Intelligent grafting cultivation method of cocktail grapefruit

Abstract

The application discloses an intelligent grafting cultivation method for cocktail grapefruit, and belongs to the technical field of intelligent cultivation of fruit trees. The method constructs a three-dimensional simulation model of a seedling raising shed through computational fluid mechanics, couples meteorological time series data and fluid topology optimization, dynamically optimizes the distribution of air flow field and temperature and humidity field, and cultivates standardized rootstocks. Distributed sensing nodes are used to collect rhizosphere soil and canopy physiological multi-modal data in real time, and a joint space-time feature tensor is generated through cross-modal feature fusion. Deep reinforcement learning is used to realize precise quantitative dripping of growth-promoting bacteria, match the root growth needs of grafting seedlings in the healing period and the slow-growing period, introduce transfer learning, federal learning and adversarial generation verification to complete stress resistance domestication, eliminate regional microclimate differences and data privacy risks, and finally output standardized grafting seedlings with uniform growth and strong stress resistance. The application solves the problems of extensive cultivation environment regulation and control, experience-based fertilization, and mismatch between environment and plant needs, and improves seedling consistency and survival rate.

Description

Technical Field

[0001] This invention relates to the field of fruit tree grafting cultivation technology, specifically to an intelligent grafting cultivation method for grapefruit, applicable to the fields of facility seedling cultivation, precise water and fertilizer regulation, stress resistance domestication, and standardized seedling propagation. Background Technology

[0002] Currently, grafting cultivation of Cocktail grapefruit is typically carried out in conventional greenhouses. During the rootstock cultivation stage, the greenhouse is equipped with ventilation openings, heaters, and irrigation equipment. Operators set fixed upper and lower temperature thresholds and lower humidity thresholds based on historical experience. When sensors detect that environmental parameters exceed the thresholds, the corresponding equipment is activated or deactivated according to a preset program. After grafting and planting, growers apply a fixed amount of growth-promoting microbial fertilizer to the rhizosphere soil according to a fixed calendar cycle. At the same time, they visually observe changes in the morphology and wilting degree of the canopy leaves and, based on experience, determine whether it is necessary to adjust the greenhouse ventilation or implement water stress training.

[0003] The aforementioned conventional cultivation methods rely on fixed threshold controls and human experience-based judgment. A core technical problem with these methods is the mismatch between environmental regulation operations and the dynamic physiological state of the plants in the spatiotemporal dimensions. Fixed thresholds cannot reflect the spatial differences in airflow and temperature / humidity fields within the greenhouse caused by equipment layout, resulting in a non-uniform growth environment for the rootstock. Applying growth-promoting microbial fertilizers at fixed intervals ignores the real-time changes in the root microecology of grafted seedlings during the healing and recovery periods. Adjusting ventilation and water stress based on experience lacks quantitative evidence regarding the synergistic relationship between the rhizosphere microbial community and canopy photosynthesis and transpiration, leading to a disconnect between the environmental parameter adjustment steps and the actual needs of the plants at specific growth stages. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent grafting cultivation method for Cocktail Grapefruit, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A smart grafting cultivation method for Cocktail Grapefruit includes:

[0007] A three-dimensional simulation model of the airflow and temperature and humidity fields inside the seedling shed was constructed using computational fluid dynamics methods. Based on the simulation output results, the operating parameters of the equipment were adjusted to cultivate standardized rootstocks.

[0008] The standardized rootstock was grafted onto a grapefruit scion to obtain grafted seedlings in the healing period.

[0009] The grafted seedlings were planted in cultivation containers with built-in distributed sensing nodes to collect real-time data on the pH, EC, moisture content, and microbial community abundance of the rhizosphere soil, as well as the photosynthetically active radiation and transpiration rate of the canopy.

[0010] Based on rhizosphere environment data, the quantitative dripping of growth-promoting bacteria is precisely triggered through a targeted regulation program to match the root growth needs during the healing and seedling establishment periods.

[0011] Based on a pre-trained stress resistance acclimatization model, combined with real-time data collected from the canopy and rhizosphere, environmental stress parameters are adjusted in stages to complete stress resistance acclimatization and output standardized grafted seedlings.

[0012] Preferably, when constructing the three-dimensional simulation model, a fluid topology optimization method that couples meteorological time series data with computational fluid dynamics is introduced. The layout of ventilation openings and temperature control equipment in the seedling shed is used as topological variables, and the wind speed uniformity and temperature gradient variance in the rootstock canopy area are used as constraints. The topological variables are updated by iteratively solving the adjoint equation, and the spatial optimal distribution parameters of the airflow field and temperature and humidity field in the seedling shed are dynamically optimized.

[0013] Preferably, after collecting rhizosphere and canopy data, a multimodal data alignment and feature fusion method is used to map rhizosphere chemical and biological data into a first-modal feature vector and canopy physiological data into a second-modal feature vector. The correlation matrix of the two feature vectors in the time dimension is calculated using a cross-modal attention mechanism, and feature-level fusion is performed to generate a joint spatiotemporal feature tensor representing the overall growth status of the grafted seedling.

[0014] Preferably, when triggering the quantitative dripping of growth-promoting bacteria, a deep reinforcement learning resource scheduling method is introduced. The joint spatiotemporal feature tensor is used as the state space, and the dripping amount and frequency of the growth-promoting bacteria are used as the action space. A delayed reward function is constructed with the root biomass increase of the grafted seedling within a future set time period as the objective. The optimal dripping action sequence is output by updating the strategy network through interaction with the environment.

