Intelligent control method and system for field environment of rice breeding based on internet of things

Abstract

The present application belongs to the field of agricultural environment intelligent regulation technology, and relates to a rice breeding field environment intelligent regulation method and system based on the Internet of Things. The method obtains an environment state vector by acquiring temperature, humidity and soil water content, calculates a canopy light transmission ratio by acquiring canopy light intensity, and obtains a stage switching flag by comparing the light transmission ratio jump amount with a stage switching threshold. An effect matrix is used to predict the environmental impact of each candidate action, and the optimal action is selected by weighted deviation. The sign of the prediction residual is obtained by counting the proportion of the same direction, and the residual bias rate is obtained by comparing the bias threshold. When both flags are true, the effect matrix is updated with a fast step length, otherwise it is updated with a slow step length. The present application eliminates the multi-actuator coupling oscillation caused by the change of canopy structure, and maintains the precision of collaborative control in the whole growth cycle.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agricultural environmental control technology, and particularly relates to a method and system for intelligent control of rice breeding field environment based on the Internet of Things. Background Technology

[0002] Rice breeding is a crucial link in ensuring food security. Breeding greenhouses provide suitable growing conditions for different rice varieties by artificially controlling environmental parameters such as temperature, humidity, and soil moisture content. The entire rice breeding process spans multiple growth stages, including tillering, jointing, and heading. The target values ​​for environmental parameters differ at each stage, and the canopy structure changes significantly with the growth process, placing high demands on the precision and adaptability of environmental control within the greenhouse.

[0003] Currently, breeding greenhouses generally adopt an independent control method based on univariate feedback. This means that a target value and control parameter for a single environmental indicator are set for each type of actuator, and each actuator adjusts independently according to its assigned environmental indicator. For example, fans adjust their speed independently based on temperature deviation, irrigation pumps adjust their speed independently based on soil moisture content deviation, and shading nets adjust their speed independently based on light intensity. There is no information exchange or collaborative decision-making between the actuators.

[0004] However, there is a physical coupling relationship between temperature, humidity, and soil moisture content inside the greenhouse, and the actions of each actuator simultaneously affect multiple environmental variables. The aforementioned independent control method ignores the coupling effect between actuators. When the rice enters the heading stage, the canopy closes rapidly, and the actual dehumidification capacity of the fans is weakened due to the canopy blocking airflow. However, the control logic still operates according to the original parameters, causing the irrigation pump to repeatedly start to compensate for the humidity deviation. This results in coupled oscillations between the actuators, causing periodic fluctuations in field temperature and humidity, which affects the heading and grain filling rate of the rice. Summary of the Invention

[0005] The purpose of this invention is to propose an intelligent control method and system for rice breeding field environment based on the Internet of Things, in order to solve the technical problem that multiple actuators in the above-mentioned breeding greenhouses oscillate due to neglecting the coupling effect and cannot adaptively adjust with changes in the canopy structure.

[0006] To address the above problems, the technical solution of the intelligent control method for rice breeding field environment based on the Internet of Things proposed in this invention is as follows: The IoT-based intelligent control method for rice breeding field environment includes the following steps: The system acquires the temperature, humidity, and soil moisture content inside the greenhouse to obtain an environmental state vector, and also acquires the light intensity at the top and bottom of the canopy. Based on these measurements, the canopy transmittance is calculated, representing the proportion of light intercepted by the canopy. The difference between the canopy transmittance and the transmittance at a preset time interval is used to obtain the transmittance jump. The absolute value of the transmittance jump is compared to a stage switching threshold to obtain a stage switching flag. Finally, an effect matrix is ​​used to predict the predicted environmental state vector after each candidate action is executed. The effect matrix characterizes the coupling effect of each actuator's gear change on each environmental variable. The algorithm calculates the weighted deviation between each predicted environmental state vector and the target state vector, selects the candidate action with the smallest weighted deviation as the optimal action, and executes it. It then calculates the difference between the measured change after executing the optimal action and the predicted change corresponding to the effect matrix, obtaining the prediction residual. The algorithm statistically analyzes the proportion of prediction residuals with the same sign in the sliding window to obtain the residual bias rate. It compares the residual bias rate with a bias threshold to obtain a bias flag. When both the stage switching flag and the bias flag are present, a fast step size is used; otherwise, a slow step size is used to update the coupling coefficient of the effect matrix along the normalized gradient direction using the prediction residual.

[0007] This invention utilizes a stage switching flag generated based on the canopy transmittance jump, which can capture changes in canopy structure at the physical level and provide a feedforward signal for updating the effect matrix. A bias flag generated based on the residual bias rate can confirm systematic deviations in the effect matrix at the prediction level. This dual verification mechanism can accurately trigger differentiated updates to the effect matrix. Fast step size updates are used only when both flags are simultaneously valid; slow step size updates are used in other scenarios. This allows the effect matrix to quickly adapt to the new growth environment when significant changes occur in the canopy structure, while maintaining the stability of the effect matrix during the stable growth period. It continuously tracks gradual changes in the canopy, eliminates multi-actuator coupling oscillations caused by canopy structure changes, and maintains the accuracy and stability of environmental regulation throughout the entire growth cycle.

