ± 2% RH, the CO2 concentration is 400 - 500 ppm, and only weak light is turned on when checking the seedling situation; in the rapid growth stage, the temperature is adjusted to 28 ± 0.5 °C / 24 ± 0.5 °C, the relative humidity is 75 ± 5% RH, the PPFD is 0 - 400 μmol·m-2·s-1, the photoperiod is 18 h / 6 h, and the CO2 concentration is 1000 - 1200 ppm.

[0158] To evaluate the system performance, the following key indicators are measured: seed emergence rate, seedling age consistency (coefficient of variation of plant height), root-shoot ratio, chlorophyll content (SPAD value), stem strength, water use efficiency, and energy consumption. The experimental results are shown in Table 1:

[0159] Table 1 Performance Comparison of Different Seedling Raising Methods

[0160] Index T1 (Field) T2 (Common greenhouse) T3 (Conventional climate chamber) T4 (This invention) Emergence rate (%) 75 83 87 92 Seedling age consistency (CV%) 18 15 11 8 Root-shoot ratio 0.35 0.38 0.43 0.48 Chlorophyll content (SPAD) 32 35 38 41 Stem strength (N) 1.2 1.4 1.6 1.8 Water use efficiency (g / L) 3.5 4 4.7 5.2 <![CDATA[Energy consumption (kWh / m 2 )]]> - 8 14 10

[0161] It can be seen from the experimental results in Table 1 that the system of the present invention performs best in all indicators:

[0162] 1. The seed emergence rate reaches 92%, which is 17 percentage points higher than that of traditional field seedling raising and 5 percentage points higher than that of conventional artificial climate chamber seedling raising, significantly reducing seed waste.

[0163] 2. The consistency of seedling age is significantly improved, with a coefficient of variation of only 8%, which is 55.6% lower than the traditional method, laying a foundation for subsequent field management.

[0164] 3. The root-shoot ratio reaches 0.48, which is 37.1% higher than the traditional method, indicating better root development and enhancing the stress resistance and growth potential of seedlings.

[0165] 4. The chlorophyll content (SPAD value of 41) and stem strength (1.8 N) are both the highest, increasing by 28.1% and 50% respectively, laying a foundation for high yield.

[0166] 5. The water use efficiency reaches 5.2 g / L, which is 48.6% higher than the traditional method, significantly saving water resources.

[0167] 6. The energy consumption is 10 kWh / m 2 , which is 28.6% lower than that of the conventional artificial climate chamber, achieving a win-win situation in economic and ecological benefits.

[0168] It is particularly worth noting that the advantages of the system of the present invention in terms of resource utilization efficiency are particularly obvious. Although the conventional artificial climate chamber can provide stable environmental conditions, due to the lack of intelligent optimization strategies, the energy consumption is relatively high; while the system of the present invention, through multi-objective optimization and predictive control, while providing better environmental conditions, reduces the energy consumption and realizes efficient and sustainable development.

[0169] In addition, the self-learning and self-adaptive capabilities of the system of the present invention have also been verified. In the later stage of seedling cultivation, the system gradually optimized the control strategy, and both the control accuracy and resource utilization efficiency have been improved. For example, in the operation of the last 10 days, the energy consumption of the system was 12% lower than that in the initial stage, and the environmental control accuracy was increased by 8%. This shows that the system can continuously learn from the operation experience and continuously self-optimize.

[0170] In summary, the control system and control method of the artificial climate chamber for seedling cultivation of the present invention, through multi-dimensional environmental perception, intelligent analysis and decision-making, and precise environmental regulation, realize the efficient and precise management of the seedling cultivation process, significantly improve the seedling cultivation quality and resource utilization efficiency, and have good application prospects and promotion value.

[0171] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A control system for an artificial climate chamber for seedling raising, characterized in that, Including: A multi-dimensional environment perception module, an intelligent analysis and decision-making module, a precise environment regulation module, a central controller, and a data storage and analysis platform; among them, The multi-dimensional environment perception module includes a temperature and humidity sensor array, a light sensor, a CO2 concentration sensor, a soil moisture sensor, and an image acquisition system, which are used to collect multi-dimensional parameter data of the seedling raising environment in real time; The intelligent analysis and decision-making module is connected to the multi-dimensional environment perception module and includes a multi-modal data fusion unit, a bidirectional recurrent neural network unit, a deep reinforcement learning unit, a plant digital twin model unit, and a multi-objective optimization unit. Among them, the multi-modal data fusion unit is used to uniformly preprocess heterogeneous environment data; the bidirectional recurrent neural network unit is used to establish a spatio-temporal growth state model of seeds; the deep reinforcement learning unit is used to generate an environment control strategy; the plant digital twin model unit is used to simulate and predict the impact of different environmental parameter combinations on plant growth; the multi-objective optimization unit is used to make a balance decision between maximizing yield and minimizing resource consumption; The precise environment regulation module is connected to the intelligent analysis and decision-making module and includes a temperature regulation system, a humidity regulation system, a light control system, a CO2 supplementation system, and a nutrient solution irrigation system, which are used to execute the environment control strategy; The central controller is used to coordinate the work of each module and perform human-computer interaction; The data storage and analysis platform is used to store environment parameters and growth data and support data cloud analysis.

