Dynamic adjusting method for concentration of hydrogen-rich water

Abstract

The invention provides a hydrogen-rich water concentration dynamic adjustment method, and relates to the field of concentration adaptive adjustment, and the method comprises the following steps: S1, data acquisition; step S2, data preprocessing; s3, key influence factors are extracted; s4, acquiring the initial concentration of the hydrogen-rich water; and S5, obtaining the optimized concentration of the hydrogen-rich water: obtaining a dynamic concentration adjustment amount, an extreme environment adjustment amount and an implicit stress adjustment amount through reinforcement learning dynamic optimization, extreme environment adaptation and implicit stress early warning, and obtaining the final concentration of the hydrogen-rich water. According to the method, self-adaptive adjustment of the concentration of the hydrogen-rich water can be performed according to the growth state of the plant and environmental factors, the accuracy of the concentration of the applied hydrogen-rich water is ensured, and the problem that normal growth of the plant is affected due to too high or too low concentration of the hydrogen-rich water is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of concentration self-adaptive adjustment, and particularly relates to a dynamic adjustment method for hydrogen-rich water concentration. BACKGROUND

[0002] Adding hydrogen-rich water in the irrigation water body, using the antioxidant properties of hydrogen in the water body to regulate the physiological metabolism of plants, that is, selectively neutralizing the excess active oxygen (ROS, such as hydroxyl radicals, superoxide anions, etc.) generated in the growth process of plants by hydrogen-rich water, reducing the oxidative damage of plants, protecting the function of organelles such as chloroplasts and mitochondria, and at the same time, using hydrogen to promote the synthesis of auxins (IAA), cell division factors (CTK) and other growth-promoting hormones in the plant body, and inhibit the accumulation of abscisic acid (ABA) and other stress hormones, thereby effectively regulating the division, growth and material synthesis of plant cells; in addition, hydrogen-rich water can also regulate the microbial community in the soil, improve the soil aeration, promote the root development, and thus improve the plant's ability to absorb water and nutrients, and improve the survival rate and growth of the plants.

[0003] Although hydrogen-rich water can effectively improve the growth of plants and achieve the yield increase of agriculture, if the concentration of hydrogen-rich water used for irrigation is too high, that is, the concentration of hydrogen-rich water does not match the growth state of the plants, not only the redox balance in the plant body will be broken (that is, the hydrogen-rich water with too high concentration will excessively remove the ROS in the plant body, affecting the cell division and gene expression, leading to the stagnation of root development and hindering the pollen germination), but also the plant growth and development will be inhibited, the photosynthesis efficiency will be reduced, and the nutrient composition of the plants will be imbalanced. At present, the concentration of hydrogen-rich water is mainly adjusted according to the experience and a large amount of test data of the operators, and the human factors are high and the error is large, so it is easy to have problems such as too high or too low concentration of hydrogen-rich water (too high concentration leads to large-area wilting or yield reduction of the plants, and too low concentration cannot achieve the purpose of applying hydrogen-rich water, which needs to be repeatedly applied for many times, which is time-consuming and laborious), which not only leads to the waste of hydrogen-rich water, but also easily causes the wilting or death of the plants, and thus affects the planting yield and economic value. SUMMARY

[0004] In view of the problems existing in the prior art, the purpose of the present application is to provide a dynamic adjustment method for hydrogen-rich water concentration, which belongs to the process control of crop planting, and can self-adaptively adjust the concentration of hydrogen-rich water according to the growth state of the plants and environmental factors, so as to effectively ensure the accuracy of the concentration of hydrogen-rich water and avoid the problems of too high or too low concentration of hydrogen-rich water.

[0005] The purpose of the present application is achieved by the following technical solutions: A dynamic adjustment method for hydrogen-rich water concentration, comprising: Step S1, data acquisition: according to the crop species and field conditions, deploying a collection sensor to collect the growth trend data of the plants and the field environmental factor data, respectively; Step S2, Data Preprocessing: The collected data undergoes noise removal, abnormal data detection and correction, and data normalization preprocessing. Step S3: Extraction of key influencing factors: Explore plant growth trend data and field environmental factor data to identify key factors affecting crop growth; Step S4: Obtaining the initial concentration of hydrogen-rich water: Input the data of key influencing factors into the fuzzy logic system, define the fuzzy subsets of each key influencing factor, and obtain the initial concentration of hydrogen-rich water. Step S5, Obtaining the optimized concentration of hydrogen-rich water: The dynamic concentration adjustment amount, extreme environment adjustment amount, and latent stress adjustment amount are obtained through reinforcement learning dynamic optimization, extreme environment adaptation, and latent stress early warning, respectively, to obtain the final concentration of hydrogen-rich water.

[0006] Based on further optimization of the above scheme, the growth trend data includes chlorophyll fluorescence, canopy multispectral data, stem microdeformation, leaf infrared temperature, root growth, etc.; field environmental factor data includes light intensity, air temperature and humidity, soil temperature and humidity, CO2 concentration, soil pH, etc.

[0007] Based on further optimization of the above scheme, the noise removal in step S2 is to remove random noise using the Kalman filter algorithm.

[0008] Based on further optimization of the above scheme, the abnormal data detection and correction in step S2 specifically includes: For input feature data X =[ x 1, x 2,…, x n The Isolation Forest algorithm was used to detect anomalies in the data after removing random noise, and anomaly scores were obtained. ; In the formula: h(X) Indicates sample X Path length in an isolated tree; E[h(X)] Indicates the expected path length; Represents Euler's constant; Preset abnormal score threshold SX d ,like If the value is 0, then the corresponding sample is determined to be an outlier. For outliers, Huber robust regression is used for correction. The Huber loss function is: ; ; In the formula: The parameters are constants; by optimizing the model parameters Improve tolerance to extreme environments.

[0009] Based on further optimization of the above scheme, in step S3, key factors are extracted using a stress-metabolism coupling network, a dynamic time-delay network, a multi-scale feature fusion network, and a stress state adaptive network, respectively. Specifically: The preprocessed data variables from step S2 are sequentially input into four networks: a stress-metabolism coupling network, a dynamic time-delay network, a multi-scale feature fusion network, and a stress state adaptive network. If a variable is identified as a key element by one of the four network structures, the corresponding variable is incremented by 1 in the key factor determination. If the total value of a variable is not less than 3, the variable is identified as a key influencing factor. If the total value of a variable is 2, it is considered a secondary influencing factor (when no key influencing factor is obtained, the secondary influencing factor is used as the key influencing factor).