[0008] Furthermore, the formula for calculating the canopy transmittance is as follows:

[0009] In the formula, The canopy transmittance is the light transmittance of the canopy. The light intensity at the top of the canopy. The light intensity at the bottom of the canopy is denoted as .

[0010] Furthermore, the stage switching threshold is determined as follows: the transmittance jump amount of each control cycle during the stable growth period in which no reproductive stage transition has occurred is collected, and the mean and standard deviation of the collected transmittance jump amount are calculated; the sum of the mean and standard deviation by a preset multiple is determined as the stage switching threshold.

[0011] This invention determines the threshold based on measured data in specific planting scenarios, which is consistent with the actual environment of the corresponding greenhouse and the growth characteristics of rice. It avoids misjudgment caused by general thresholds, and can take into account normal fluctuations during the stable growth period. This greatly reduces the probability that normal fluctuations in canopy transmittance are misjudged as changes in growth stage, and ensures the accuracy of stage change marker triggering.

[0012] Furthermore, the weighted bias is calculated as follows: for each environmental variable, the difference between the component of the predicted environmental state vector and the corresponding component in the target state vector is calculated; the square of each difference is multiplied by the corresponding tolerance weight and then summed to obtain the weighted bias; the tolerance weight is the reciprocal of the square of the tolerance of the corresponding environmental variable, and the tolerance is the maximum allowable deviation of the corresponding environmental variable from the target state vector.

[0013] This invention can unify the deviations of environmental variables with different dimensions into directly comparable scalars, while giving higher priority to environmental variables with smaller tolerances in the candidate action selection process. It automatically ensures the control accuracy of the environmental indicators that have the most significant impact on rice growth, without the need for manual setting of weight coefficients, reducing the complexity of parameter setting and making the selection of the optimal action more in line with the physiological needs of rice growth.

[0014] Furthermore, the step of using the effect matrix to predict the predicted environment state vector after each candidate action is executed includes: subtracting the current actuator gear from each actuator gear of the candidate action to obtain the adjustment vector; multiplying the effect matrix with the adjustment vector and then adding it to the environment state vector to obtain the predicted environment state vector corresponding to the candidate action.

[0015] Furthermore, the residual bias rate is calculated as follows: within the sliding window, the number of positive periods with predicted residuals greater than zero and the number of negative periods with predicted residuals less than zero are counted respectively; the ratio of the larger of the positive period number and the negative period number to the total number of periods in the sliding window is used to obtain the residual bias rate.

[0016] This invention focuses only on the positive and negative directions of the prediction residuals, without relying on the absolute size of the residuals. It can effectively distinguish between random disturbances and systematic biases, and is not sensitive to individual extreme outliers, thus reducing the risk of false triggering caused by instantaneous sensor anomalies.

[0017] Furthermore, the bias flag is determined as follows: calculate the residual bias rate for temperature and humidity respectively; when the residual bias rate corresponding to temperature reaches the bias threshold, or the residual bias rate corresponding to humidity reaches the bias threshold, the bias flag is determined to be valid.

[0018] Furthermore, the step of updating the coupling coefficients of the effect matrix with the predicted residuals along the normalized gradient direction includes updating them using the following formula:

[0019] In the formula, The updated coupling coefficients, The coupling coefficients before the update. For fast or slow stride length, For environment variables The predicted residuals For actuator The adjustment amount, To adjust the square norm of the quantity vector, This is a regularization constant; updates are skipped when the square norm of the adjustment vector is zero.

[0020] Furthermore, the fast step size is greater than the slow step size, and both the fast step size and the slow step size are constants greater than 0 and less than 1.

[0021] This invention enables differentiated updates to the effect matrix. The fast step size allows for rapid correction of the coupling coefficient when changes in the canopy structure cause systematic inaccuracies in the effect matrix, enabling the effect matrix to quickly adapt to the new canopy environment. The slow step size allows for minor adjustments to the coupling coefficient during the stable growth period, tracking the gradual growth changes of the canopy while avoiding fluctuations in the effect matrix caused by random disturbances.

[0022] The technical solution of the intelligent control system for rice breeding field environment based on the Internet of Things proposed in this invention is as follows: The IoT-based intelligent control system for rice breeding field environment includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are run by the processor, the IoT-based intelligent control method for rice breeding field environment described in any of the above technical solutions is implemented.