2. The seedling raising artificial climate chamber control system according to claim 1, characterized in that, The multi-modal data fusion unit adopts a data fusion algorithm combining Kalman filtering and deep belief network to fuse the numerical data from the temperature and humidity sensor array, light sensor, CO2 concentration sensor, and soil moisture sensor, as well as the image data from the image acquisition system, to solve the problems of different sampling frequencies and noise interference of different modal data, and generate a standardized environment state vector.

3. The seedling raising artificial climate chamber control system according to claim 1, characterized in that, The bidirectional recurrent neural network unit constructs a bidirectional long short-term memory network containing 128 hidden units, with a temporal-spatial double-recurrent structure, which can capture both short-term fluctuations and long-term trends of plant growth simultaneously. The temporal modeling range is from 1 hour to 7 days, the spatial modeling resolution is 5 cubic centimeters, and the prediction accuracy reaches 87.5%.

4. The seedling cultivation artificial climate chamber control system according to claim 1, characterized in that The deep reinforcement learning unit constructs an environment regulation decision-making system based on an improved deep Q network, constructs a multi-dimensional state space including temperature, humidity, light, and CO2 concentration, innovatively introduces plant physiological indicators as direct inputs to the reward function, and uses an experience replay pool technology to store and optimize past decision-making experiences to achieve continuous optimization.

5. The control system of the artificial climate chamber for seedling cultivation according to claim 1, wherein The plant digital twin model unit integrates a plant physiology model and a machine learning model, can simulate the impact of more than 1000 environmental parameter combinations on plant growth in a virtual environment, mathematically model the key physiological processes of photosynthesis, transpiration, and nutrient absorption, with a prediction time of up to 7 days, and the prediction error is controlled within the range of ±8%.

6. The seedling cultivation artificial climate chamber control system according to claim 1, characterized in that, The multi-objective optimization unit adopts an improved NSGA-III algorithm to construct a multi-dimensional objective function including yield, energy consumption, water consumption, and fertilizer consumption, search for Pareto optimal solutions, achieve comprehensive decision-making for maximizing yield, minimizing resource consumption, and minimizing environmental impact, and improve the comprehensive resource utilization efficiency by 38.5%.

7. The seedling raising artificial climate chamber control system according to claim 1, characterized in that, The image acquisition system includes a high-definition camera and a multi-spectral camera. The resolution of the high-definition camera is not less than 4K, and a panoramic image is taken once an hour. The multi-spectral camera includes three bands: visible light, near-infrared, and red edge, and multi-spectral images are collected at 10:00, 14:00, and 18:00 every day to capture visible changes and non-visible physiological characteristics of the plant growth state.

8. The seedling raising artificial climate chamber control system according to claim 1, characterized in that, The lighting control system includes full-spectrum LED lights and a mobile supplementary lighting vehicle. The full-spectrum LED lights include four primary color LEDs: red, blue, green, and far-red, and achieve stepless dimming from 0 to 100% through PWM dimming technology, with a dimming accuracy of 1%. The mobile supplementary lighting vehicle can perform directional supplementary lighting on local areas according to the plant growth distribution map provided by the image acquisition system.

9. The seedling cultivation artificial climate chamber control system according to claim 1, characterized in that, The central controller is connected to the multi-dimensional environment perception module, the intelligent analysis and decision-making module, and the precise environment regulation module, and includes an industrial-grade PLC, a human-machine interface, and a remote monitoring module. Among them, the industrial-grade PLC is responsible for the overall coordination and real-time control of the system; the human-machine interface is a touch screen display for parameter setting and status monitoring; the remote monitoring module supports real-time viewing and controlling the operation status of the system through a mobile APP.

10. A control method for a control system of a seedling artificial climate chamber according to any one of claims 1-9, characterized in that, It includes the following steps: 1) Real-time collect multi-dimensional parameter data of the seedling raising environment through the multi-dimensional environment perception module; 2) Perform heterogeneous fusion processing on the multi-dimensional parameter data through the multi-modal data fusion unit; 3) Establish a seed spatio-temporal growth state model through the bidirectional recurrent neural network unit; 4) Generate an environment control strategy through the deep reinforcement learning unit; 5) Simulate and predict the growth effects of different environmental parameter combinations through the plant digital twin model unit; 6) Make a balance decision between yield and resource consumption through the multi-objective optimization unit and optimize the environment control strategy; 7) Execute the optimized environment control strategy through the precise environment regulation module; 8) Update the plant digital twin model and the experience replay pool to achieve continuous optimization of the control system.

Images