[0010] Based on further optimization of the above scheme, the specific method for identifying key elements in the stress-metabolism coupling network is as follows: First, the preprocessed data variables from step S2 X’ Based on the effects of ROS (reactive oxygen species) on plants, the plant is divided into an environmental layer (CeX), a physiological layer (CeY), and an effect layer (CeZ). The environmental layer (CeX) corresponds to environmental indicators that directly induce ROS generation (i.e., ROS generation drivers, including data such as light intensity, air temperature and humidity, soil temperature and humidity, CO2 concentration, and soil pH). The physiological layer (CeY) reflects indicators of ROS damage to plant function (i.e., ROS damage targets, including data such as chlorophyll fluorescence and leaf infrared temperature). The effect layer (CeZ) reflects indicators of the final traits of plants affected by ROS (i.e., ROS regulation results, including data such as root growth, canopy multispectral data, and stem microdeformation). Then, two interlayer unidirectional influence pathways were established, including the first pathway CeX→CeY (the influence of environmental factors on physiological state) and the second pathway CeY→CeZ (the influence of physiological state on growth effect). Then, the path coefficients of the two paths are obtained using the partial least squares path model (PLS-PM): , , Finally, preset path coefficient threshold ,like Then the stress-metabolism coupling network will identify the corresponding variables as key elements.

[0011] Based on further optimization of the above scheme, the specific method for identifying key elements using the dynamic time-delay network is as follows: First, obtain the preprocessed data variables from step S2. X’(t)(This indicates that the characteristic variables were collected at time t and preprocessed in step S2) and concentration requirements. In time lag Mutual information under TDMI for:, ; In the formula: P represents the probability distribution; Representing variables X’ In time t Pick a Value and concentration requirement B in time Pick b The probability of the value; Representing variables X’ In time t Pick a The probability of the value This indicates that the concentration requirement B is in time. Pick b The probability of the value; Plant growth trend data and field environmental factor data were obtained at time lags. The TDMI is set below, with a preset TDMI peak threshold. max If the mutual information TDMI of a certain variable is greater than TDMI max Then the corresponding variable is determined to be a key element by the dynamic time-delay network.

[0012] Based on further optimization of the above scheme, the specific method for identifying key elements using the multi-scale feature fusion network is as follows: For time series variables X’(t) A three-level wavelet packet decomposition was performed (sampling frequency was once per hour, with a cumulative sampling period of 4 weeks to ensure data coverage of the "hour-day-week" scale); the db4 wavelet basis function was used for decomposition, dividing the time series variables into three levels: instantaneous (hourly), short-term (daily), and long-term (weekly), with a frequency band of 8; the 8 frequencies were sorted from high to low. j =1~2 represents high frequency, corresponding to the hour pole. j =3 to 6 represent medium frequency, corresponding to short-term day-level frequencies. j =7~8 are low frequencies, corresponding to long-term cycles), obtaining the energy values ​​of eight frequency bands. E i,j : ; In the formula: x i,j (t) Indicates the first i The first variable j Sub-signals in each frequency band; t 0、 t nThese represent the start and end times of the acquisition cycle corresponding to the sub-signal, respectively. The sub-signals of each frequency band are input as independent features into a random forest to obtain the Gini importance of each sub-signal; weights are assigned to different frequency bands based on the metabolic cycle caused by hydrogen-rich water. To obtain the overall importance of the corresponding variables: ; In the formula: Gini i,j Indicates the first i The first variable j The importance of the Gini coefficient in frequency band characteristics; Preset comprehensive importance threshold I d If the overall importance of a variable is greater than the overall importance threshold, then the corresponding variable is determined to be a key element by the multi-scale feature fusion network.

[0013] Based on further optimization of the above scheme, the specific method for identifying key elements by the stress state adaptive network is as follows: Introducing the Stress Index SI: ; In the formula: T leaf Indicates the infrared temperature of the blade; T opt This indicates the optimal temperature for the leaves; T max Indicates the critical high temperature; Indicates soil volumetric moisture content. Indicates field holding capacity; K 4. K 5 represents the corresponding weight coefficients; Preset stress index threshold SI d : when When the time is right, it indicates that the plant's metabolism is basically normal. Variables related to photosynthesis (such as canopy multispectral, light intensity, CO2 concentration, chlorophyll fluorescence, root growth, etc.) are used as input variables of the Bayesian network. The posterior probability of the variables and the concentration of hydrogen-rich water is calculated through the Bayesian network. The variables are arranged from largest to smallest and the top three variables are taken as key elements. when When the plant's photosynthesis / respiration is inhibited, stress-related variables (such as the degree of stem microdeformation, leaf infrared temperature, soil temperature and humidity, air temperature and humidity, soil pH, etc.) are used as input variables of the Bayesian network. The posterior probability of the variables and the concentration of hydrogen-rich water is calculated through the Bayesian network, and the variables are arranged from largest to smallest. The top three variables are selected as key elements. A Bayesian network is a probabilistic graphical model that outputs posterior probabilities based on the causal relationship between "field environmental factors / plant growth trends" and "hydrogen-rich water concentration requirements".

[0014] Based on further optimization of the above scheme, step S4 specifically includes: Define the membership function of the fuzzy subset of key influencing factors, including a small number of memberships. U dr (S) Suitable membership degree U su (S) Excessive membership U we (S) : ; In the formula: S Indicates the current measured values ​​of key influencing factors; S min This indicates the critical value when the key influencing factor is too low. S a Indicates the appropriate central value of key influencing factors; S max This indicates the critical value when key influencing factors are too high. A fuzzy rule base is constructed by integrating multi-factor correlation and expert practical experience, and the centroid method is used to defuzzify and obtain the initial concentration. ; In the formula: U i Indicates the first i The membership degree of a rule; C i Indicates the first i The reference concentration for this rule; m This indicates the number of rules that are active.

[0015] Based on the above scheme, the reinforcement learning dynamic optimization specifically includes: Using the vectors of key influencing factors (initial concentration and real-time monitoring) as the state space Y, and the concentration adjustment of hydrogen-rich water as the action space A, a reward function is constructed as follows: ; In the formula: This indicates the change in photosynthetic efficiency; This indicates the change in root growth. This indicates the adjustment range of the hydrogen-rich water concentration; These represent the corresponding weight coefficients; Update the Q value using the Q-Learning algorithm: ; In the formula: This represents the learning rate, typically between 0.1 and 0.3. This represents the discount factor, typically ranging from 0.7 to 0.9. Q(Y t , A t ) Represents the state at the current time t. Y t -action A t The corresponding Q value; Indicates the next state Y t+1 Next, iterate through all possible actions. a Choose the one with the largest Q value; When the change in Q value is less than a preset threshold, Q value converges; after Q value convergence, for any current state... Y t Choose the action that maximizes the Q value. A * : ; action A * This refers to the adjustment amount for dynamic optimization of hydrogen-rich water concentration in the current state using reinforcement learning. .