[0023] The beneficial effects of this invention are as follows: This invention can effectively eliminate the coupling oscillations between multiple actuators in rice breeding greenhouses, maintain the stability and precise control of environmental parameters within the greenhouse throughout the entire rice growth cycle, adapt to the changes in environmental coupling characteristics caused by changes in canopy structure at different growth stages of rice, ensure the accuracy and stability of multi-actuator coordinated regulation, avoid the adverse effects of periodic fluctuations in environmental parameters on rice growth, enhance the adaptive ability of environmental regulation during rice breeding, provide a stable and suitable growth environment for rice breeding, and ensure the stable development of rice breeding work. Attached Figure Description

[0024] Figure 1 A flowchart of the IoT-based intelligent control method for rice breeding field environment provided in this embodiment; Figure 2 A comparison curve of the adaptive update process of the coupling coefficient provided in this embodiment. Detailed Implementation

[0025] In one embodiment, the IoT-based intelligent field environment control method for rice breeding provided by this invention is applied to the environmental control scenario of a rice breeding greenhouse. Multiple rice varieties are grown in the greenhouse, which is equipped with temperature and humidity sensors, soil moisture sensors, and light intensity sensors to collect real-time data on temperature, humidity, soil moisture content, and light intensity. The greenhouse is equipped with three types of actuators: fans, irrigation pumps, and shading nets, used to regulate ventilation and heat dissipation, irrigation, and shading and cooling, respectively. Rice breeding spans multiple growth stages, including tillering, jointing, heading, and grain-filling stages, each with significantly different canopy structures: during the tillering stage, the canopy is sparse and has high light transmittance, resulting in significant ventilation and cooling effects from fans; during the heading stage, the canopy is dense and has low light transmittance, obstructing airflow and reducing the actual cooling and dehumidification capabilities of fans, while the indirect impact of irrigation pumps on humidity increases. This change in canopy structure causes the actual influence coefficients of each actuator on temperature, humidity, and soil moisture content to vary with the growth stage. The present invention provides an intelligent field environment control method for rice breeding based on the Internet of Things. It uses an effect matrix to uniformly characterize the coupled influence of each actuator on each environmental variable, and adaptively updates the effect matrix according to changes in canopy structure and prediction deviations during operation, thereby realizing the coordinated control of multiple actuators.

[0026] Specifically, such as Figure 1 As shown in this embodiment, the intelligent control method for rice breeding field environment based on the Internet of Things includes the following steps: S1. Obtain the temperature, humidity and soil moisture content inside the greenhouse to obtain the environmental state vector, and obtain the light intensity at the top of the canopy and the light intensity at the bottom of the canopy.

[0027] In one embodiment, at the beginning of each control cycle, the current temperature and humidity are read from temperature and humidity sensors installed inside the greenhouse, and the current soil moisture content is read from soil moisture sensors. The temperature, humidity, and soil moisture content are then arranged sequentially to form a three-dimensional environmental state vector. Temperature is expressed in degrees Celsius, humidity in percentage, and soil moisture content in percentage. The control cycle is the time interval between two adjacent actuator adjustments, preferably 10 minutes. The control cycle is determined based on the fact that the response time of temperature and humidity in the greenhouse to actuator actions is typically 5 to 8 minutes. A 10-minute control cycle ensures that the effect of the previous actuator action is fully reflected in the environment before the next adjustment is made, avoiding the situation where the effect of the actuator action is overwritten by a new action before the control cycle has stabilized due to an excessively short cycle.

[0028] Simultaneously, light intensity readings are obtained from a light intensity sensor installed at the top of the canopy and from a light intensity sensor installed at the bottom of the canopy. The top light intensity sensor is installed approximately 20 cm above the rice canopy, while the bottom light intensity sensor is installed approximately 10 cm below the ground. The placement of these two sensors ensures that the top light intensity reflects the total amount of light incident on the upper surface of the canopy, while the bottom light intensity reflects the amount of light that penetrates the canopy and reaches the ground.

[0029] It should be noted that the order of the components in the environmental state vector remains fixed throughout the entire operation; that is, the first component is always temperature, the second component is always humidity, and the third component is always soil moisture content. This order corresponds one-to-one with the row indices of the effect matrix, ensuring that the physical meaning of each component is not misaligned when the effect matrix is ​​multiplied by the adjustment vector in the subsequent step S3.

[0030] In addition, this step also obtains the target state vector, effect matrix, and tolerances for each environmental variable. The target state vector is a three-dimensional vector, with three components representing the target values ​​for temperature, humidity, and soil moisture content at the current growth stage. The target state vector is pre-set by breeders based on the current planting strain and growth stage and stored in the controller. For example, during the tillering stage, the target temperature is 28 degrees Celsius, the target humidity is 70%, and the target soil moisture content is 50%; during the heading stage, the target temperature is 26 degrees Celsius, the target humidity is 65%, and the target soil moisture content is 55%.

[0031] The effect matrix is ​​a 3x3 matrix, where row indices correspond to three environment variables and column indices correspond to three types of executors. The effect matrix contains the first... Line number The elements of the column are coupling coefficients, representing the first... Each time the actuator is increased by one level, the first... The expected change of an environmental variable within a control period. For example, the coupling coefficient in the first row and first column of the effect matrix represents the expected change in temperature within a control period when the fan speed increases by one level. If this value is -0.8 degrees Celsius per level, it means that the temperature is expected to decrease by 0.8 degrees Celsius within a control period when the fan speed increases by one level.

