Remote intelligent seedling raising control system

Abstract

The invention discloses a remote intelligent seedling raising control system, and relates to the technical field of seedling raising environment regulation and control, and the system comprises a data collection module, a data prediction module, a stage division module, a threshold adjustment module and a control strategy module. The data acquisition module acquires environmental parameters, soil states, plant physiological indexes and equipment control variables through a multi-dimensional sensor matrix; the data prediction module outputs future environmental parameters by using the LSTM-attention mechanism model; the stage division module constructs a state vector based on the image, and divides seedling growth stages through improved K-means clustering; the threshold adjustment module dynamically adjusts parameter thresholds in combination with the growth rate model; the control strategy module adopts model predictive control, and combines a predictive value and a threshold value to generate an optimization regulation and control measure. Accurate sensing, dynamic division and self-adaptive regulation and control in the seedling raising process are achieved, and the seedling raising efficiency and quality are improved.

Description

Technical Field

[0001] This invention relates to the field of seedling environment control technology, specifically a remote intelligent seedling control system. Background Technology

[0002] Remote intelligent seedling raising technology, through automated control of environmental parameters, can significantly improve seedling raising efficiency and seedling quality, and has become an important development direction for modern agriculture. However, existing remote intelligent seedling raising methods still have many shortcomings:

[0003] In existing technologies, data collection is mostly limited to a single environmental parameter, lacking comprehensive collection of soil conditions, plant physiological indicators and equipment control variables, resulting in insufficient data dimensions and an inability to fully characterize the dynamic characteristics of the seedling system.

[0004] In terms of environmental prediction, traditional methods often use simple time series models, which are difficult to capture the complex time dependence between environmental parameters and plant growth, resulting in low prediction accuracy and the inability to predict environmental change trends in advance, leading to regulatory lag.

[0005] In terms of stage division, it often relies on fixed time nodes and does not combine multi-dimensional growth characteristics, resulting in a rough division result that cannot accurately match the environmental requirements of different growth stages.

[0006] The control strategies often employ simple feedback control, which triggers actions solely based on whether the current parameters exceed the threshold. This lack of integration with predictive information and energy consumption optimization can easily lead to frequent equipment start-ups and shutdowns, excessive energy consumption, and insufficient control precision. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a remote intelligent seedling control system, which solves the problems of insufficient data dimensions, prediction lag, and low control accuracy in existing technologies.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a remote intelligent seedling control system, comprising:

[0009] The data acquisition module deploys a multi-dimensional sensor matrix in the seedling area. The multi-dimensional sensor matrix specifically includes a temperature and humidity sensor, a photosynthetically active radiation sensor, a CO2 sensor, a soil EC / PH composite probe, and a miniature dew point water potential meter.

[0010] The data prediction module inputs real-time data collected by the multi-dimensional sensor matrix and corresponding control variables into the trained LSTM, and outputs air temperature T, relative humidity RH, photosynthetically active radiation intensity L, CO2 concentration C, and water potential ψ at several future time points.

[0011] The stage division module deploys a top-mounted RGB camera, combined with LED ring lighting, to automatically capture images of seedling trays daily. A four-dimensional state vector S is constructed from these images, and an improved K-means clustering method is used to divide the seedling growth process into K stages.

[0012] The threshold adjustment module establishes a seedling growth rate model and calculates the dynamic threshold for each environmental parameter based on this model.

[0013] The control strategy module uses Model Predictive Control (MPC) combined with LSTM predictions and parameter thresholds to derive corresponding control measures.

[0014] As a further aspect of the present invention, the real-time data specifically includes air temperature T, relative humidity RH, photosynthetically active radiation intensity L, CO2 concentration C, soil electrical conductivity EC, soil pH, and water potential ψ, with the corresponding control variables being shading net opening degree Sno, fan power Fp, and irrigation amount Iv.

[0015] As a further aspect of the present invention, the specific steps for training the LSTM are as follows:

[0016] Data is collected in real time using a multi-dimensional sensor matrix to form a time series dataset. These represent parameters T, RH, L, C, EC, PH, ψ, Sno, Fp, and Iv, respectively.

[0017] Perform outlier detection, missing value handling, and normalization on the time series dataset D;

[0018] Training samples are generated using a sliding window. The input window is Tin, the output window is Tout, and samples (Xi, yi) are generated, where X... i =[x i-Tin ,x i-Tin-1 ,...,x i-1 ], y i =[x i ,x i+1 ,...,x i+Tout-1 ];

[0019] An LSTM-attention mechanism network was constructed, and the mean squared error (MSE) was selected as the loss function.