[0016] Based on further optimization of the above scheme, the extreme environment adaptation specifically includes: Collect historical extreme environment data to construct a meta-task set: ; In the formula: This represents a small amount of training data for the k-th extreme scenario; This represents the test data for the k-th extreme scenario; Use MAML (Model-Agnostic Meta-Learning) to pre-train general initial parameters : For each metatask k ,use Data is updated within the task: ; In the formula: Indicates the learning rate of the inner loop; This represents the training loss for task k; This represents the gradient operator, i.e., with respect to parameters. Find the partial derivative; use Data is updated meta-level and initial parameters are optimized: ; In the formula: This represents the outer loop learning rate; The novelty of environmental states is tested using isolated forests, and a novelty score is output. A preset novelty score threshold is established: if the novelty score exceeds the threshold, it is considered a new extreme environment; in the new extreme environment, real-time data is collected for meta-learning updates. ; In the formula: The loss function represents the data from the new environment; ; In the formula: Indicates the model uses the current parameters And based on the characteristics of the new environment X new Predicted demand for hydrogen-rich water concentration; y new This represents the optimal concentration requirement actually obtained in the corresponding extreme environment (obtained through historical experience data). N new Indicates the number of samples in the new environmental data; Use the adapted parameters Combined with reinforcement learning dynamic optimization to achieve Q ( Y t ,A t ; The biggest action A 2 * : ; action A 2 * That is, the adjustment amount of hydrogen-rich water concentration to adapt to extreme environments. .

[0017] Based on further optimization of the above scheme, the implicit stress early warning specifically refers to: Microscopic features of early latent stress were extracted from chlorophyll fluorescence and Raman spectroscopy data to construct a stress feature vector. Z x ; A variational autoencoder (VAE) is used to learn the feature distribution under normal conditions to obtain early latent stress: the VAE encoder will learn the feature distribution under normal conditions. Z x Mapping to latent space Output mean and variance; the VAE decoder will output the latent spatial variables. Reconstructed as features (in, and Z x (For vehicles with the same dimensions), the loss function of the VAE is: ; In the formula: This represents the "conditional probability distribution" defined by the decoder, which describes the reconstructed features. The probability distribution of the original input feature Z; E[·] represents the expected value of the reconstruction loss; This represents the approximate posterior distribution of the encoder output; represents the prior distribution of the latent variable; KL[·] represents the KL divergence; This represents the balance coefficient (usually 1). Calculate the reconstruction error REW: ; In the formula: Representing the eigenvector Z x The dimension; Preset error threshold If the reconstruction error REW is greater than the error threshold, then the feature is determined. Z x This is early-stage, latent stress; For characteristics identified as early latent stress Z x Calculate its relationship with the historical coercion database { Z 1, Z 2,…, Z n The Euclidean distance of} ; Choose the one with the smallest distance. h 1 sample, to obtain the initial adjustment amount: ; In the formula: C h,i Representation and Features Z x The smallest distance i The historical adjustment amount corresponding to each historical sample; Combining reinforcement learning dynamic optimization, that is, the action space A in reinforcement learning dynamic optimization is... If the fine-tuning interval is centered, the reward function is updated as follows: ; In the formula: This indicates the change in photosynthetic efficiency; This indicates the change in root growth. This indicates the adjustment range of the hydrogen-rich water concentration; These represent the corresponding weight coefficients; Latent stress adjustment for: ; In the formula: It represents the gradient of the Q-value with respect to the reward function, reflecting the impact of changes in the adjustment amount on the reward; Indicates the fine-tuning coefficient; The specific method for obtaining it is as follows: First, the Q-value is defined as "the long-term cumulative reward for performing adjustment action A in state Y," and its expression is the classic form of reinforcement learning, namely... ; but: ; In the formula: This represents the discount factor.

[0018] The following are the technical effects of the present invention: This invention identifies key elements through four networks: a stress-metabolism coupling network, a dynamic time-delay network, a multi-scale feature fusion network, and a stress state adaptive network. This allows for the output of key influencing factors for hydrogen-rich water. The identification process not only anchors the core physiological mechanisms, supplements the dynamics of the time dimension, covers the completeness of the time scale, and adapts to the variability of plant states, but also avoids problems such as detachment from physiological mechanisms, time mismatch, and inability to adapt to complex and changing field environments caused by outputting key influencing factors through a single network. This overcomes the limitations of traditional static analysis, ultimately identifying key influencing factors strongly correlated with hydrogen-rich water concentration requirements, achieving precise monitoring and regulation, eliminating redundant factor detection, and improving monitoring and control efficiency. Meanwhile, this invention dynamically adjusts the initial concentration of hydrogen-rich water output by the fuzzy logic system through reinforcement learning dynamic optimization, extreme environment adaptation, and implicit stress early warning. This effectively solves problems such as static rules being unable to adapt to real-time changes, conventional rules being unable to cover extreme scenarios, and explicit rules being unable to capture early implicit stress damage. It ensures that the hydrogen-rich water concentration not only matches the real-time state but also covers extreme situations and can predict implicit risks, achieving real-time and precise adjustment of the hydrogen-rich water concentration and avoiding problems of excessively low or high hydrogen-rich water concentration.

[0019] This invention can adjust the concentration of hydrogen-rich water in real time, continuously and accurately according to the growth status of plants and environmental factors, thereby improving the accuracy, adaptability and stability of hydrogen-rich water irrigation, thus meeting the needs of complex and ever-changing field scenarios, and avoiding problems such as reduced yield and death of plants due to excessively high hydrogen-rich water concentration or poor irrigation effect and increased cost due to excessively low hydrogen-rich water concentration. Attached Figure Description

[0020] Figure 1 This is a structural block diagram illustrating the dynamic adjustment of hydrogen-rich water concentration in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. In the following description, specific details such as specific system structures and technologies are presented for illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention.

[0022] Example 1: A method for dynamically adjusting the concentration of hydrogen-rich water, comprising: Step S1, Data Acquisition: Based on the crop type and field conditions, deploy acquisition sensors to collect plant growth trend data and field environmental factor data respectively; growth trend data includes chlorophyll fluorescence, canopy multispectral data, stem microdeformation degree, leaf infrared temperature, root growth, etc.; field environmental factor data includes light intensity, air temperature and humidity, soil temperature and humidity, CO2 concentration, soil pH value, etc.