[0015] Preferably, when adjusting environmental stress parameters in stages, a graph neural network construction method is adopted to map the species interaction relationships of the rhizosphere microbial community and the canopy stress response characterization set into a dynamic graph structure. Microbial abundance and physiological indicators are used as node attributes, and material exchange and signal transmission relationships are used as edge attributes. The spatiotemporal evolution characteristics of the dynamic graph structure are extracted through a message passing mechanism to guide the adjustment of stress parameters.

[0016] Preferably, when extracting spatiotemporal evolution features, the graph attention mechanism and the temporal convolutional network sequence prediction method are combined. The graph attention mechanism assigns adaptive weights to nodes in the dynamic graph structure to focus on the core growth-promoting bacterial community. The weighted node features are input into the temporal convolutional network to capture the local temporal dependence of the material flow conduction efficiency at the rootstock-sprout healing interface and output the predicted sequence of the conduction efficiency at the healing interface.

[0017] Preferably, the predicted sequence based on the conduction efficiency of the healing interface adopts an adaptive multi-objective optimization algorithm with the dual objective function of maximizing the peak conduction efficiency in the predicted sequence and minimizing the total consumption of probiotics output by the deep reinforcement learning resource scheduling method. The Pareto front solution set is obtained by non-dominated sorting and crowding calculation, and the triggering strategy of the probiotic-targeting regulation program is dynamically modified from the solution set.

[0018] Preferably, before implementing stress resistance training, transfer learning and few-shot learning strategies are introduced to extract model parameters accumulated during multi-objective optimization and graph neural network training of grafted seedlings of different citrus varieties in history as prior knowledge of the source domain. The divergence between the source domain and the current distribution of grapefruit data is minimized through domain adaptation methods, and the hyperparameters of the stress resistance training model are quickly fine-tuned with a small sample size.

[0019] Preferably, when fine-tuning the stress resistance domestication model, a privacy protection and collaborative training method under the federated learning framework is introduced. The model gradient is calculated locally on multiple distributed cultivation container nodes using desensitized perception data. Only the encrypted gradient parameters are uploaded to the central server for aggregation. This eliminates the model bias caused by different regional microclimates without sharing the original rhizosphere and canopy multimodal data.

[0020] Preferably, when aggregating gradient parameters, an adversarial generation and verification method is introduced. A generator and a discriminator are constructed in the federated learning framework. The generator is used to fit the true distribution of gradient parameters uploaded by each node, and the discriminator is used to evaluate the consistency between the aggregated gradient and the local true gradient distribution. The gradient noise introduced by abnormal nodes is eliminated through adversarial training, so as to ensure the robustness of the output standardized grafted seedling quality evaluation index.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. By introducing computational fluid dynamics and fluid topology optimization methods, the layout of ventilation openings and temperature control equipment is used as topological variables for iterative solution. With wind speed uniformity and temperature gradient variance as constraints, the limitations of fixed threshold adjustment are broken, and the spatially optimal distribution parameters matching the rootstock canopy area are obtained, realizing the spatial matching optimization of the rootstock cultivation environment. After the grafted seedlings are planted, the data of rhizosphere pH, conductivity, water content, and microbial community abundance are mapped with the data of canopy photosynthetically active radiation and transpiration rate as multimodal feature vectors. A joint spatiotemporal feature tensor is generated through a cross-modal attention mechanism. A deep reinforcement learning resource scheduling method is used to construct a delayed reward function by using the dripping amount and frequency of growth-promoting bacteria as the action space. This allows the dripping action sequence of growth-promoting bacteria to be adjusted according to the dynamic growth needs of the root system during the healing and seedling establishment periods, changing the extensive mode of fixed-cycle fertilization.

[0023] 2. A graph neural network was used to map the rhizosphere microbial species interactions and canopy stress response into a dynamic graph structure. Combining graph attention mechanisms and temporal convolutional networks, adaptive weights were assigned to focus on core growth-promoting bacterial communities, extracting the local temporal dependence of material flow conduction efficiency at the rootstock-scion healing interface. Based on this predicted sequence, an adaptive multi-objective optimization algorithm was used to construct a dual-objective function with peak conduction efficiency and total consumption of growth-promoting bacterial agents to obtain the Pareto front solution set. This modified the triggering strategy of the growth-promoting bacterial targeted regulation program, maintaining healing efficiency while limiting resource consumption. During the stress resistance acclimatization stage, transfer learning and few-shot learning strategies were used to eliminate data distribution differences between historical and current citrus varieties. A privacy-preserving collaborative training method under a federated learning framework was used to aggregate encrypted gradient parameters computed locally by multiple nodes. A generator and discriminator were introduced for adversarial generation verification to eliminate abnormal gradient noise, ensuring the robustness of the stress resistance acclimatization model to evaluation indicators under multi-regional microclimate conditions. Attached Figure Description

[0024] Figure 1 This is a diagram illustrating the standardized rootstock cultivation method of the present invention.

[0025] Figure 2 This is a diagram illustrating the real-time acquisition of multimodal data according to the present invention.

[0026] Figure 3 This is a diagram illustrating the dynamic graph neural network and spatiotemporal feature extraction of the present invention.