[0032] The initial values ​​of the effect matrix were obtained through a calibration period, a specific timeframe (preferably 5 to 7 days) before the greenhouse was officially put into breeding operation, dedicated to measuring the actual impact of each actuator on various environmental variables. During the calibration period, only the setting of a single actuator was adjusted each time, while keeping the settings of other actuators constant. The changes in the response of each environmental variable before and after the change in the actuator's setting were recorded. Specifically, on the first day of the calibration period, the irrigation pump and shade net settings were kept constant, and the fan settings were sequentially adjusted from 0 to 1, from 1 to 2, and from 2 to 3. After each adjustment, a complete control cycle was waited, and the changes in temperature, humidity, and soil moisture content were recorded. On the second and third days of the calibration period, the same single adjustment operation was performed on the irrigation pump and shade net, respectively. The calibration period covered different weather conditions, including sunny, cloudy, and overcast days, to reduce the systematic interference of external meteorological factors on the calibration results. After the calibration period, the gear adjustment of each actuator was used as the independent variable and the response change of each environmental variable was used as the dependent variable. The standard least squares method was used to fit the initial coupling coefficient of each actuator to each environmental variable, and the fitting results were filled into the effect matrix as the initial value.

[0033] Understandably, the purpose of adjusting only a single actuator each time during the calibration period is to decouple the independent effects of each actuator. If multiple actuators are adjusted simultaneously, the changes in the response of each environmental variable are the superposition result of the combined effects of multiple actuators, making it impossible to distinguish the independent contributions of each actuator, leading to confusion in the estimation of the initial coupling coefficient. By adjusting a single actuator, the recorded response change each time only reflects the independent effect of that actuator, and the initial coupling coefficient obtained by the least squares fitting is more accurate. The purpose of covering different weather conditions during the calibration period is that: on sunny days, solar radiation is strong, the temperature baseline inside the greenhouse is higher, and the cooling effect of the fan may differ from that on cloudy days; by repeatedly calibrating under different weather conditions, the result of the least squares fitting is the average effect under each weather condition, reducing the bias of the calibration results under a single weather condition.

[0034] The tolerance for environmental variables represents the maximum allowable deviation of that variable from the target value, set by breeders based on the physiological characteristics of the rice strain. The preferred tolerances are 1 degree Celsius for temperature, 3 percentage points for humidity, and 3 percentage points for soil moisture content. The basis for determining these tolerances is that rice pollen viability is most sensitive to temperature; a temperature deviation exceeding 1 degree Celsius from the target value can have an observable impact on pollen germination rate. Fluctuations in humidity and soil moisture content within a 3 percentage point range have relatively small effects on the vegetative and reproductive growth of rice. The weight of each environmental variable's tolerance is the reciprocal of the square of the corresponding tolerance: temperature has a weight of 1, humidity has a weight of 1 / 9, and soil moisture content has a weight of 1 / 9.

[0035] S2. Calculate the canopy transmittance based on the light intensity at the top and bottom of the canopy. The canopy transmittance is the proportion of light intercepted by the canopy. Subtract the canopy transmittance from the canopy transmittance before the preset time interval to obtain the transmittance jump. Compare the absolute value of the transmittance jump with the stage switching threshold to obtain the stage switching flag.

[0036] In one embodiment, in each control cycle, the canopy transmittance is calculated based on the light intensity at the top and bottom of the canopy obtained in step S1. The canopy transmittance is calculated as follows:

[0037] in, Canopy transmittance, The light intensity at the top of the canopy. Light intensity at the base of the canopy. Canopy transmittance. The physical meaning of canopy transmittance is the proportion of light intercepted by the canopy, that is, the proportion of the total amount of light incident on the upper surface of the canopy that is absorbed and reflected by the canopy leaves and fails to penetrate to the bottom of the canopy. The value of canopy transmittance ranges from 0 to 1. The closer the value is to 1, the denser the canopy and the more light it traps. The closer to 0, the sparser the canopy, and the more light penetrates.

[0038] Understandably, the above formula originates from a simplified application of the classic Beer-Lambert law of light attenuation in plant canopies. Beer-Lambert's law describes the exponential decay of light intensity with path length as light propagates in a homogeneous medium, and is widely used in agronomy to describe the attenuation process of light in crop canopies. Canopy transmittance This directly reflects the leaf area index and spatial distribution of leaves in the canopy: During the tillering stage, rice plants are short and sparse, the canopy's ability to retain light is weak, and the η value is low, usually between 0.3 and 0.5; as rice enters the jointing and heading stages, the plants grow taller, tillers increase, leaves unfold, and the canopy gradually closes. The value gradually increased to between 0.7 and 0.9. The trend of canopy transmittance is highly correlated with the growth process of rice, and therefore can be used as a physical indicator for detecting the transition of growth stages.