[0023] Step S2, Data Preprocessing: The collected data undergoes preprocessing including Kalman filtering to remove random noise (using existing conventional Kalman filtering methods), outlier detection and correction, and data normalization (using existing conventional normalization methods). Specifically, outlier detection and correction are performed as follows: For input feature data X =[ x 1, x 2,…, x n The Isolation Forest algorithm was used to detect anomalies in the data after removing random noise, and anomaly scores were obtained. ; In the formula: h(X) Indicates sample X Path length in an isolated tree; E[h(X)] Indicates the expected path length; Represents Euler's constant; Preset abnormal score threshold SX d (Usually 0.7), if If the value is 0, then the corresponding sample is determined to be an outlier. For outliers, Huber robust regression is used for correction. The Huber loss function is: ; ; In the formula: This is a constant, typically taken as 1.35; the model parameters are optimized. Improve tolerance to extreme environments.

[0024] Step S3: Extraction of key influencing factors: Explore plant growth trend data and field environmental factor data to identify key factors affecting crop growth; Key factors were extracted using stress-metabolism coupling networks, dynamic time-delay networks, multi-scale feature fusion networks, and stress state adaptive networks, respectively. Specifically: The preprocessed data variables from step S2 are sequentially input into four networks: a stress-metabolism coupling network, a dynamic time-delay network, a multi-scale feature fusion network, and a stress state adaptive network. If a variable (i.e., any variable in the growth trend data and field environmental factor data) is identified as a key element by one of the four network structures, the corresponding variable's value is incremented by 1 in the key factor determination. If the total value of a variable is not less than 3, the variable is considered a key influencing factor. If the total value of a variable is 2, it is considered a secondary influencing factor (in the absence of a key influencing factor, the secondary influencing factor is used as the key influencing factor). For example, if leaf infrared temperature is identified as a key element by the stress-metabolism coupling network, the dynamic time-delay network, and the stress state adaptive network, but not by the multi-scale feature fusion network, then the leaf infrared temperature value is 3, making it a key influencing factor.

[0025] The specific method for identifying key elements in stress-metabolism coupling networks is as follows: First, the preprocessed data variables from step S2 X’ Based on the effects of ROS (reactive oxygen species) on plants, the plant is divided into an environmental layer (CeX), a physiological layer (CeY), and an effect layer (CeZ). The environmental layer (CeX) corresponds to environmental indicators that directly induce ROS generation (i.e., ROS generation drivers, including data such as light intensity, air temperature and humidity, soil temperature and humidity, CO2 concentration, and soil pH). The physiological layer (CeY) reflects indicators of ROS damage to plant function (i.e., ROS damage targets, including data such as chlorophyll fluorescence and leaf infrared temperature). The effect layer (CeZ) reflects indicators of the final traits of plants affected by ROS (i.e., ROS regulation results, including data such as root growth, canopy multispectral data, and stem microdeformation). Then, two interlayer unidirectional influence pathways were established, including the first pathway CeX→CeY (the influence of environmental factors on physiological state) and the second pathway CeY→CeZ (the influence of physiological state on growth effect). Then, the path coefficients of the two paths are obtained using the partial least squares path model (PLS-PM): , : If the stress-metabolism coupling network (i.e., the partial least squares path model) is a mapping of "latent variables to manifest variables", then the sign of the latent variables in the environmental layer is... The corresponding explicit variable is the light intensity after preprocessing in step S2. air temperature and humidity Soil temperature and humidity CO2 concentration Soil pH Etc., that is, the explicit variables of the environment layer are ( j =1, 2, 3, 4, 5); similarly, the symbols for latent variables in the physiological layer are... The corresponding dominant variable is chlorophyll fluorescence. Blade infrared temperature etc., that is ( j =1, 2); the latent variables in the effect layer are , and the corresponding manifest variables are root growth potential. Canopy Multispectral Degree of slight deformation of stem etc., that is ( j =1, 2, 3); For example: the latent variable score for the environmental layer is: ; In the formula: In the CeX environment layer, the first j The first environment variable, the first k The weight coefficients for the next iteration; Simultaneous generation of latent variable scores for the physiological layer: ; PLS-PM iterates through a loop of "external weights → internal weights → path coefficients" until the weight change is less than a threshold (e.g., 10). -4 That is, the weight difference between two adjacent iterations. Stop the iteration and output the final path coefficients; the path coefficients represent the strength and direction of the influence between latent variables, and their essence is the "standardized regression coefficient of the target latent variable on the source latent variable".

[0026] in: External weights of the environment layer: ; Internal weights: ; In the formula: Cov() Represents covariance; Var() Indicates variance; because Standardization has been performed, and the variance of the standardized variables can be simplified to 1, i.e.: ; ; In the formula: n represents the sample size (i.e., the total number of plants collected in the field, such as n=100), and X and Y represent the corresponding variables (such as the latent variable scores of the environmental layer). Physiological layer latent variable score wait); X i Representing variables X The i Observed values ​​of plant samples; Representing variables X The sample mean; Similarly, the second path coefficients are obtained. ; Finally, preset path coefficient threshold (Usually 0.5), if In this case, the stress-metabolism coupling network identifies the corresponding variables as key elements; for example: (That is, in this path, CeX is the input layer variable and CeY is the output layer variable), and the corresponding variable is the data in the environment layer CeX, then the variable in the environment layer is determined to be a key element; if (That is, CeY is the input layer variable and CeZ is the output layer variable in this path), and the corresponding variable is the data in the physiological layer CeY. Then the variable in the physiological layer is determined to be the key element.