[0027] Figure 4 This is a diagram illustrating the collaborative training of resilience domestication and federated learning in this invention. Detailed Implementation

[0028] Please refer to the attached document. Figure 1 A three-dimensional simulation model of the airflow and temperature / humidity fields inside a seedling shed is constructed using computational fluid dynamics (CFD) methods. Based on the simulation output, equipment operating parameters are adjusted to cultivate standardized rootstocks. Specifically, in this embodiment, the geometric model of the seedling shed is obtained through three-dimensional laser scanning. The computational domain of the constructed three-dimensional simulation model includes the entire interior space of the seedling shed and the air domain within a 10m radius outside the shed, to eliminate the influence of boundary effects on the simulation results. The governing equations of the three-dimensional simulation model adopt the Reynolds-averaged Navier-Stokes equations, and the turbulence model is selected from... The two-equation model uses a component transport equation to describe the water vapor transport process and an energy equation to describe the spatiotemporal evolution of the temperature field. The Reynolds-averaged Navier-Stokes equations are expressed as follows:

[0029]

[0030] in, air density, , In Cartesian coordinate system , The time-averaged velocity component in the direction, For time, Average pressure over time Aerodynamic viscosity, For the Kronecker function, , for , The directional pulsating velocity component, For Reynolds stress term, for The component of gravitational acceleration in the direction of gravity.

[0031] In the k-ε two-equation model, turbulent kinetic energy With dissipation rate The transport equations are as follows:

[0032]

[0033] in, For turbulent viscosity, ; The turbulent kinetic energy generated by the average velocity gradient The turbulent kinetic energy generated by buoyancy, The contribution of fluctuating expansion in compressible turbulence to the total dissipation rate; , , , These are model constants. , They are respectively and The Prandtl number, in this embodiment, uses the standard The model uses default constant values. The boundary conditions for the simulation are set as follows: the external entrance boundary uses local real-time meteorological data, including ambient temperature, relative humidity, wind speed, and wind direction; the ventilation outlet boundary is set as a pressure outlet boundary; the temperature control equipment and evaporative cooling unit boundary is set as a velocity inlet boundary; and the wall boundary uses a no-slip wall condition. After solving the simulation model, the airflow field and temperature / humidity field distribution data of the rootstock canopy area are extracted. With a wind speed deviation of less than ±0.2 m / s and a temperature deviation of less than ±0.5℃ in the canopy area as control targets, the equipment operating parameters—ventilation outlet opening, temperature control unit power, and evaporative cooling unit operating frequency—are adjusted to ensure that the environmental parameters of the rootstock growth area in the seedling shed meet the standardized cultivation requirements. In this embodiment, trifoliate orange rootstock is selected, and the cultivation period is 90 days. During the cultivation process, the meteorological data is updated every 24 hours, and the simulation is re-executed to dynamically adjust the equipment operating parameters, obtaining standardized rootstocks with uniform growth.

[0034] Select standardized trifoliate orange rootstocks with uniform growth, and graft the grapefruit scion onto the main trunk of the rootstock using the cleft grafting method. After aligning the cambium layers of the rootstock and scion, wrap them tightly with grafting film for fixation. After grafting, place the plants in a high-humidity, low-light healing environment in a seedling shed, controlling the ambient temperature at 25-28℃ and the relative humidity above 90%. After 7-10 days of care, when callus tissue forms at the graft union and the rootstock and scion have initially healed, the grafted seedlings in the healing stage are obtained.

[0035] Secondly, the grafted seedlings were planted in cultivation containers with built-in distributed sensing nodes to collect real-time data on the pH, EC, moisture content, and microbial community abundance of the rhizosphere soil, as well as the photosynthetically active radiation and transpiration rate of the canopy. Specifically, in this embodiment, the grafted seedlings were selected from the healing period of the standardized trifoliate orange rootstock grafted with grapefruit scions. The cultivation container was a root-limited cultivation container, with three sets of rhizosphere sensing nodes arranged vertically at 10cm intervals inside the container, corresponding to the growth areas of the surface roots, middle roots, and deep roots, respectively. Each set of rhizosphere sensing nodes simultaneously collected data on the pH, EC, volumetric moisture content, and microbial community abundance of the corresponding root layer soil. Among them, the microbial community abundance data was obtained by fluorescence quantitative PCR detection using a microbial abundance detection chip. The detection targets included the core growth-promoting bacteria of Bacillus, Pseudomonas, and Trichoderma, and the abundance data was measured in CFU / g per gram of dry soil. The photosynthetically active radiation sensor of the canopy physiological monitoring unit is placed at the top of the grafted seedling canopy to collect photosynthetically active radiation values ​​in the 400-700nm wavelength range, in units of... ; Transpiration rate data were obtained by converting stem flow rates collected by a stem flow meter. The conversion formula is as follows:

[0036]

[0037] in, The transpiration rate of the plant is expressed in units of t / v. ; The stem flow rate is expressed in g / h. The total leaf area of ​​the plant is expressed in m², which is obtained by converting the leaf area index into the canopy projection area.

[0038] In this embodiment, the data acquisition frequency of the rhizosphere sensing node is once per hour, and the data acquisition frequency of the canopy physiological monitoring unit is once every 30 minutes. All acquired data is uploaded to the central processing subsystem via the LoRa wireless communication protocol to complete data caching and preprocessing. Preprocessing includes outlier removal and missing value interpolation. Outliers are... Criteria for elimination; missing values ​​are filled in using linear interpolation. Please refer to [reference needed]. Figure 2 .