[0020] The test data is substituted into the trained LSTM to calculate the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) to evaluate the real-time performance of the trained LSTM model.

[0021] As a further aspect of the present invention, the specific parameters for constructing the LSTM-attention mechanism network include 2 LSTM layers, 128 hidden units per layer, bidirectional LSTM, a dropout rate of 0.2, a cyclic dropout rate of 0.1, and an attention dimension of 64.

[0022] As a further aspect of the present invention, the four-dimensional state vector S specifically includes bud length L1, number of true leaves N, stem diameter D, and water potential ψ.

[0023] As a further aspect of the present invention, the specific steps for dividing the seedling growth process into K stages using improved K-means clustering are as follows:

[0024] The continuous variables L1, D, and ψ in S are standardized using Z-score, and the discrete variable N is encoded using one-hot encoding.

[0025] Based on botanical knowledge and seedling cultivation experience, a preset K value is determined.

[0026] Based on historical high-quality seedling data, the mean of the feature vector for each stage is calculated as the initial centroid.

[0027] The relationship between S and the centroid C is calculated using a weighted combination of Euclidean and Manhattan distances. k The distance is given by the formula: d(S,C) k )=0.6×d euclidean (S,C k )+0.4×d manhattan (S,C k ), where d euclidean Calculate the distance d between continuous variables (L1, D, ψ). manhattan Calculate the distance between discrete variables N;

[0028] Each sample is assigned to the class of the nearest centroid, and the mean of the feature vector of each class is recalculated as the new centroid. If the centroid change rate... If the maximum number of iterations is reached, then stop iterating;

[0029] According to the formula Calculate the sum of squared errors (SSE) for different K values;

[0030] Plot the K-SSE curve and select the K value at the inflection point of the curve as the number of stages to divide the curve.

[0031] As a further aspect of the present invention, high-quality batches with a survival rate ≥95% and uniformity ≥85% are selected from historical seedling data. Temperature T, humidity RH, light L, CO2 concentration, and water potential ψ are extracted for each stage. For each parameter, the mean ±1.5σ is calculated as the optimal range for each stage. After each high-quality batch of seedlings is completed, the mean and σ of each stage parameter need to be recalculated.

[0032] As a further aspect of the present invention, according to the formula The changes in seedling shoot length over time were calculated, and then... The ratio divides each growth stage into three relative periods:

[0033] like This is the growth initiation period;

[0034] like This is the rapid growth period;

[0035] like This is the period of growth stagnation;

[0036] Where Ke is the maximum theoretical shoot length, r is the growth rate parameter, t0 is the growth inflection point time, and t is the growth time.

[0037] As a further aspect of the present invention, for each environmental parameter x, according to formula x target =x base The threshold of this parameter is calculated using the formula ×(1-α×Δr), where x base The parameter is the base threshold, α is the correction coefficient, and Δr is the current growth rate r. 实测 The proportion of deviation from the expected value r, according to the formula To obtain.

[0038] As a further aspect of the present invention, the specific steps for using model predictive control (MPC) for corresponding regulation are as follows:

[0039] Compare the predicted environmental parameter value x from the LSTM model with the corresponding parameter threshold x. target Compare;

[0040] Filter out | xx target The parameter |>ε is selected, and max{|xx} is chosen. target The parameters corresponding to |};

[0041] Define the prediction time domain N p With control time domain N c ;

[0042] Define the objective function J for tracking error and energy consumption, with the specific formula as follows: Where, x tx is the predicted value of the environmental parameters at time t. target For the parameter threshold, u t It is the control quantity at time t, u t-1 It is the control quantity at the previous moment, and ρ is the energy consumption weight;

[0043] Hard constraints are set for the control quantities: shading net opening Sno, fan power Fp, and irrigation volume Iv.

[0044] Transform the objective function J and control constraints into a standard QP problem;

[0045] The optimal control sequence is obtained by calling the QP solver. And extract the first control variable from the sequence. Distribute to the execution device.

[0046] This invention provides a remote intelligent seedling control system, which has the following advantages compared with the prior art:

[0047] (1) By deploying a multi-dimensional sensor matrix, the present invention synchronously collects environmental parameters, soil conditions, plant physiological indicators and equipment control variables, thereby achieving comprehensive coverage of seedling data, overcoming the shortcomings of the single data dimension in the existing technology, providing complete data support for precise regulation, and improving the system's ability to perceive the seedling status.