[0027] The specific method for identifying key elements using dynamic time-delay networks is as follows: First, obtain the preprocessed data variables from step S2. X’(t) (This indicates that the characteristic variables were collected at time t and preprocessed in step S2) and concentration requirements. In time lag Mutual information under TDMI for:, ; In the formula: P represents the probability distribution; Representing variables X’ In time t Pick a Value and concentration requirement B in time Pick b The probability of the value; Representing variables X’ In time t Pick a The probability of the value This indicates that the concentration requirement B is in time. Pick b The probability of the value; where concentration requirement B is the target concentration of hydrogen-rich water required by the plant under the corresponding growth state; (the probability can be obtained using the frequency probability approximation method, for example: based on the time series data collected in step S1: the growth trend / environmental factor variable X is collected once per hour for a cumulative period of 4 weeks; the time series data of concentration requirement B comes from the matching records of "key factors-optimal concentration" in the historical database, and the continuous variables after preprocessing in step S2 are...) X’ The concentration requirement B was divided into 5–10 equally spaced intervals, with the number of intervals adjusted according to the data distribution density; for the variable… X’ The result is obtained by counting the frequency of the value 'a' within a certain interval in hour t and dividing by the total amount of data. For concentration requirement B, its statistical value is... Take the value within the interval within the hour b The frequency percentage, that is, Statistical variables X’ Take value a within hour t, and the concentration requirement B is... Take values ​​within an hour b The frequency of synchronous occurrences divided by the total amount of data yields the result. (Data interval partitioning and frequency statistics can be performed using Python's pandas library, and probability calculations can be performed using the scipy.stats library.) Plant growth trend data and field environmental factor data were obtained at time lags. TDMI under the range of 0–48 h, with a preset TDMI peak threshold. max (Typically 0.8 bits), if the mutual information TDMI of a certain variable is greater than TDMI max Then the corresponding variable is determined to be a key element by the dynamic time-delay network.

[0028] The specific method for identifying key elements using multi-scale feature fusion networks is as follows: For time series variables X’(t)A three-level wavelet packet decomposition was performed (sampling frequency was once per hour, with a cumulative sampling period of 4 weeks to ensure data coverage across hourly, daily, and weekly scales; the db4 wavelet basis function was used for decomposition, dividing the time-series variables into three levels: instantaneous (hourly), short-term (daily), and long-term (weekly). The first level, instantaneous response: sudden increase in leaf temperature (direct sunlight) → hydrogen-rich water needs to quickly alleviate transpiration stress, with 2 frequency bands; the second level, short-term response: slight stem deformation (water deficit) → hydrogen-rich water needs to continuously regulate turgor pressure, with 4 frequency bands; the third level, long-term response: root growth (nutrient absorption) → hydrogen-rich water needs to promote metabolic repair, with 8 frequency bands; the 8 frequencies were sorted from high to low. j =1~2 represents high frequency, corresponding to the hour pole. j =3 to 6 represent medium frequency, corresponding to short-term day-level frequencies. j =7~8 are low frequencies, corresponding to long-term cycles), obtaining the energy values ​​of eight frequency bands. E i,j : ; In the formula: x i,j (t) Indicates the first i The first variable j Sub-signals in each frequency band; t 0、 t n These represent the start and end times of the acquisition cycle corresponding to the sub-signal, respectively. The sub-signals of each frequency band are used as independent features and input into a random forest (the random forest model chosen is RandomForestRegressor from the Python scikit-learn library, using 8 frequency band sub-signals as 8 independent feature inputs and "hydrogen-rich water concentration requirement" as the label). A label vector is formed by matching the probability of "8 sub-signal states → optimal irrigation hydrogen-rich water concentration" in field experiments. The preprocessed feature matrix is ​​then used as the input. X’ The label vectors are input into a random forest model for training. After the random forest is trained, the Gini importance of each of the eight frequency band sub-signals is obtained directly through the model's built-in feature_importance_ attribute. Based on the metabolic cycle caused by hydrogen-rich water, weights are assigned to different frequency bands. (Generally high frequency) Mid-frequency low frequency ), to obtain the overall importance of the corresponding variables: ; In the formula: Gini i,j Indicates the first iThe first variable j The importance of the Gini coefficient in frequency band characteristics; Preset comprehensive importance threshold I d If the overall importance of a variable is greater than the overall importance threshold, then the corresponding variable is determined to be a key element by the multi-scale feature fusion network.

[0029] The specific method for identifying key elements using a stress-state adaptive network is as follows: Introducing the Stress Index SI: ; In the formula: T leaf Indicates the infrared temperature of the blade; T opt This indicates the optimal temperature for the leaves; T max Indicates the critical high temperature; Indicates soil volumetric moisture content. Indicates field holding capacity; K 4. K 5 represents the corresponding weight coefficients; Preset stress index threshold SI d (Usually 0.3): when When the plant's metabolism is basically normal, variables related to photosynthesis (such as canopy multispectral density, light intensity, CO2 concentration, chlorophyll fluorescence, root growth, etc.) are used as input variables of the Bayesian network. The Bayesian network calculates the posterior probability of the variables and the concentration of hydrogen-rich water, and the variables are arranged from largest to smallest. The top three variables are selected as key elements (e.g., canopy multispectral density, light intensity, CO2 concentration). when When the plant's photosynthesis / respiration is inhibited, stress-related variables (such as the degree of stem microdeformation, leaf infrared temperature, soil temperature and humidity, air temperature and humidity, soil pH, etc.) are used as input variables of a Bayesian network. The Bayesian network calculates the posterior probability of the variables and the concentration of hydrogen-rich water, and the variables are arranged from largest to smallest. The top three variables are selected as key elements (e.g., the degree of stem microdeformation, leaf infrared temperature, and soil temperature and humidity). A Bayesian network is a probabilistic graphical model that outputs a posterior probability based on the causal relationship between "field environmental factors / plant growth trends" and "hydrogen-rich water concentration requirements," specifically: First, growth trend data and corresponding variables of field environmental factors are used as input variables for the parent node, and hydrogen-rich water concentration requirements are used as output variables for the child nodes. Based on the mechanism of action of hydrogen-rich water and past empirical experimental data, a directed acyclic graph (DAG) is drawn (the structure can be defined using the BayesianNetwork class of the Python pgmpy library, for example: light intensity ↑ → concentration requirement ↓, soil moisture content ↓ → concentration requirement ↑, chlorophyll fluorescence ↓ → concentration requirement ↑, stem microdeformation degree ↑ → concentration requirement ↑). Then, the conditional probability of the child node (concentration requirement) under different parent node states is calculated to form a conditional probability table (CPT, which can be implemented using the MaximumLikelihoodEstimator class of the pgmpy library). Finally, the measured values ​​of variables at the current moment in the field are collected and converted into network-recognizable states (consistent with the discretization / continuation rules in the data preparation stage). Based on Bayes' theorem, the posterior probability of each variable and the current concentration requirement is inferred (using the VariableElimination inferencer of the pgmpy library). ; In the formula: These represent the prior probability of the variable (obtained by the frequency percentage of each state of the variable occurring naturally in the field), the conditional probability of the concentration requirement (obtained by statistically analyzing the corresponding frequency percentages based on historical matching data of "variable state - optimal concentration requirement"), and the marginal probability of the concentration requirement (obtained by directly statistically analyzing the frequency percentages of each state of the concentration requirement in historical data).