[0039] Furthermore, based on rhizosphere environmental data, the quantitative dripping of growth-promoting agents is precisely triggered through a growth-promoting bacteria-targeted regulation program to match the root growth needs during the healing and seedling establishment periods. Specifically, in this embodiment, the growth-promoting bacteria-targeted regulation program is pre-stored in the computing unit of the central processing subsystem. The healing period is 0-14 days after grafting, and the seedling establishment period is 15-30 days after grafting. For the healing period, the root growth needs are wound healing and new root germination, with preset rhizosphere environmental regulation targets for the corresponding growth stage. For the seedling establishment period, the root growth needs are root expansion and root hair formation, with preset rhizosphere environmental regulation targets for the corresponding growth stage. The growth-promoting bacteria-targeted regulation program acquires pre-processed rhizosphere environmental data in real time and compares the real-time data with the regulation targets for the corresponding growth stage. When the abundance of the core growth-promoting bacteria community detected in real time is lower than the lower limit of the regulation target, or the pH / EC value exceeds the regulation target range, the dripping action of the growth-promoting agent is triggered. The dripping volume is calculated using a linear proportional-integral-derivative control algorithm, and the calculation formula is:

[0040]

[0041] in, for The volume of the growth-promoting agent dripped at any given time, in liters (L). This is the proportionality coefficient. The integral time constant is... As the differential time constant, corresponding control parameters are set for the healing period and the seedling establishment period in this embodiment; for Timing deviation, , The target value for the corresponding regulatory indicator. This refers to the real-time detection value of the corresponding control index. In this embodiment, the growth-promoting bacterial agent is a compound bacterial agent containing Bacillus, Pseudomonas, and Trichoderma, with an effective viable count ≥2×10^9 CFU / mL; the dripping action is performed by a quantitative peristaltic pump, and the maximum volume of a single drip does not exceed 80% of the water holding capacity of the substrate in the cultivation container, and the dripping frequency does not exceed once every 12 hours to avoid root hypoxia caused by water accumulation in the rhizosphere.

[0042] Finally, based on the pre-trained stress resistance acclimatization model and combined with real-time data collected from the canopy and rhizosphere, environmental stress parameters were adjusted in stages to complete the stress resistance acclimatization and output standardized grafted seedlings.

[0043] Specifically, in this embodiment, the stress resistance acclimatization model is a pre-trained backpropagation neural network model. The input layer parameters include rhizosphere pH, EC value, water content, microbial community abundance, canopy photosynthetically active radiation, and transpiration rate. The output layer parameters are environmental stress parameters, including daytime temperature, nighttime temperature, relative humidity, photosynthetically active radiation supplementation, and substrate water content threshold in the seedling shed. The stress resistance acclimatization is divided into three stages: the first stage is the acclimatization adaptation period, corresponding to 31-45 days after grafting, with a stress intensity of mild stress; the second stage is the acclimatization strengthening period, corresponding to 46-60 days after grafting, with a stress intensity of moderate stress; and the third stage is the acclimatization stabilization period, corresponding to 61-75 days after grafting, with the stress intensity gradually decreasing to the normal level. The pre-trained stress resistance acclimatization model has preset parameter constraint ranges for each acclimatization stage. The model acquires real-time data from the canopy and rhizosphere, outputs environmental stress parameter adjustment values ​​for the corresponding stage, and performs parameter adjustments through the stress resistance acclimatization execution subsystem and the seedling shed control subsystem. After acclimatization, the new shoot length, stem diameter, number of leaves, and root vitality of the grafted seedlings are detected. Plants that meet the preset standards are standardized grafted cultivation finished seedlings.

[0044] As a preferred embodiment, for the construction of the three-dimensional simulation model, this embodiment introduces a fluid topology optimization method that couples meteorological time series data with computational fluid dynamics. The layout of ventilation openings and temperature control equipment in the seedling shed is used as topological variables, and the wind speed uniformity and temperature gradient variance in the rootstock canopy area are used as constraints. The topological variables are updated by iteratively solving the adjoint equation, and the spatial optimal distribution parameters of the airflow field and temperature and humidity field in the seedling shed are dynamically optimized.

[0045] Specifically, in this embodiment, the design domain for fluid topology optimization is the area within the seedling shed where ventilation openings and temperature control equipment can be installed, including the walls and roof. The topology variables are pseudo-density variables from the density method. , ,when At that time, the corresponding location is the installation position of a physical ventilation opening / temperature control equipment. At that time, the corresponding location is a solid wall. The objective function of topology optimization is to maximize the wind speed uniformity and minimize the temperature gradient variance in the rootstock canopy region. The objective function expression is:

[0046]

[0047] in, The objective function value, For wind speed uniformity in the rootstock canopy area, The variance of the temperature gradient in the rootstock canopy region. , These are the weighting coefficients. .

[0048] Among them, wind speed uniformity The calculation formula is:

[0049]

[0050] in, This represents the number of grid nodes in the canopy region. For the first Wind speed values ​​at each node, This represents the average wind speed across all nodes in the canopy region.

[0051] Temperature gradient variance The calculation formula is:

[0052]

[0053] in, For the first Temperature gradient magnitude at each node This represents the average temperature gradient magnitude across all nodes in the canopy region. The topology optimization constraints include: total vent area constraint. Total power constraint of temperature control equipment And the constraints of the fluid control equations. During the iterative solution process, the adjoint equation method is used to solve for the gradient of the objective function with respect to the topological variables. The general form of the adjoint equation is:

[0054]

[0055] in, Let be the residual vector of the fluid control equations. The state variable vector for solving the flow field. For the accompanying variable vector, Let be the partial derivative of the objective function with respect to the state variables. After obtaining the gradient of the objective function with respect to the topological variables, the moving asymptote method is used to update the topological variables. The iterative convergence criterion is that the change in the objective function value between two consecutive iterations is less than 1. The maximum number of iterations was set to 200. After the iterations converged, the optimal layout parameters of the ventilation openings and temperature control equipment, as well as the corresponding equipment operating parameters, were obtained, further improving the standardization of rootstock cultivation.