[0039] After calculating the canopy transmittance ratio for the current control cycle, the difference between the current canopy transmittance ratio and the canopy transmittance ratio before a preset time interval is taken to obtain the transmittance jump. The preset time interval is preferably 72 hours, or 432 control cycles. The preset time interval is determined based on the fact that when rice transitions from one growth stage to the next, significant changes in canopy structure typically occur within 48 to 96 hours. 72 hours falls in the middle of this time window, capturing the main changes in canopy transmittance ratio during growth stage transitions. If the preset time interval is too short, diurnal fluctuations in canopy transmittance ratio may be misinterpreted as growth stage transitions; if the preset time interval is too long, the signal of growth stage transitions may be diluted.

[0040] The light transmittance jump is calculated as follows: the canopy light transmittance of the current control period minus the canopy light transmittance of 432 control periods prior. A positive light transmittance jump indicates that the canopy has become denser in the past 72 hours; a negative light transmittance jump indicates that the canopy has become sparser in the past 72 hours, which may occur during the late grain-filling stage when leaves age and fall off.

[0041] The absolute value of the increase in light transmittance is compared with the stage switching threshold. When the absolute value of the increase in light transmittance is greater than or equal to the stage switching threshold, the stage switching indicator is established, indicating that the canopy structure has undergone rapid changes in the past 72 hours and may be undergoing a change in the reproductive stage. When the absolute value of the increase in light transmittance is less than the stage switching threshold, the stage switching indicator is not established, indicating that the changes in the canopy structure in the past 72 hours are within the normal slow growth range.

[0042] The stage switching threshold is determined as follows: Light transmittance jumps are collected during each control cycle in the stable growth period before any stage transition. The mean and standard deviation of the collected light transmittance jumps are calculated, and the sum of the mean and standard deviation by a preset multiple is determined as the stage switching threshold. Specifically, during the first complete growth cycle in the greenhouse, breeders mark the start and end times of each growth stage, selecting the stable growth period in the middle of each stage as the data collection window, excluding the 5-day transition period before and after stage switching. Within the stable growth period, the light transmittance jump for each control cycle is calculated, resulting in a sample of light transmittance jumps. The mean and standard deviation of this sample are calculated, and the mean plus three times the standard deviation is determined as the stage switching threshold. The preset multiple is preferably three times, based on the assumption of a normal distribution. The probability that the data falls outside the range of the mean plus or minus three times the standard deviation is approximately 0.3 percentage points, meaning the probability of the light transmittance jump exceeding the stage switching threshold during the stable growth period is extremely low, keeping the false trigger rate at a low level.

[0043] Thus, by calculating the canopy transmittance and comparing the transmittance jump with the stage switching threshold, the switching of rice growth stages can be automatically detected based on the physical and optical characteristics of the canopy.

[0044] S3. Use the effect matrix to predict the predicted environment state vector after each candidate action is executed. The effect matrix represents the coupling coefficient of each actuator gear change to each environmental variable. Calculate the weighted deviation between each predicted environment state vector and the target state vector, and select the candidate action with the smallest weighted deviation as the optimal action and execute it.

[0045] In one embodiment, in each control cycle, based on the current environment state vector, target state vector, and effect matrix obtained in step S1, all candidate actions are evaluated traversally, and the optimal action that makes the environment state closest to the target state is selected. Candidate actions are combinations of target gears for each actuator, with each type of actuator having four working gears: 0, 1, 2, and 3. To control the traversal scale and avoid frequent large adjustments by the actuators, the gear adjustment amount for each type of actuator within a single control cycle is limited to three values: -1, 0, and 1. That is, the target gear of each actuator in a candidate action is only allowed to increase by 1 gear, remain unchanged, or decrease by 1 gear from the current gear. Each of the three types of actuators has three possible values, resulting in a total of 27 candidate actions.

[0046] It should be noted that when an actuator is currently in the highest gear (gear 3), its target gear cannot be gear 4 because it exceeds the physical gear range; similarly, when an actuator is currently in the lowest gear (gear 0), its target gear cannot be -1. During the traversal of candidate actions, actions that violate gear boundary constraints are automatically excluded, so the actual number of candidate actions evaluated may be less than 27.

[0047] For each candidate action, the adjustment vector is obtained by subtracting the current actuator speed from the current actuator speed for each candidate action. This adjustment vector is a three-dimensional vector with three components: the fan speed adjustment, the irrigation pump speed adjustment, and the shade net speed adjustment. Each component has a value of -1, 0, or 1. Multiplying the effect matrix by the adjustment vector yields a three-dimensional predicted change vector. This predicted change vector has three components: the predicted change in temperature, humidity, and soil moisture content after the candidate action is executed. Adding the current environmental state vector to the predicted change vector yields the predicted environmental state vector for the candidate action. This predicted environmental state vector is also a three-dimensional vector with three components: the predicted temperature, humidity, and soil moisture content after the candidate action is executed.