[0030] Step S4: Obtaining the initial concentration of hydrogen-rich water: Input the data of key influencing factors into the fuzzy logic system, define the fuzzy subsets of each key influencing factor, and obtain the initial concentration of hydrogen-rich water. Specifically: Define the membership function (range [0,1]) of the fuzzy subset of key influencing factors, including a small number of memberships. U dr (S) Suitable membership degree U su (S) Excessive membership U we (S) : ; In the formula: S Indicates the current measured values ​​of key influencing factors; S min This indicates the critical value when the key influencing factor is too low. S a Indicates the appropriate central value of key influencing factors; Smax This indicates the critical value when key influencing factors are too high. A fuzzy rule base is constructed by integrating multi-factor correlations and expert practical experience. (When constructing the fuzzy rule base, each rule is a combination of the membership degrees of multiple factors; for example, a rule could be: if the light intensity is "excessive", the soil moisture is "slight", and the chlorophyll fluorescence is "suitable", then the reference concentration of hydrogen-rich water is...) C i The initial concentration was 1.5 mg / L, and the centroid method was used to defuzzify the sample to obtain the preliminary concentration. ; In the formula: U i Indicates the first i The membership degree of a rule; C i Indicates the first i The reference concentration for this rule; m This indicates the number of rules that are active.

[0031] Step S5, Obtaining the optimized concentration of hydrogen-rich water: The dynamic concentration adjustment amount, extreme environment adjustment amount, and latent stress adjustment amount are obtained through reinforcement learning dynamic optimization, extreme environment adaptation, and latent stress early warning, respectively, to obtain the final concentration of hydrogen-rich water.

[0032] Specifically, reinforcement learning dynamic optimization includes: The initial concentration and the vector of key influencing factors in real-time monitoring are used as the state space Y (state Y = {initial concentration of hydrogen-rich water}). C st +Key influencing factor vector}), using the concentration adjustment of hydrogen-rich water as the action space A, construct the reward function: ; In the formula: This indicates the change in photosynthetic efficiency; This indicates the change in root growth. This indicates the adjustment range of the hydrogen-rich water concentration; These represent the corresponding weight coefficients; The Q-value is updated using the Q-Learning algorithm (Q-Learning is a classic value-based, model-free algorithm in reinforcement learning. Its core objective is to learn a "state-action" value function, or Q-function, through trial and error between the agent and the environment, thereby guiding the agent to choose the optimal action in any state to maximize long-term cumulative reward). ; In the formula: This represents the learning rate, typically between 0.1 and 0.3. This represents the discount factor, typically ranging from 0.7 to 0.9.Q(Y t , A t ) Represents the state at the current time t. Y t -action A t The corresponding Q value (to the left of the arrow) Q(Y t ,A t ) The updated, more optimal value estimate, i.e., the "corrected new understanding"; the right side of the arrow Q(Y t ,A t ) This represents the value estimate of action A performed in state Y before the update, i.e., the old knowledge; both the left and right sides of the formula show... Q(Y t ,A t ) This is because the Q-Learning algorithm uses an "incremental update" logic, which means it does not directly replace the old Q value, but rather makes corrections based on the original estimate and new experience. Indicates the next state Y t+1 Next, iterate through all possible actions. a Select the one with the largest Q-value (a core feature of the Q-Learning algorithm, demonstrated through offline learning, representing the new state). Y t+1 (the optimal expected reward); When the change in Q value is less than a preset threshold (typically 0.01), Q value converges; after Q value convergence, for any current state... Y t Choose the action that maximizes the Q value. A * : ; action A * This refers to the adjustment amount for dynamic optimization of hydrogen-rich water concentration in the current state using reinforcement learning. .

[0033] Extreme environment adaptation specifically refers to: Collect historical extreme environment data to construct a meta-task set: ; In the formula: This represents a small amount of training data for the k-th extreme scenario; This represents the test data for the k-th extreme scenario; Use MAML (Model-Agnostic Meta-Learning) to pre-train general initial parameters : For each metatask k ,use Data is updated within the task: ; In the formula: Indicates the learning rate of the inner loop; This represents the training loss for task k; This represents the gradient operator, i.e., with respect to parameters. Find the partial derivative; use Data is updated meta-level and initial parameters are optimized: ; In the formula: This represents the outer loop learning rate; The novelty of environmental states is assessed using isolated forests, and a novelty score is output, which includes: Collect at least three types of historical extreme scenario data (such as high temperature and drought, heavy rain and high humidity, extreme pH stress, etc.), with no fewer than 50 data samples for each scenario, and remove completely duplicate samples to serve as the training set; collect real-time data of the current field environment as the test set, with its feature dimensions completely consistent with the training set (e.g., if the training set contains 5 features, the test set cannot be increased or decreased), and the collection frequency of the real-time data is consistent with the training set (e.g., 1 data sample per hour); then, use the data preprocessing method of step S2 (Kalman filtering algorithm and Huber robust regression to preprocess the collected data), and map the features of the training set and test set to the [0,1] interval; use the IsolationForest class of the Python scikit-learn library as the isolated forest model, and train the model through the training set; the novelty score of the isolated forest is based on the "average path length of the sample in the isolated tree" transformation, and the score range is [0,1]. For example, for real-time test samples... x test The isolated forest model was used to calculate its performance. N tree Average path length in an isolated tree Thus, a novelty score is obtained. S xyx : ; In the formula: E[h(x)] This represents the average expected path length of the training set samples (automatically calculated by the model and related to the sample size and number of features). A preset novelty score threshold (typically 0.9) is set: if the novelty score exceeds the threshold, it is considered a new extreme environment; in the new extreme environment, real-time data is collected for meta-learning updates. ; In the formula: The loss function represents the data from the new environment; ; In the formula: Indicates the model uses the current parameters And based on the characteristics of the new environment X new Predicted demand for hydrogen-rich water concentration; y new This represents the optimal concentration requirement actually obtained in the corresponding extreme environment (obtained through historical experience data). The model parameters are: At that time, in the new extreme environment, the first i The predicted output value for each sample; N new Indicates the number of samples in the new environmental data; Use the adapted parameters Combined with reinforcement learning dynamic optimization to achieve Q ( Y t ,A t ; The biggest action A 2 * (That is, when the change in Q value is less than a preset threshold of 0.01, Q value converges; after Q value convergence, for any current state) Y t Choose the action that maximizes the Q value. A 2 * (Similar to the reinforcement learning dynamic optimization above) ; action A 2 * That is, the adjustment amount of hydrogen-rich water concentration to adapt to extreme environments. .