[0048] (2) This invention combines the LSTM-attention mechanism model with improved K-means clustering to achieve accurate prediction of environmental parameters and dynamic division of seedling growth stages. Combined with a threshold adjustment mechanism based on growth rate, it makes the regulation target highly matched with the real-time needs of seedlings, solving the problem of inaccurate regulation caused by the coarse stage division and fixed threshold in traditional methods.

[0049] (3) The present invention adopts the Model Predictive Control (MPC) strategy, which balances the tracking accuracy of the objective function with the energy consumption of the equipment by optimizing the objective function, and combines rolling optimization to dynamically generate control sequences. This not only ensures the stable control of environmental parameters, but also reduces the invalid actions of the equipment, thereby improving the operating efficiency and economy of the system. Attached Figure Description

[0050] Figure 1 This is the system principle block diagram of the present invention. Detailed Implementation

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] like Figure 1 This invention provides a remote intelligent seedling control system, comprising:

[0053] The data acquisition module deploys a multi-dimensional sensor matrix in the seedling area. The multi-dimensional sensor matrix specifically includes a temperature and humidity sensor, a photosynthetically active radiation sensor, a CO2 sensor, a soil EC / PH composite probe, and a miniature dew point water potential meter.

[0054] Temperature and humidity sensors, photosynthetically active radiation sensors, and CO2 sensors belong to the category of environmental sensor deployment;

[0055] For temperature and humidity sensors, air temperature T and relative humidity RH are collected;

[0056] Temperature (T) reflects the environmental thermal state and affects plant respiration rate, enzyme activity, and transpiration. During seed germination (such as in tomatoes), a constant temperature environment of 25–28°C is often required. Temperature fluctuations exceeding ±3°C will significantly reduce the germination rate.

[0057] RH determines the air-water vapor pressure difference, which directly regulates the opening and closing of stomata. High humidity environments are prone to fungal diseases, while low humidity will lead to excessive transpiration in seedlings and wilting of leaves.

[0058] For the photosynthetically active radiation sensor, the photosynthetically active radiation intensity L is collected;

[0059] Light directly determines the rate of photosynthesis in plants. Different crops have significantly different light saturation points. Light exceeding the light saturation point will lead to light inhibition and waste of energy. On the other hand, the duration of light needs to be precisely controlled during the seedling stage. For example, pepper seedlings need 14 hours of light. Insufficient light will lead to excessive growth of seedlings.

[0060] For a CO2 sensor, the CO2 concentration C in the air is collected;

[0061] C directly affects the carbon assimilation efficiency of plants. Plant photosynthesis consumes CO2. If it is not replenished, the indoor concentration can drop to 300 ppm, which is lower than the atmospheric level of 400 ppm, leading to a decrease in photosynthetic rate. Moreover, most crops have the best photosynthetic efficiency at a CO2 concentration of 800-1000 ppm. Too high a concentration will cause stomatal closure and carbon metabolism imbalance.

[0062] The soil EC / PH composite probe is part of the soil parameter sensor deployment.

[0063] For the soil EC / PH composite probe, soil electrical conductivity (EC) and soil pH are collected.

[0064] EC reflects the salt concentration in the soil solution and indirectly indicates nutrient availability. Excessively high EC can lead to reverse osmosis of water from the roots, manifested as scorching of leaf edges, while excessively low EC reflects insufficient nutrients.

[0065] pH affects the solubility of mineral elements and the absorption capacity of plant roots. Different crops prefer different pH environments. For example, blueberries prefer acidic soil with a pH of 4.3 to 5.3, while tomatoes adapt to neutral soil with a pH of 5.5 to 6.8. A pH deviation from the suitable range can lead to absorption barriers of specific elements.

[0066] Miniature dew point water potential meters are part of the deployment of plant physiological index sensors;

[0067] For a miniature dew point water potential meter, the water potential ψ of plant tissue is collected;

[0068] ψ reflects the water state within the plant, comprehensively reflecting soil water availability, transpiration pull, and root activity. Traditional leaf water potential meters require mature leaves (7-10 days after germination), while miniature dew point water potential meters can measure the water potential of the embryo (before germination), radicle (during germination), and cotyledon (after germination), covering the entire seedling cycle.

[0069] The data prediction module inputs real-time data collected by the multi-dimensional sensor matrix and corresponding control variables into the trained LSTM, and outputs T, RH, L, C, and ψ for the next few hours.