[0056] Regarding the processing of rhizosphere and canopy data, this embodiment employs a multimodal data alignment and feature fusion method after collecting rhizosphere and canopy data. Rhizosphere chemical and biological data are mapped to a first-modal feature vector, and canopy physiological data are mapped to a second-modal feature vector. A cross-modal attention mechanism is used to calculate the correlation matrix of the two feature vectors in the time dimension, and feature-level fusion is performed to generate a joint spatiotemporal feature tensor that characterizes the overall growth status of the grafted seedling.

[0057] Specifically, the process begins with time alignment of the multimodal data. Rhizosphere and canopy data collected at different frequencies are unified into time-series data (1 time / hour) through linear interpolation, resulting in rhizosphere and canopy data sequences with equal time steps. For the rhizosphere data sequence, pH, EC, water content, and abundance data of four core microbial communities are extracted, totaling seven features. These features are then mapped to a dimension of [dimensional value missing] through a one-dimensional convolutional layer and layer normalization. First mode eigenvector ,in The length of the time series. Let be the feature dimension. The mapping formula for the first modality feature vector is:

[0058]

[0059] in, The aligned rhizosphere data sequence, , The weights and bias parameters of a one-dimensional convolutional layer. This is a one-dimensional convolution operation. For layer normalization operations, photosynthetically active radiation and transpiration rate data are extracted from the canopy data sequence, forming a total of 2D features. These features are then processed through a 1D convolutional layer of the same dimension and layer normalization, mapping them to a dimension of [missing information]. Second mode eigenvector The mapping formula is:

[0060]

[0061] in, The aligned canopy data sequence, , These are the weights and bias parameters corresponding to a one-dimensional convolutional layer.

[0062] Subsequently, a cross-modal attention mechanism is used to calculate the correlation matrix between the two feature vectors. This mechanism employs a scaled dot product attention structure, with the query vector... From the second mode eigenvector The key vector is obtained through linear transformation. AND value vector From the first mode eigenvector It is obtained through linear transformation, and the specific formula is as follows:

[0063]

[0064]

[0065] in, , , The weight matrix is ​​the linear transformation matrix. In this embodiment, the dimension of the key vector is... The output attention feature matrix is ​​a weighted representation of the attention of rhizosphere features to canopy features, with dimensions of [missing information]. The feature matrix output by cross-modal attention is concatenated with the original first-modal feature vector and second-modal feature vector in terms of feature dimension, and then subjected to feature dimensionality reduction through a fully connected layer to generate a joint spatiotemporal feature tensor. The formulas for splicing and dimensionality reduction are:

[0066]

[0067]

[0068] in, For feature concatenation operations, , These are the weights and bias parameters of the fully connected layer. The generated joint spatiotemporal feature tensor synchronously characterizes the spatiotemporal evolution of the rhizosphere microecological environment and the canopy physiological state, and is used for subsequent regulation of growth-promoting bacteria and stress resistance acclimatization processes.

[0069] As a preferred embodiment, this embodiment introduces a deep reinforcement learning resource scheduling method for the triggering process of quantitative dripping of growth-promoting bacteria agent. The joint spatiotemporal feature tensor is used as the state space, and the dripping amount and dripping frequency of growth-promoting bacteria agent are used as the action space. A delayed reward function is constructed with the root biomass increment of the grafted seedling within a future set time period as the objective. The optimal dripping action sequence is output by updating the strategy network through interaction with the environment.

[0070] Specifically, this embodiment employs a deep deterministic policy gradient algorithm to construct a deep reinforcement learning scheduling model. The model includes an actor network and a critic network. The actor network outputs deterministic dripping actions, while the critic network evaluates the value of the actions and guides the actor network's updates. State Space The joint spatiotemporal feature tensor generated for Example 2 The state dimension is In this embodiment, a time window is used. The hour, i.e., the state space is the joint spatiotemporal feature tensor of the 24 hours preceding the current moment. The action space... The continuous action space includes two action dimensions: the amount of growth-promoting agent injected in a single dose. The range of values ​​is , The maximum single drip volume of the cultivation container; drip interval. The range of values ​​is The unit is hours. The output of the action space is mapped to the corresponding value range through the Sigmoid activation function. The action output formula is:

[0071]

[0072] in, for Action vector at time step, for The state vector at time t, , For the mapping function of the actor network, These are the weight parameters for the actor network. Delayed reward function. The construction of the reward function is based on the increase in root biomass of grafted seedlings over the next 72 hours, while also incorporating a penalty term for the consumption of growth-promoting bacteria. The reward function expression is as follows:

[0073]

[0074] in, for Time to The increase in root biomass at time t, in units of Root biomass is related to rhizosphere water content and The value is obtained by combining time-series data with a pre-trained biomass prediction model; for The volume of growth-promoting agent consumed corresponding to each dripping action performed at any given time, in units of ; This is the penalty coefficient.

[0075] During model training, the actor network and the critic network sample through the experience replay pool, and the weights of the target network are updated using a soft update method. The update formula is as follows:

[0076]

[0077]

[0078] in, , These are the weight parameters for the target actor network and the target critic network, respectively. The soft update coefficient is used. After the model converges, the output optimal dripping sequence can dynamically match the root growth needs during the healing and seedling establishment periods, achieving precise quantitative dripping of the growth-promoting bacteria.