[0034] The latent stress early warning mechanism is a mechanism for identifying potential stresses (such as mild drought, latent nutrient imbalance, and micro-abiotic stress) in the early stages of dynamic regulation of hydrogen-rich water concentration, before any visible damage is observed in the plant. Its core is to accurately determine the presence and severity of stress by capturing abnormal changes in the plant's microscopic physiological characteristics, before the stress manifests as visible symptoms (such as wilting, yellowing, or growth stagnation). This provides a "predictive basis" for adjusting the hydrogen-rich water concentration, preventing the stress from developing into visible damage. The latent stress early warning mechanism is the perception mechanism for "early detection of potential stresses," while the hydrogen-rich water adjustment amount is the execution parameter for "precisely correcting the concentration for potential stresses." The two are causally linked. Microscopic features of early latent stress were extracted from chlorophyll fluorescence and Raman spectroscopy data to construct a stress feature vector. Z x ; A variational autoencoder (VAE) is used to learn the feature distribution under normal conditions to obtain early latent stress: the VAE encoder will learn the feature distribution under normal conditions. Z x Mapping to latent space Output mean and variance; the VAE decoder will output the latent spatial variables. Reconstructed as features (in, and Z x (For vehicles with the same dimensions), the loss function of the VAE is: ; In the formula: This represents the "conditional probability distribution" defined by the decoder, which describes the reconstructed features. The probability distribution of the original input feature Z; E[·] represents the expected value of the reconstruction loss; This represents the approximate posterior distribution of the encoder output; represents the prior distribution of the latent variable; KL[·] represents the KL divergence; This represents the balance coefficient (usually 1). Calculate the reconstruction error REW: ; In the formula: Representing the eigenvector Z x The dimension; Preset error threshold (Generally 0.8), if the reconstruction error REW is greater than the error threshold, then the feature is determined. Z x This is early-stage, latent stress; For characteristics identified as early latent stress Zx Calculate its relationship with the historical coercion database { Z 1, Z 2,…, Z n The Euclidean distance of} ; Choose the one with the smallest distance. h 1 sample, to obtain the initial adjustment amount: ; In the formula: C h,i Representation and Features Z x The smallest distance i The historical adjustment amount corresponding to each historical sample; Combining reinforcement learning dynamic optimization, that is, the action space A in reinforcement learning dynamic optimization is... If the fine-tuning interval is centered, the reward function is updated as follows: ; In the formula: This indicates the change in photosynthetic efficiency; This indicates the change in root growth. This indicates the adjustment range of the hydrogen-rich water concentration; These represent the corresponding weight coefficients; Latent stress adjustment for: ; In the formula: It represents the gradient of the Q-value with respect to the reward function, reflecting the impact of changes in the adjustment amount on the reward; This represents the fine-tuning coefficient (typically 0.01). The specific method for obtaining it is as follows: First, the Q-value is defined as "the long-term cumulative reward for performing adjustment action A in state Y," and its expression is the classic form of reinforcement learning, namely... ; but: ; In the formula: This represents the discount factor (typically 0.7 to 0.9).

[0035] The final concentration of hydrogen-rich water obtained is: ; In the formula: This represents the correction factor (i.e., the correction factor based on the physiological characteristics of the crop at its current growth stage); when the plant is in the flowering stage, When the plant is in the seedling stage, .

[0036] Example 2: As a further optimization of the scheme in this application, based on the scheme in Example 1, the chlorophyll fluorescence parameters include the maximum optical efficiency ( F v / F m The potential activity of the photosystem II reaction center is reflected in the photosystem II protein complex located on the thylakoid membrane during photosynthesis. Its main function is to absorb light energy and catalyze the photolysis of water to produce oxygen, protons, and electrons. The photochemical quenching coefficient (qP, reflecting the open proportion of the photosystem II reaction center) is also considered.

[0037] In the formula: F 0 indicates the initial fluorescence intensity; F m Indicates the maximum fluorescence intensity; F v Indicates variable fluorescence intensity; F t Indicates the steady-state fluorescence intensity under light adaptation; Indicates the maximum fluorescence intensity under light adaptation; Indicates the initial fluorescence intensity under light adaptation; Canopy multispectral parameters include the Normalized Difference Vegetation Index (NDVI) and Red Edge Position (REP, a sensitive indicator reflecting changes in chlorophyll content):

[0038] In the formula: Indicates near-infrared reflectance; Indicates the reflectivity in the red light band; , , These represent the reflectivity at wavelengths of 680nm, 705nm, and 750nm, respectively. Degree of micro-deformation of the stem (measured by strain sensor):

[0039] In the formula: This indicates the change in stem length; Indicates the initial length of the stem; Leaf infrared temperature (canopy temperature, measured by an infrared sensor):

[0040] In the formula: C1 and C2 represent the parameters related to Planck's constant (C1 is...). C2 is ); This represents the spectral radiance received by the sensor; Indicates the wavelength of the infrared band; Root growth parameters (obtained through 3D image scanning) include root length density (RLD, reflecting the root absorption range) and root surface area index (RSI, reflecting absorption capacity):

[0041] In the formula: L 根 Indicates the root system length; V 土壤 This indicates the corresponding soil volume; r 根 Indicates the root radius.

[0042] Field environmental data, such as light intensity, air temperature and humidity, soil temperature and humidity, CO2 concentration, and soil pH, can be obtained using corresponding light intensity sensors, air temperature and humidity sensors, soil temperature and humidity sensors, CO2 concentration sensors, and soil pH sensors.

[0043] Example 3: As a further optimization of the present application, a dynamic adjustment system for hydrogen-rich water concentration is also included, which uses any one of the methods in Example 1 or Example 2 to adjust the concentration of hydrogen-rich water during the crop planting process.

Claims

1. A method for dynamically adjusting the concentration of hydrogen-rich water, characterized by: The application relates to a method for obtaining the optimal concentration of hydrogen-rich water for crops, comprising the following steps: Step S1, data collection: according to the crop type and field conditions, a collection sensor is arranged to collect the growth trend data of the plants and the field environmental factor data; Step S2, data preprocessing: the collected data is preprocessed through noise removal, abnormal data detection and correction, and data normalization; Step S3, key influence factor extraction: the plant growth trend data and the field environmental factor data are explored to determine the key factors affecting the growth of crops; Step S4, preliminary concentration of hydrogen-rich water acquisition: the key influence factor data is input into a fuzzy logic system, the fuzzy subsets of the key influence factors are defined, and the preliminary concentration of the hydrogen-rich water is obtained; Step S5, optimal concentration of hydrogen-rich water acquisition: the dynamic concentration adjustment amount, the extreme environment adjustment amount and the implicit stress adjustment amount are respectively obtained through reinforcement learning dynamic optimization, extreme environment adaptation and implicit stress early warning, and the final concentration of the hydrogen-rich water is obtained.