[0048] After obtaining the predicted environmental state vector, the weighted bias corresponding to the candidate action is calculated. The weighted bias is used to measure the overall deviation between the predicted environmental state vector and the target state vector. For each environmental variable, the difference between the component of the predicted environmental state vector and the corresponding component in the target state vector is calculated. The squares of each difference are multiplied by their corresponding tolerance weights and then summed to obtain the weighted bias of the candidate action. The tolerance weight is the reciprocal of the square of the tolerance of the corresponding environmental variable, which has been determined in step S1. The tolerance for temperature is 1 degree Celsius, and the tolerance weight for temperature is 1; the tolerance for humidity is 3 percentage points, and the tolerance weight for humidity is 1 / 9; the tolerance for soil moisture content is 3 percentage points, and the tolerance weight for soil moisture content is 1 / 9.

[0049] Understandably, the tolerance weighting design ensures that environmental variables with smaller tolerances have a larger weight in the weighted bias. In this embodiment, the contribution of a 1-degree Celsius deviation of temperature from the target value to the weighted bias is equivalent to the contribution of a 3-percentage-point deviation of humidity or soil moisture content from the target value to the weighted bias.

[0050] The physical meaning of this weighting method is that environmental variables with smaller tolerances are more sensitive to crop growth and should be given higher priority in the selection process. In rice breeding, temperature has the most significant impact on pollen viability and seed setting rate; a temperature deviation of 1 degree Celsius can have an observable effect on pollen germination rate. Therefore, temperature receives the highest weight to ensure that the selection of the optimal action prioritizes the control accuracy of temperature indicators. Compared with the method of directly using equal-weighted summation, the tolerance-based weighting method does not require breeders to manually set the weight coefficients of each environmental variable. Only the tolerance of each environmental variable needs to be set to automatically generate a reasonable weight allocation, reducing the complexity of parameter setting.

[0051] For example, suppose the current environmental state vector is 29 degrees Celsius, humidity 72%, and soil moisture 48%, and the target state vector is 28 degrees Celsius, humidity 70%, and soil moisture 50%. The current fan speed is set to level 1, the irrigation pump speed is set to level 1, and the shade net speed is set to level 1. For candidate actions such as increasing the fan speed to level 2, increasing the irrigation pump speed to level 2, and keeping the shade net at level 1, the adjustment vector is 1, 1, and 0. Assuming the predicted change vector obtained by multiplying the effect matrix by this adjustment vector is a temperature decrease of 0.8 degrees Celsius, a humidity increase of 1.5%, and a soil moisture increase of 1.2%, then the predicted environmental state vector is 28.2 degrees Celsius, 73.5% humidity, and 49.2% soil moisture. The temperature deviation is 0.2 degrees Celsius, the humidity deviation is 3.5%, the soil moisture deviation is -0.8%, and the weighted deviation is 1.47.

[0052] After calculating the weighted deviation for each candidate action, the candidate action with the smallest weighted deviation is selected as the optimal action. The target gear for each actuator corresponding to the optimal action is sent to the controllers of the fan, irrigation pump, and shading net. After receiving the target gear, the actuator adjusts to the corresponding gear and maintains that gear throughout the entire control cycle. At the same time, the predicted environmental state vector and adjustment vector corresponding to the optimal action are recorded for use in step S4 when calculating the prediction residual.

[0053] In another embodiment, when multiple candidate actions have the same weighted deviation and are all minimum values, the candidate action with the smallest sum of absolute values ​​of each component of the adjustment vector is selected first, that is, the action with the smallest actuator gear adjustment range is selected first, so as to reduce unnecessary adjustment of the actuator and extend the service life of the actuator.

[0054] In this way, the environmental state after the execution of each candidate action is predicted by the effect matrix, and the optimal action is selected by weighted deviation as the selection criterion. This realizes the coordinated decision-making of three types of actuators: fans, irrigation pumps and shade nets. This allows the actuator action combination in each control cycle to comprehensively consider the coupled effects of each actuator on temperature, humidity and soil moisture content, and avoids the mutual interference caused by ignoring the coupling effect when each actuator is controlled independently.

[0055] S4. Calculate the difference between the measured change after executing the optimal action and the predicted change corresponding to the effect matrix to obtain the prediction residual; calculate the proportion of the predicted residuals with the same sign within the sliding window to obtain the residual bias rate; compare the residual bias rate with the bias threshold to obtain the bias flag.