[0070] The aforementioned real-time data specifically refers to T, RH, L, C, EC, PH, and ψ, with the corresponding control variables being the shading net opening degree Sno, the fan power Fp, and the irrigation amount Iv;

[0071] The opening degree of the shade net directly determines the light transmittance, which in turn affects the temperature and transpiration rate. If the opening degree of the shade net is not input, the model cannot distinguish between natural light fluctuations and artificially controlled light, resulting in deviations in temperature prediction.

[0072] Fan power affects airflow speed, directly altering humidity distribution and evaporation efficiency. By inputting fan power, LSTM can predict humidity changes in advance, improving control response speed.

[0073] The amount of irrigation directly determines the moisture content of the substrate, which in turn affects root water absorption and leaf water potential. By inputting the amount of irrigation, the error in leaf water potential prediction can be reduced.

[0074] For example, if the shade net is closed at 14:00 on a certain day (0% opening), the temperature will suddenly rise from 25°C to 32°C. If there is no input on the opening of the shade net, the model will attribute the temperature rise to the increase in the external air temperature and predict that the temperature will continue to rise for several hours. However, if there is an input on the opening of the shade net, the model will identify the causal relationship between the output and the closing of the shade net and predict that the temperature will drop after the shade net is reopened. By adding control variables, the prediction error of sudden temperature changes is reduced.

[0075] The specific steps for training an LSTM model are as follows:

[0076] Data is collected in real time using a multi-dimensional sensor matrix at a frequency of f minutes / time (which can be 10), and combined with corresponding control variables to form a time series dataset. These represent T, RH, L, C, EC, PH, ψ, Sno, Fp, and Iv, respectively. Key events, such as equipment failure and extreme weather, also need to be manually labeled to generate anomaly labels yanomaly for subsequent robustness training.

[0077] Multi-dimensional data covers four elements: environment, soil, plants, and equipment, avoiding the one-sidedness of traditional univariate models. At the same time, anomaly annotation helps the model learn dynamic features under extreme scenarios, such as the response pattern when the shade net is fully opened during a rainstorm, causing a sudden drop in sunlight.

[0078] After obtaining the time series dataset D, it is necessary to perform data cleaning and preprocessing, specifically including outlier detection, missing value handling, and normalization.

[0079] For outlier detection, the interquartile range (IQR) method is used. For each feature, Q3+1.5IQR and Q1-1.5IQR are calculated, and values ​​outside the range are filled with the average of the preceding and following time points.

[0080] The IQR method effectively filters out outliers caused by transient equipment failures, such as spurious data from sudden power surges during wind startup, thus reducing model training errors.

[0081] For handling missing values, if there are ≤3 consecutive missing time points, linear interpolation is used; if there are >3 missing time points, the historical mean of the corresponding growth stage is used to fill the missing values.

[0082] For normalization, temperature and humidity are standardized using Z-score, while light intensity and shade net opening are scaled using Min-Max.

[0083] Feature normalization can better preserve the physical meaning of data. For example, after scaling the light intensity from 0 to 1, the model can easily identify the 0.8 threshold corresponding to the light saturation point.

[0084] Training samples are generated using a sliding window. The input window Tin is defined as 144 (24 hours, 10 minutes / point), and the output window Tout is defined as 12 (for the next 2 hours). Samples (Xi, yi) are generated, where X... i =[x i-Tin ,x i-Tin-1 ,...,x i-1 ], y i =[x i ,x i+1 ,...,x i+Tout-1 ];

[0085] The output features are five dimensions: temperature, humidity, light intensity, CO2 concentration, and leaf water potential. Since soil EC / pH and equipment status are control inputs, they do not need to be predicted.

[0086] The LSTM-attention mechanism network is set up with the following specific parameters:

[0087] LSTM has two layers. While deeper networks can learn more complex time dependencies, the actual network performance needs to be considered.

[0088] The number of hidden units in each layer is 128. The number of hidden units determines the model capacity. When the number of units increases from 64 to 128, the validation set MSE decreases, while when it increases to 256, the risk of overfitting increases.

[0089] Choosing a bidirectional LSTM is beneficial in the seedling cultivation scenario, where the future state of equipment affects the prediction of the current environment. The bidirectional structure allows the model to utilize both historical and future information simultaneously.

[0090] The dropout rate is 0.2, randomly discarding 20% ​​of the input units to prevent overfitting;

[0091] The loop dropout rate is 0.1. Dropout on the loop connection is specifically used to suppress sequence overfitting of LSTM and improve the model's adaptability to sudden weather changes.

[0092] The attention dimension is 64, which controls the computational complexity of the attention weights;

[0093] The attention score uses the dot product score function because the dot product is computationally efficient and can effectively capture the correlation between features.