[0079] To address the phased adjustment process of environmental stress parameters, this embodiment employs a graph neural network construction method. It maps the species interactions of the rhizosphere microbial community to the canopy stress response characterization set as a dynamic graph structure. Microbial abundance and physiological indicators are used as node attributes, and material exchange and signal transmission relationships are used as edge attributes. A message passing mechanism is used to extract the spatiotemporal evolution characteristics of the dynamic graph structure to guide stress parameter adjustment. Please refer to the appendix. Figure 3 .

[0080] Specifically, dynamic graph structure At each time step Update, where the set of nodes This includes rhizosphere microbial nodes and canopy physiological nodes. The rhizosphere microbial nodes correspond to core microbial communities such as Bacillus, Pseudomonas, and Trichoderma, totaling [number missing]. There are 10 nodes; the canopy physiological nodes correspond to stress response indicators such as photosynthetically active radiation, transpiration rate, and leaf temperature. Number of nodes, total number of nodes .

[0081] Node attribute matrix , For node attributes, the attributes of rhizosphere microbial nodes are the real-time abundance values ​​of the corresponding microbial communities, while the attributes of canopy physiological nodes are the real-time detection values ​​of the corresponding physiological indicators. (Edge set) This includes intermicrobial interaction edges, signal transmission edges between microorganisms and the canopy, and edge attribute matrices. This is an adjacency matrix, where the weights of the edges between microorganisms are calculated using the Spearman correlation coefficient between species, and the weights of the edges between microorganisms and the canopy are calculated using the Granger causality test. The range of edge weight values ​​is... Positive values ​​represent a positive correlation, while negative values ​​represent a negative correlation.

[0082] A graph convolutional neural network is used to implement the message passing mechanism, and the update formula for node features is as follows:

[0083]

[0084] in, For the first The node feature matrix of the layer network, initial layer ; To add a self-loop to the adjacency matrix, It is the identity matrix; for The degree matrix; , For the first Weights and bias parameters of the layer network; The ReLU function is used as the non-linear activation function. The spatial feature matrix of the dynamic graph structure is extracted through multi-layer graph convolution operations. , This is the output feature dimension.

[0085] Furthermore, when extracting spatiotemporal evolution features, this embodiment combines a graph attention mechanism with a temporal convolutional network sequence prediction method. The graph attention mechanism assigns adaptive weights to nodes in the dynamic graph structure to focus on the core growth-promoting bacterial community. The weighted node features are then input into a temporal convolutional network to capture the local temporal dependence of the material flow conduction efficiency at the rootstock-stalk healing interface, outputting a predicted sequence of the healing interface conduction efficiency. Specifically, the graph attention mechanism is a multi-head graph attention structure. For each node, its attention weight with neighboring nodes is calculated. The formula for calculating the attention coefficient is:

[0086]

[0087] in, , for Time Node With nodes eigenvectors, The weight matrix is ​​a linear transformation matrix. For attention weight vectors, For feature concatenation operations, It is a non-linear activation function. for Time Node For nodes The normalized attention weights. The output of multi-head attention is the concatenation of the outputs of multiple attention heads, as shown in the formula:

[0088]

[0089] in, For the number of attention heads, For nodes The set of adjacent nodes, For the first Attention weights for each attention head For the first The linear transformation weight matrix of each attention head is used. Through the graph attention mechanism, higher attention weights are assigned to core growth-promoting bacterial community nodes to suppress interference from irrelevant bacterial communities and noise nodes.

[0090] The weighted node features at each time step are subjected to global average pooling to obtain the global feature vector for the corresponding time step. The time series feature matrix is ​​composed of global feature vectors at each time step. The input is processed by the Temporal Convolutional Network (TCN). TCN employs causal convolution and dilated convolution structures to prevent future information leakage while simultaneously expanding the receptive field. The formula for convolution operation is:

[0091]

[0092] in, The kernel size is [size]. For the first The weights of each convolutional kernel, The coefficient of thermal expansion is 1 / 3. This is the current time step. Through multi-layer TCN operations, local dependencies of temporal features are captured, and a predicted sequence of material flow conduction efficiency at the rootstock-scion healing interface for the next 24 hours is output. The material flow conduction efficiency is the rate of transmission of water, nutrients, and signaling substances between the rootstock and the scion, measured in units of... .

[0093] Furthermore, based on the predicted sequence of the healing interface conduction efficiency, this embodiment adopts an adaptive multi-objective optimization algorithm with the dual objective function of maximizing the peak conduction efficiency in the predicted sequence and minimizing the total consumption of probiotics output by the deep reinforcement learning resource scheduling method. The Pareto front solution set is obtained through non-dominated sorting and crowding calculation, and the triggering strategy of the probiotic-targeted regulation program is dynamically modified from the solution set.

[0094] Specifically, the objective function for bi-objective optimization is:

[0095]

[0096]

[0097] in, The first objective function is the peak value of the predicted sequence of healing interface conduction efficiency. The second objective function is the total amount of growth-promoting agent consumed. The optimization variables are the trigger threshold, infusion volume ratio coefficient, and upper limit of infusion frequency for the growth-promoting agent targeted regulation program.

[0098] The non-dominated sorting genetic algorithm NSGA-II with an elitist strategy is used to solve the above multi-objective optimization problem. The solution process includes: initializing the population; performing selection, crossover and mutation operations; performing non-dominated sorting on the individuals in the population and dividing them into Pareto ranks; calculating the crowding degree of each individual; selecting elite individuals based on Pareto rank and crowding degree to generate the next generation population; the iteration termination condition is that the number of iterations reaches a preset upper limit, or the change in the Pareto front solution set between two adjacent generations is less than a preset threshold.