[0056] In one embodiment, after the optimal action in step S3 is completed and a full control cycle has passed in each control cycle, temperature, humidity, and soil moisture content are read again from the temperature and humidity sensors and soil moisture sensors, and arranged sequentially to form a measured environmental state vector. The measured environmental state vector reflects the actual state of the greenhouse environment after the optimal action is actually executed, while the predicted environmental state vector calculated in step S3 reflects the expected estimate of the environmental state after the optimal action is executed by the effect matrix. The difference between the measured change after executing the optimal action and the predicted change corresponding to the effect matrix is ​​calculated to obtain the prediction residual. Specifically, the measured change is the difference between the measured environmental state vector and the environmental state vector before the optimal action is executed; the predicted change is the result of multiplying the effect matrix by the adjustment vector recorded in step S3; and the prediction residual is the difference between the measured change and the predicted change. The prediction residual is a three-dimensional vector, with three components: the prediction residual for temperature, the prediction residual for humidity, and the prediction residual for soil moisture content. The sign of each component reflects the direction of the effect matrix's prediction deviation from that environmental variable: if the temperature prediction residual is positive, it means that the measured temperature change is greater than the predicted temperature change, that is, the effect matrix underestimates the actual temperature change during the control period; if the temperature prediction residual is negative, it means that the measured temperature change is less than the predicted temperature change, that is, the effect matrix overestimates the actual temperature change.

[0057] It should be noted that the prediction residuals of a single control period may be affected by transient disturbances, such as a sudden gust of wind causing a brief drop in the greenhouse temperature, resulting in a large negative value in the temperature prediction residuals for that period. However, this does not necessarily mean that the effect matrix itself has a systematic bias. Therefore, the inaccuracy of the effect matrix cannot be judged solely based on the prediction residuals of a single control period. Statistical analysis of the prediction residuals over multiple control periods is required to distinguish between random disturbances and systematic biases.

[0058] To achieve the above statistical analysis, after calculating the predicted residuals in each control period, the predicted residual vector is stored in the predicted residual history queue. The predicted residual history queue adopts a first-in, first-out sliding window structure, with the window length preferably being the most recent 36 control periods. The determination of the window length of 36 control periods is based on the following: under the condition of a control period of 10 minutes, 36 control periods correspond to 6 hours, covering 1 / 4 of the main daily variation cycle of environmental variables in the greenhouse. Selecting 6 hours as the statistical window can accumulate a sufficient sample size to smooth the impact of random disturbances, while avoiding the inclusion of outdated residual information in the statistics due to an excessively long window, which would lead to a lag in the judgment of the current bias state. In actual operation, if the control period is adjusted to other durations, the window length should be adjusted accordingly to maintain the statistical window covering a time span of approximately 6 hours.

[0059] After accumulating 36 predicted residual vectors within the sliding window, the proportion of predicted residuals with the same sign within the sliding window is calculated to obtain the residual bias rate. The residual bias rate is calculated as follows: within the sliding window, the number of periods with positive and negative predicted residuals are counted separately, and the larger of these two values ​​is taken as the ratio to the total number of periods in the sliding window. Specifically, for a given environmental variable, the predicted residuals are iterated through the 36 control periods within the sliding window. The number of periods with predicted residuals greater than zero is recorded as the positive period number, and the number of periods with predicted residuals less than zero is recorded as the negative period number. Periods with predicted residuals exactly equal to zero are not included in the positive or negative period number. The larger of the positive and negative period numbers is divided by the total number of periods in the sliding window, 36, to obtain the residual bias rate for that environmental variable.

[0060] In essence, the physical meaning of residual bias ratio is the degree to which the predicted residuals are concentrated in the same direction within a sliding window. If the effect matrix prediction has no systematic bias, the predicted residuals in each period are affected by random disturbances, and the positive and negative directions should be roughly evenly distributed, with the number of positive periods and negative periods each accounting for approximately half, and the residual bias ratio should be close to 0.5. If the effect matrix has a systematic bias, the predicted residuals will continue to be biased in the same direction, with the number of positive or negative periods significantly exceeding the other, and the residual bias ratio will be significantly greater than 0.5. The residual bias ratio ranges from 0.5 to 1, with values ​​closer to 1 indicating a more severe bias. Compared to directly calculating the mean of the predicted residuals, the residual bias ratio only focuses on the sign of the predicted residuals and not their absolute magnitude, thus it is not sensitive to individual extreme outliers. For example, if a sensor momentary malfunction generates a very large positive prediction residual in a certain control cycle, this outlier will significantly increase the mean of the prediction residual, but will only increase the number of positive cycles by 1, and will have a negligible impact on the residual bias rate.

[0061] The bias flag is determined as follows: calculate the residual bias rate for temperature and humidity respectively. When either the residual bias rate corresponding to temperature or the residual bias rate corresponding to humidity reaches the bias threshold, the bias flag is established; when neither the residual bias rate corresponding to temperature nor the residual bias rate corresponding to humidity reaches the bias threshold, the bias flag is not established.