[0094] The loss function chosen is the mean squared error (MSE), which measures the regression error between the predicted and the true values.

[0095] The test set is substituted into the trained LSTM to calculate the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 R 2 Reflecting the model's ability to interpret data variance, RMSE and MAE provide absolute error references, and the combination of the three comprehensively evaluates model performance.

[0096] The phase division module deploys a top-mounted RGB camera, along with LED ring lighting, to automatically capture images of the seedling trays daily.

[0097] By segmenting the image and performing skeletonization, the pixel length of the main stem is calculated and converted into the actual length by combining the calibration coefficient to obtain the bud length L1, which can reflect the longitudinal growth rate of the seedling and is the core indicator for judging whether growth has stopped.

[0098] The model identifies true leaves based on the YOLOv8 object detection model and outputs the number of leaves (N). It is directly related to the developmental stage, such as N=0 for the germination stage, N=1-2 for the cotyledon stage, and N≥3 for the true leaf stage, which is more accurate than time nodes.

[0099] The stem region was extracted 2 mm below the cotyledon node. The circular outline was detected by Hough transform, and the diameter was calculated to obtain the stem diameter D, which can reflect the seedling vigor and indirectly reflect the seedling stage.

[0100] Using a miniature dew point water potential meter, the leaf tip tissue of the newly unfolded true leaf was selected to measure the osmotic potential and convert it into total water potential ψ. It is an important physiological indicator for measuring the water status of plant leaf cells. During the seedling growth process, there are significant differences in physiological activities and water requirements at different stages. For example, seedlings in the germination period are sensitive to water shortage and the water potential drops slightly; while seedlings in the hardening-off period can tolerate lower water potential after water control training.

[0101] Construct a four-dimensional state vector S = [L1, N, D, ψ], which represents morphological growth, developmental stage, structural robustness, and water physiological state, respectively.

[0102] An improved K-means clustering method is used to divide the growth process into K stages based on S. The specific operation is as follows:

[0103] Z-score standardization is applied to the continuous variables L1, D, and ψ in S, and one-hot encoding is applied to the discrete variable N.

[0104] The dimensions of shoot length L1 and water potential ψ are very different. After standardization, we can avoid large numerical features from dominating the clustering results. The numerical difference of true leaf number N does not represent the size relationship. It is more reasonable to use one-heat encoding.

[0105] Based on botanical knowledge and seedling cultivation experience, a preset K value is used. For example, tomato seedling cultivation can be divided into four stages: germination period (0-3 days), cotyledon period (4-7 days), true leaf period (8-20 days), and hardening-off period (21-28 days).

[0106] Based on historical high-quality seedling data, the mean of the feature vector for each stage is calculated as the initial centroid.

[0107] It overcomes the shortcomings of traditional K-means random initialization, avoids getting trapped in local optima, and accelerates convergence;

[0108] The distance to the centroid is calculated using a weighted combination of Euclidean and Manhattan distances, specifically using the formula: d(S,C) k )=0.6×d euclidean (S,C k )+0.4×d manhattan (S,C k ), where deuclidean Calculate the distance d between continuous variables (L1, D, ψ). manhattan Calculate the distance between discrete variables N;

[0109] Euclidean distance is suitable for capturing the spatial distribution of continuous values, such as differences in water potential, while Manhattan distance is more sensitive to the number of true leaves N in integer encoding. Therefore, the two distances are combined to comprehensively evaluate the distance to the centroid of each stage.

[0110] Each sample is assigned to the class of the nearest centroid, and the mean of the feature vector of each class is recalculated as the new centroid. If the centroid change rate... If the maximum number of iterations is reached, then stop iterating;

[0111] According to the formula Calculate the sum of squared errors (SSE) for different K values;

[0112] Plot the K-SSE curve and select the K value at the inflection point of the curve to avoid overfitting or underfitting due to subjective setting of the K value.

[0113] The clustering results can be mapped to the actual growth stage, and the accuracy of the division can be verified through expert evaluation and cross-validation.

[0114] For each stage that has been divided, high-quality batches with a survival rate of ≥95% and uniformity of ≥85% are selected from historical seedling data, and environmental parameters for each stage are extracted, namely temperature T, humidity RH, light intensity L, CO2 concentration, and water potential ψ.

[0115] Whether seedlings can survive is the most basic test standard for environmental control. If the survival rate of a certain batch is low, it indicates that there may be fatal defects in its environmental parameters, such as low temperature during the cotyledon stage causing frost damage, or drought during the true leaf stage causing wilting. Such data cannot reflect the environmental laws of suitable growth and must be excluded.