[0099] After iterative convergence, the Pareto front solution set is obtained. The optimal compromise solution is selected from the solution set using the fuzzy membership method. The triggering strategy of the growth-promoting bacteria targeted regulation program is dynamically modified with the optimization variables corresponding to the optimal compromise solution, so as to reduce the total consumption of growth-promoting bacteria agent while ensuring the conduction efficiency of the rootstock-scion healing interface.

[0100] As a preferred embodiment, please refer to Figure 4 Before implementing stress resistance training, this embodiment introduces transfer learning and few-shot learning strategies. It extracts model parameters accumulated during multi-objective optimization and graph neural network training of grafted seedlings of different citrus varieties in history as prior knowledge of the source domain. The divergence between the source domain and the current distribution of grapefruit data is minimized through domain adaptation method, so as to quickly fine-tune the hyperparameters of the stress resistance training model with a small sample size.

[0101] Specifically, source domain dataset This dataset contains multimodal growth data of grafted seedlings from different historically cultivated citrus varieties, along with corresponding model training parameters. The sample size is [missing data]. ; Target domain dataset This is a small sample dataset of grafted grapefruit seedlings, with a sample size of [number missing]. , The resilient learning model's network structure includes a feature extraction layer and a prediction layer. During transfer learning, the weight parameters of the pre-trained feature extraction layer from the source domain are used as prior knowledge. The weights of the lower-level network of the feature extraction layer are frozen, and only the weight parameters of the top-level network and the prediction layer are fine-tuned. Maximum mean difference (MMD) is used to minimize the divergence between the feature distributions of the source and target domains. The formula for calculating MMD is:

[0102]

[0103] in, The mapping function for the feature extraction layer. Let the regenerative kernel Hilbert space be used. MMD is used as the domain adaptive loss function and added to the total loss function during model training. The expression for the total loss function is:

[0104]

[0105] in, The task loss function for the model is the mean squared error loss. The weights are the coefficients for the domain adaptive loss. Using the transfer learning strategy described above, only a small amount of sample data from "cocktail grapefruit" is needed to quickly fine-tune the resilience training model, solving the problems of slow convergence and poor generalization ability in small sample scenarios.

[0106] When fine-tuning the stress resistance domestication model, this embodiment introduces a privacy protection and collaborative training method under the federated learning framework. The model gradient is calculated locally on multiple distributed cultivation container nodes using desensitized perception data. Only the encrypted gradient parameters are uploaded to the central server for aggregation. This eliminates the model bias caused by different regional microclimates without sharing the original rhizosphere and canopy multimodal data.

[0107] Specifically, the federated learning framework employs the FedAvg algorithm, which includes... Each of the distributed cultivation container nodes corresponds to an independent cultivation site and possesses a local desensitized knowledge dataset. , The central server is an aggregation node and does not store any raw sensing data.

[0108] The collaborative training process includes: the central server assigning the weight parameters of the current global model... Distribute to all distributed nodes; each distributed node utilizes its local dataset. Perform model training and calculate local gradients. ,in The loss function is used for the local model; each node uses a homomorphic encryption algorithm to optimize its local gradient. Encryption is performed to obtain the encryption gradient. The data is then uploaded to the central server; the central server performs an aggregation operation on all encrypted gradients to obtain the global aggregated gradient. The central server updates the global model weight parameters by globally aggregating gradients. ,in The learning rate is used; the updated global model weight parameters are distributed to each node, and the above iterative process is repeated until the model converges.

[0109] The formula for calculating the global aggregate gradient is as follows:

[0110]

[0111] in, For the first The number of local samples per node The total number of samples across all nodes. Through the federated learning method described above, each node does not need to share its local raw sensing data, but only uploads encrypted gradient parameters, thus achieving data privacy protection. Simultaneously, it aggregates cultivation data from different regions, eliminating model bias caused by microclimate differences and improving the model's generalization ability.

[0112] Furthermore, when aggregating gradient parameters, this embodiment introduces an adversarial generation and verification method. A generator and a discriminator are constructed in the federated learning framework. The generator is used to fit the true distribution of gradient parameters uploaded by each node, and the discriminator is used to evaluate the consistency between the aggregated gradient and the local true gradient distribution. The gradient noise introduced by abnormal nodes is eliminated through adversarial training, ensuring the robustness of the output standardized grafted seedling quality evaluation index.

[0113] Specifically, generator It is a fully connected neural network, and the input is a random noise vector. The output is the fitted gradient vector. Used to fit the true distribution of the gradient uploaded by normal nodes. Discriminator This is a fully connected binary classification neural network. The input is a gradient vector, and the output is the probability that the gradient comes from the true normal gradient distribution. .

[0114] The objective function for adversarial training is a minimax game function, expressed as:

[0115]

[0116] in, Upload the true distribution of gradients for normal nodes. For the prior distribution of the input noise, the standard normal distribution is used in this embodiment.

[0117] After adversarial training is completed, the trained discriminator is used. The gradient parameters uploaded by each node are verified. When the probability value output by the discriminator is lower than a preset threshold, the gradient is determined to be an abnormal gradient, and the gradient parameters of that node are removed and not included in this global aggregation. Through the above adversarial generation and verification method, gradient noise introduced by malicious or faulty nodes is effectively eliminated, ensuring the stability and robustness of the global aggregation model, and thus ensuring the consistency of quality of grafted seedlings cultivated in different regions.