2. The method of claim 1, wherein the concentration of hydrogen-rich water is dynamically adjusted. In the step S2, the Kalman filtering algorithm is used to remove random noise.

3. The method of claim 1 or 2, wherein the method comprises: In the step S2, the abnormal data detection and correction are specifically as follows: For input feature data X [ x 1, x 2,…, x n ], the isolated forest algorithm is used for anomaly detection on the data removed from random noise to obtain data anomaly scores: ； wherein: h(X) denotes the sample X path length in an isolated tree; E[h(X)] denotes the expected path length; denotes the Euler constant; Pre-set abnormal score threshold SX d , if , the corresponding sample is determined as an outlier. For abnormal values, Huber robust regression is used for correction, and the Huber loss function is as follows: ； ； In the formula: is a constant; the model parameters are optimized Improved fault tolerance to extreme environments.

4. The method of claim 2 or 3, wherein: In the step S3, the stress-metabolism coupling network, the dynamic time delay network, the multi-scale feature fusion network and the stress state adaptive network are used to extract the key factors, and the specific method is as follows: The data variables preprocessed in the step S2 are sequentially input into the stress-metabolism coupling network, the dynamic time delay network, the multi-scale feature fusion network and the stress state adaptive network, if a variable is identified as a key element by a network in the four network structures, the numerical value of the corresponding variable is added by 1 in the key factor judgment; If the total numerical value of a variable is not less than 3, the variable is determined as a key influence factor; if the total numerical value of a variable is 2, the variable is determined as a secondary influence factor.

5. The method of claim 1 or 4, wherein the method comprises: The specific method for identifying the key elements by the stress-metabolism coupling network is as follows: First, the data variable after pretreatment in step S2 is changed as follows X’ According to the influence of ROS on the plant, it is divided into environmental layer CeX, physiological layer CeY and effect layer CeZ; wherein, the environmental layer CeX corresponds to the environmental index directly inducing the generation of ROS, the physiological layer CeY reflects the index of ROS damage to the function of the plant, and the effect layer CeZ reflects the trait index finally embodied by the ROS influence on the plant. Then, two interlayer one-way influence paths are established, including a first path CeX->CeY and a second path CeY->CeZ. After that, the path coefficients of the two paths were obtained by using the partial least squares path model (PLS-PM): , , Finally, a preset path coefficient threshold value , if , the stress-metabolism coupling network will determine the corresponding variable as a key element.

6. The method of claim 4, wherein the concentration of hydrogen-rich water is dynamically adjusted. The specific method for identifying the key elements by the dynamic time delay network is as follows: First, the data variable pre-processed in the acquisition step S2 X'(t) with the concentration requirement , mutual information under time delay TDMI is: ； where P denotes a probability distribution; denotes a variable X’ at time t takes a a value and the concentration demand B at time takes b a value with probability; denotes a variable X’ at time t the probability a that the variable takes a value, denotes the probability that the concentration demand B at time b takes Plant growth trend data and field environmental factor data were obtained at time lags. The TDMI is set below, with a preset TDMI peak threshold. max If the mutual information TDMI of a certain variable is greater than TDMI max Then the corresponding variable is determined to be a key element by the dynamic time-delay network.

7. The method of claim 4, wherein the concentration of hydrogen-rich water is dynamically adjusted. The specific method for identifying the key elements by the multi-scale feature fusion network is as follows: On timing variables X'(t) A three-layer wavelet packet decomposition is performed to obtain energy values of eight frequency bands E i,j : ； In the formula: x i,j (t) represents the sub-signal of the i-th frequency band of the j-th variable; i j t 0、 t n respectively represent the start time and the end time of the acquisition period corresponding to the sub-signal.​​ The sub-signals of each frequency band are input into the random forest as independent features to obtain Gini importance Gini of each sub-signal; and weights are assigned to different frequency bands based on the metabolic cycle of the hydrogen-rich water action to obtain comprehensive importance of the corresponding variables: ； In the formula: Gini i,j represents the first i variable of the first j band feature Gini importance; Pre-set comprehensive importance threshold I d If the comprehensive importance of the variable is greater than the comprehensive importance threshold, the corresponding variable is determined as a key element by the multi-scale feature fusion network.

8. The method of claim 4, wherein the concentration of hydrogen-rich water is dynamically adjusted. The specific method for identifying the key elements by the stress state adaptive network is as follows: A stress index SI is introduced: ； wherein: T leaf represents the leaf infrared temperature; T opt represents the optimum leaf temperature; T max represents the critical high temperature; represents the soil volumetric water content, represents the field capacity; K 4、 K 5 respectively represent the corresponding weight coefficients; Predefined stress index threshold SI d : When the plant metabolism is basically normal, the variables related to photosynthesis are taken as the input variables of the Bayesian network, the posterior probability of the variables and the concentration of hydrogen-rich water is calculated by the Bayesian network, the variables are arranged from large to small, and the top three variables are taken as the key elements; When the plant photosynthesis / respiration is inhibited, the stress resistance related variables are taken as the input variables of the Bayesian network, the posterior probability of the variables and the concentration of hydrogen-rich water is calculated by the Bayesian network, and the top three variables are taken as the key elements in descending order. The Bayesian network is a kind of probability graph model, which outputs the posterior probability through the causal relationship between the "field environmental factor / plant growth trend" and the "hydrogen-rich water concentration demand".

9. The method of claim 1, wherein the method further comprises: determining a concentration of hydrogen in the water; and adjusting the concentration of hydrogen in the water based on the determined concentration of hydrogen in the water. The step S4 is specifically as follows: Defining membership functions for the fuzzy subsets of key influencers, including low, suitable, and excessive membership U dr (S) , suitable membership U su (S) and excessive membership U we (S) : ； wherein: S represents the current measured value of the key influencer; S min represents the critical value when the key influencer is too low; S a represents the suitable central value of the key influencer; S max represents the critical value when the key influencer is too high; A fuzzy rule base is constructed by fusing the multi-factor correlation and the expert practical experience, and the gravity method is used to defuzzify to obtain the preliminary concentration: ； wherein: U i represents the membership of the rule of item i C i represents the reference concentration of the rule of item i m represents the number of activated rules.​​

Images