[0116] Uniformity reflects the consistency of the growth status of the seedling group. If the uniformity is low, it indicates that there are significant fluctuations in environmental parameters in space or time. For example, local high temperature in the greenhouse causes some seedlings to grow too tall, and local insufficient light causes some seedlings to be stunted. Such data cannot reflect stable and controllable environmental patterns.

[0117] Only by combining the two can the overall quality of the seedlings be reflected. If the survival rate is high but the uniformity is low, it means that the environmental parameters can ensure survival but cannot support uniform growth. Such data lacks value for mass production. If the uniformity is high but the survival rate is low, it means that although the environmental parameters are stable, there is a fatal flaw. Its stability comes at the cost of survival and is not of reference value.

[0118] For each parameter in each stage, the mean ± 1.5σ is calculated as the optimal interval. The mean ± 1.5σ balances strictness and tolerance, avoiding both overly strictness and loose intervals. Here, σ is the standard deviation.

[0119] After each high-quality batch of seedlings is completed, the mean and σ of the parameters at each stage are recalculated, and the optimal range of each parameter at different stages is dynamically adjusted.

[0120] The threshold adjustment module establishes a seedling growth rate model, using the Logistic growth equation to describe the change in seedling shoot length over time. The specific formula is as follows:

[0121]

[0122] Where Ke is the carrying capacity (maximum theoretical shoot length), r is the growth rate parameter, reflecting the degree of vigorous growth, and t0 is the time of the growth inflection point.

[0123] The Logistic equation naturally describes the S-shaped growth curve, which perfectly matches the "slow, fast, slow" growth pattern of seedlings.

[0124] pass The ratio quantifies the relative growth state, dividing each growth stage into three relative periods:

[0125] like This is the growth initiation period;

[0126] like This is the rapid growth period;

[0127] like This is the period of growth stagnation;

[0128] Within the same macroscopic stage (e.g., cotyledon stage), (Just entering this growth stage) and Seedlings about to leave this growth stage have different environmental requirements; the former need high humidity to promote growth, while the latter need humidity to be gradually reduced to prevent excessive growth. This difference can be refined;

[0129] For each environmental parameter x, according to formula x target =x base The threshold of this parameter is calculated using the formula ×(1-α×Δr), where x base The basic threshold for the parameter is determined by the relative period. Let α be the correction coefficient and Δr be the current growth rate r. 实测 The proportion of deviation from the expected value r, according to the formula Seek;

[0130] If Ke decreases due to environmental stress, then the current The ratio will enter the later stage prematurely, triggering threshold adjustment;

[0131] The optimal range is set based on the general laws of seedling growth stages, while the threshold is dynamically adjusted according to the real-time status of the seedlings and environmental fluctuations, which better meets the adaptive requirements.

[0132] For example, suppose the optimal light range for tomatoes during the true leaf stage is [300, 500]. The initial control threshold is set to 280, which is 20 below the lower limit of the optimal range. If real-time monitoring shows that the water potential of the seedling leaves is decreasing, the light control threshold will be dynamically adjusted to 320 to avoid strong light exacerbating transpiration and water loss. This means that supplemental lighting will be started earlier, allowing the light to fluctuate within [320, 500] instead of waiting until 280. At this time, the optimal range remains unchanged, but the threshold is adjusted according to the seedling status, achieving precise seedling protection.

[0133] The control strategy module uses Model Predictive Control (MPC) combined with LSTM predictions and parameter thresholds to derive corresponding control measures.

[0134] Compare the predicted environmental parameter value x from the LSTM model with the corresponding parameter threshold x. target Compare and calculate |xx target |, if|xx target If | > ε, then corresponding adjustments need to be made:

[0135] Temperature and humidity directly affect the transpiration rate and are key inputs for irrigation decisions. If either of them deviates too much from the threshold, the irrigation amount will be adjusted.

[0136] The parameter light intensity determines the photosynthetic rate. If it differs too much from the threshold, the opening of the shade net should be adjusted.

[0137] Regarding the parameter CO2 concentration, it is easy for it to be insufficient in a greenhouse environment due to plant consumption. If the difference from the threshold is too large, the fan power should be adjusted.

[0138] Leaf water potential is a direct indicator of plant physiological state. If it differs too much from the threshold, the irrigation amount should be adjusted.