Claims

1. A smart grafting cultivation method for Cocktail Grapefruit, characterized in that, include: A three-dimensional simulation model of the airflow and temperature and humidity fields inside the seedling shed was constructed using computational fluid dynamics methods. Based on the simulation output results, the operating parameters of the equipment were adjusted to cultivate standardized rootstocks. The standardized rootstock was grafted onto a grapefruit scion to obtain grafted seedlings in the healing period. The grafted seedlings were planted in cultivation containers with built-in distributed sensing nodes to collect real-time data on the pH, EC, moisture content, and microbial community abundance of the rhizosphere soil, as well as the photosynthetically active radiation and transpiration rate of the canopy. Based on rhizosphere environment data, the quantitative dripping of growth-promoting bacteria is precisely triggered through a targeted regulation program to match the root growth needs during the healing and seedling establishment periods. Based on a pre-trained stress resistance acclimatization model, combined with real-time data collected from the canopy and rhizosphere, environmental stress parameters are adjusted in stages. The stress resistance acclimatization model is a pre-trained backpropagation neural network model. The input layer parameters include rhizosphere pH, EC value, water content, microbial community abundance, canopy photosynthetically active radiation, and transpiration rate. The output layer parameters are environmental stress parameters, including daytime temperature, nighttime temperature, relative humidity, photosynthetically active radiation supplementation, and substrate water content threshold in the seedling shed. When constructing the three-dimensional simulation model, a fluid topology optimization method coupled with meteorological time series data and computational fluid dynamics is introduced. The layout of ventilation openings and temperature control equipment in the seedling shed is used as topological variables. The wind speed uniformity and temperature gradient variance in the rootstock canopy area are used as constraints. The topological variables are updated by iteratively solving the adjoint equation, and the spatial optimal distribution parameters of airflow field and temperature and humidity field in the seedling shed are dynamically optimized. After collecting rhizosphere and canopy data, a multimodal data alignment and feature fusion method was used to map rhizosphere chemical and biological data into first-modal feature vectors and canopy physiological data into second-modal feature vectors. The correlation matrix of the two feature vectors in the time dimension was calculated using a cross-modal attention mechanism, and feature-level fusion was performed to generate a joint spatiotemporal feature tensor characterizing the overall growth status of the grafted seedling. When triggering the quantitative dripping of growth-promoting bacteria, a deep reinforcement learning resource scheduling method is introduced. The joint spatiotemporal feature tensor is used as the state space, and the dripping amount and frequency of the growth-promoting bacteria are used as the action space. A delayed reward function is constructed with the increase in root biomass of the grafted seedling within a future set time period as the objective. The optimal dripping action sequence is output by updating the strategy network through interaction with the environment.

2. The intelligent grafting cultivation method for Cocktail Grapefruit according to claim 1, characterized in that, When adjusting environmental stress parameters in stages, a graph neural network construction method is adopted to map the species interaction relationships of the rhizosphere microbial community and the canopy stress response characterization set into a dynamic graph structure. Microbial abundance and physiological indicators are used as node attributes, and material exchange and signal transmission relationships are used as edge attributes. The spatiotemporal evolution characteristics of the dynamic graph structure are extracted through a message passing mechanism to guide the adjustment of stress parameters.

3. The intelligent grafting cultivation method for Cocktail Grapefruit according to claim 2, characterized in that, When extracting spatiotemporal evolution features, the graph attention mechanism and the temporal convolutional network sequence prediction method are combined. The graph attention mechanism assigns adaptive weights to nodes in the dynamic graph structure to focus on the core growth-promoting bacterial community. The weighted node features are input into the temporal convolutional network to capture the local temporal dependence of the material flow conduction efficiency at the rootstock-sprout healing interface and output the predicted sequence of the conduction efficiency at the healing interface.

4. The intelligent grafting cultivation method for Cocktail Grapefruit according to claim 3, characterized in that, Based on the predicted sequence of healing interface conduction efficiency, an adaptive multi-objective optimization algorithm is adopted. The dual objective function is to maximize the peak conduction efficiency in the predicted sequence and minimize the total consumption of probiotic agents output by the deep reinforcement learning resource scheduling method. The Pareto front solution set is obtained by non-dominated sorting and crowding calculation. The triggering strategy of the probiotic targeted regulation program is dynamically modified from the solution set.

5. The intelligent grafting cultivation method for Cocktail Grapefruit according to claim 4, characterized in that, Before implementing stress resistance training, transfer learning and few-shot learning strategies were introduced. The model parameters accumulated during the multi-objective optimization and graph neural network training of grafted seedlings of different citrus varieties in history were extracted as prior knowledge of the source domain. The divergence between the source domain and the current distribution of grapefruit data was minimized through the domain adaptation method, and the hyperparameters of the stress resistance training model were quickly fine-tuned with a small sample size.

6. The intelligent grafting cultivation method for Cocktail Grapefruit according to claim 5, characterized in that, When fine-tuning the resilience domestication model, a privacy protection and collaborative training method under the federated learning framework is introduced. The model gradient is calculated locally on multiple distributed cultivation container nodes using desensitized perception data, and only the encrypted gradient parameters are uploaded to the central server for aggregation.

7. The intelligent grafting cultivation method for Cocktail Grapefruit according to claim 6, characterized in that, When aggregating gradient parameters, an adversarial generation and verification method is introduced. In the federated learning framework, a generator and a discriminator are constructed. The generator is used to fit the true distribution of gradient parameters uploaded by each node, and the discriminator is used to evaluate the consistency between the aggregated gradient and the local true gradient distribution.

Images