[0139] Filter out | xx target The parameter |>ε is selected, and max{|xx} is chosen. target The parameters corresponding to |} are adjusted accordingly. The specific steps are as follows:

[0140] Define the prediction time domain N p With control time domain N c N p 60, N can be selected. c 10 is acceptable;

[0141] Predicting time domain N pUsing seconds as the unit, covering the next 10 minutes, one prediction point is generated every second to obtain the environmental parameter change curve within the next 10 minutes;

[0142] In the seedling environment, the effect of equipment regulation on temperature lags by about 3 to 5 minutes. A 10-minute prediction can completely capture the response process from control action to environmental change, avoiding incomplete regulation caused by short time domain.

[0143] Control time domain N c In the 60-step prediction, only the control actions of the first 10 steps are planned, such as the opening degree of the shade net and the power of the fan in the first 1 to 10 seconds. The remaining 50 steps use the control values ​​of the 10th step.

[0144] The 10-step control quantity planning can refine the control details, such as making small adjustments to the shade net in the first 3 seconds and stabilizing the opening in the next 7 seconds, while avoiding the surge in computational load caused by the full 60-step planning.

[0145] Define the objective function J for tracking error and energy consumption, with the specific formula as follows:

[0146] The first term is the tracking error term, x t x is the predicted value of the environmental parameters at time t. target The parameter threshold is the sum of squares and the penalty for deviations from the threshold, ensuring that environmental parameters remain stable within the threshold and avoiding stress on seedlings;

[0147] The second term is the energy consumption penalty term, u t It is the control quantity at time t, u t-1 It is the control quantity of the previous moment, the square of the sum of penalties for drastic changes in the control quantity, to suppress frequent jumps in the control quantity and reduce equipment energy consumption;

[0148] ρ is the energy consumption weight, which can be taken as 0.3;

[0149] Based on the physical limits of the equipment, hard constraints are set on the control quantities. For example, for the shade net opening Sno, 0% ≤ Sno ≤ 100%, the fan power Fp, 0W ≤ Fp ≤ 750W, and the irrigation volume Iv, 0L / h ≤ Iv ≤ 50L / h.

[0150] Hard constraints are the bottom line for the safe operation of equipment, avoiding failures caused by neglecting physical limits in optimization calculations;

[0151] Transform the objective function J and control constraints into a standard QP problem;

[0152] When the objective function is quadratic and the constraints are linear, QP can efficiently solve this type of problem.

[0153] The optimal control sequence is obtained by calling the QP solver. Extract only the first control variable in the sequence Send to the execution device;

[0154] The environment is dynamic and changes, so we retain room for adjusting subsequent control variables. We adapt to new changes through rolling optimization, resolving QP every second to avoid planning N all at once. c The lag caused by the step;

[0155] Repeat the above steps every second, update the LSTM prediction value based on the latest sensor data, reconstruct the objective function and constraints, solve the new optimal control sequence, compare the first control variable of the new sequence with the value of the previous time step, and if the rate of change is less than the rate of change threshold, maintain the original control variable to avoid frequent actions.

[0156] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0157] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A remote intelligent seedling control system, characterized in that, include: The data acquisition module deploys a multi-dimensional sensor matrix in the seedling area. The multi-dimensional sensor matrix specifically includes a temperature and humidity sensor, a photosynthetically active radiation sensor, a CO2 sensor, a soil EC / PH composite probe, and a miniature dew point water potential meter. The data prediction module inputs real-time data collected by the multi-dimensional sensor matrix and corresponding control variables into the trained LSTM, and outputs air temperature T, relative humidity RH, photosynthetically active radiation intensity L, CO2 concentration C, and water potential ψ at several future time points. The stage division module deploys a top-mounted RGB camera, combined with LED ring lighting, to automatically capture images of seedling trays daily. A four-dimensional state vector S is constructed from these images, and an improved K-means clustering method is used to divide the seedling growth process into K stages. The threshold adjustment module establishes a seedling growth rate model and calculates the dynamic threshold for each environmental parameter based on this model. The control strategy module uses Model Predictive Control (MPC) combined with LSTM predictions and parameter thresholds to derive corresponding control measures.

2. The remote intelligent seedling control system according to claim 1, characterized in that, The real-time data specifically includes air temperature T, relative humidity RH, photosynthetically active radiation intensity L, CO2 concentration C, soil electrical conductivity EC, soil pH, and water potential ψ. The corresponding control variables are shading net opening degree Sno, fan power Fp, and irrigation amount Iv.

3. The remote intelligent seedling control system according to claim 1, characterized in that, The specific steps for training an LSTM are as follows: Data is collected in real time using a multi-dimensional sensor matrix to form a time series dataset. These represent parameters T, RH, L, C, EC, PH, ψ, Sno, Fp, and Iv, respectively. Perform outlier detection, missing value handling, and normalization on the time series dataset D; Training samples are generated using a sliding window. The input window is Tin, the output window is Tout, and samples (Xi, yi) are generated, where X... i =[x i-Tin ,x i-Tin-1 ,...,x i-1 ], y i =[x i ,x i+1 ,...,x i+Tout-1 ]; An LSTM-attention mechanism network was constructed, and the mean squared error (MSE) was selected as the loss function. The test data is substituted into the trained LSTM to calculate the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) to evaluate the real-time performance of the trained LSTM model.

4. The remote intelligent seedling control system according to claim 1, characterized in that, The specific parameters for building the LSTM-attention mechanism network include 2 LSTM layers, 128 hidden units per layer, bidirectional LSTM, a dropout rate of 0.2, a recurrent dropout rate of 0.1, and an attention dimension of 64.

5. The remote intelligent seedling control system according to claim 1, characterized in that, The four-dimensional state vector S specifically includes bud length L1, number of true leaves N, stem diameter D, and water potential ψ.

6. The remote intelligent seedling control system according to claim 1, characterized in that, The specific steps for dividing the seedling growth process into K stages using improved K-means clustering are as follows: The continuous variables L1, D, and ψ in S are standardized using Z-score, and the discrete variable N is encoded using one-hot encoding. Based on botanical knowledge and seedling cultivation experience, a preset K value is determined. Based on historical high-quality seedling data, the mean of the feature vector for each stage is calculated as the initial centroid. The relationship between S and the centroid C is calculated using a weighted combination of Euclidean and Manhattan distances. k The distance is given by the formula: d(S,C) k )=0.6×d euclidean (S,C k )+0.4×d manhattan (S,C k ), where d euclidean Calculate the distance d between continuous variables (L1, D, ψ). manhattan Calculate the distance between discrete variables N; Each sample is assigned to the class of the nearest centroid, and the mean of the feature vector of each class is recalculated as the new centroid. If the centroid change rate... If the maximum number of iterations is reached, then stop iterating; According to the formula Calculate the sum of squared errors (SSE) for different K values; Plot the K-SSE curve and select the K value at the inflection point of the curve as the number of stages to divide the curve.

7. The remote intelligent seedling control system according to claim 1, characterized in that, High-quality batches with a survival rate of ≥95% and uniformity of ≥85% are selected from historical seedling data. Temperature T, humidity RH, light intensity L, CO2 concentration, and water potential ψ are extracted for each stage. For each parameter, the mean ±1.5σ is calculated as the optimal range for each stage. After each high-quality batch of seedlings is completed, the mean and σ of each stage parameter need to be recalculated.

8. The remote intelligent seedling control system according to claim 1, characterized in that, According to the formula The changes in seedling shoot length over time were calculated, and then... The ratio divides each growth stage into three relative periods: like This is the growth initiation period; like This is the rapid growth period; like This is the period of growth stagnation; Where Ke is the maximum theoretical shoot length, r is the growth rate parameter, t0 is the growth inflection point time, and t is the growth time.

9. The remote intelligent seedling control system according to claim 1, characterized in that, For each environmental parameter x, according to formula x target =x base The threshold of this parameter is calculated using the formula ×(1-α×Δr), where x base The parameter is the base threshold, α is the correction coefficient, and Δr is the current growth rate r. 实测 The proportion of deviation from the expected value r, according to the formula To obtain.

10. The remote intelligent seedling control system according to claim 1, characterized in that, The specific steps for using model predictive control (MPC) to perform corresponding regulation are as follows: Compare the predicted environmental parameter value x from the LSTM model with the corresponding parameter threshold x. target Compare; Filter out | xx target The parameter |>ε is selected, and max{|xx} is chosen. target The parameters corresponding to |}; Define the prediction time domain N p With control time domain N c ; Define the objective function J for tracking error and energy consumption, with the specific formula as follows: Where, x t x is the predicted value of the environmental parameters at time t. target For the parameter threshold, u t It is the control quantity at time t, u t-1 It is the control quantity at the previous moment, and ρ is the energy consumption weight; Hard constraints are set for the control quantities: shading net opening Sno, fan power Fp, and irrigation volume Iv. Transform the objective function J and control constraints into a standard QP problem; The optimal control sequence is obtained by calling the QP solver. And extract the first control variable from the sequence. Distribute to the execution device.

Images