[0062] It should be noted that the bias indicator is determined only for the two environmental variables of temperature and humidity, excluding soil moisture content. This is because, in rice breeding greenhouses, fans and shade nets have no direct impact on soil moisture content; only irrigation pumps directly regulate it. The predicted residuals for soil moisture content primarily reflect the deviation of the coupling coefficient of a single actuator of the irrigation pump, rather than a systematic misalignment of the multi-actuator coupling relationship. Temperature and humidity, however, are simultaneously affected by the coupling of fans, irrigation pumps, and shade nets. When changes in the canopy structure cause a systematic shift in the multi-actuator coupling relationship, the predicted residuals for temperature and humidity first exhibit a bias characteristic with the same sign. Therefore, using the residual bias rates of temperature and humidity as the basis for determining the bias indicator can more accurately capture the systematic misalignment signal of the multi-actuator coupling relationship.

[0063] The bias threshold is determined based on the statistical inference principle of the binomial distribution. Under the null hypothesis that the effect matrix prediction has no systematic bias, the probability of the prediction residual being positive or negative in each control period is 0.5. The number of positive periods in the 36 control periods within the sliding window follows a binomial distribution with parameters 36 and 0.5. Under this distribution, the probability that the larger of the positive or negative period number exceeds a certain threshold can be accurately calculated. The preferred bias threshold is 0.72, corresponding to a larger of 26 or more positive or negative periods in the 36 periods of the sliding window. Under the binomial distribution, the probability of achieving 26 or more successes in 36 independent trials is approximately 1 percentage point. That is, under the condition of no bias in the effect matrix, the probability of the residual bias rate reaching 0.72 is approximately 1 percentage point, keeping the false trigger rate at a low level.

[0064] For example, if, within the 36 control cycles of the sliding window, the number of cycles with positive temperature prediction residuals is 28 and the number of cycles with negative residuals is 8, then the temperature residual bias rate is approximately 0.78, which is greater than the bias threshold of 0.72, and the bias flag is established. If the number of cycles with positive temperature prediction residuals is 20 and the number of cycles with negative residuals is 16, then the temperature residual bias rate is approximately 0.56, which is less than the bias threshold of 0.72; simultaneously, if the humidity residual bias rate is also less than 0.72, then the bias flag is not established.

[0065] Understandably, the bias flag and the stage switching flag in step S2 reflect the risk of inaccuracy in the effect matrix from different dimensions. The stage switching flag detects whether the canopy structure has changed rapidly at the physical level, acting as a feedforward signal; the bias flag detects whether there is a systematic deviation in the output of the effect matrix at the prediction level, acting as a feedback signal. The combination of the two enables the effect matrix update strategy in step S5 to provide early warning when the canopy structure changes and to confirm the prediction deviation in a timely manner when it actually occurs, forming a dual verification mechanism that combines feedforward and feedback.

[0066] In another embodiment, the bias threshold can be dynamically adjusted over time. When the sliding window length changes due to adjustments in actual operating conditions, the bias threshold should be recalculated based on the binomial distribution of the corresponding parameters to maintain the false trigger rate at approximately 1 percentage point.

[0067] Thus, by comparing the measured change after the optimal action is executed with the predicted change to calculate the prediction residual, and by statistically analyzing the proportion of the predicted residuals with the same sign within the sliding window to obtain the residual bias rate, it is possible to distinguish between random disturbances and systematic biases in a robust manner against outliers. This provides a data-level trigger signal for the adaptive update of the effect matrix in step S5, and together with the stage switching flag in step S2, constitutes a dual verification condition for the update of the effect matrix.

[0068] S5. When the stage switching flag and the bias flag are both established, a fast step size is used; otherwise, a slow step size is used to predict the coupling coefficient of the effect matrix by updating the residual along the normalized gradient direction.

[0069] In one embodiment, in each control cycle, based on the stage switching flag obtained in step S2 and the bias flag obtained in step S4, the current inaccuracy state of the effect matrix is ​​determined, and the update method of the effect matrix is ​​determined accordingly. The effect matrix update method is divided into two types: fast step size update and slow step size update. These two methods have different triggering conditions, different update magnitudes, and are applicable to different scenarios.

[0070] A fast step size is used when both the stage switching flag and the bias flag are present; otherwise, a slow step size is used. Specifically, the trigger condition for fast step size updates is that both the stage switching flag and the bias flag are present. When the stage switching flag is present, it indicates that the canopy structure has changed rapidly within the past 72 hours; when the bias flag is present simultaneously, it indicates that the effect matrix prediction has developed a systematic bias. The simultaneous presence of both flags means that the rapid change in the canopy structure has actually led to inaccuracies in the effect matrix, and in this case, the effect matrix needs to be updated with a fast step size to allow it to adapt to the new canopy structure as quickly as possible.

[0071] The triggering condition for slow step size updates is any other situation where neither the stage switching flag nor the bias flag is simultaneously met. Specifically, this includes three scenarios: First, if neither the stage switching flag nor the bias flag is met, it indicates that the canopy structure has not changed rapidly, the effect matrix prediction has not shown systematic bias, and the effect matrix as a whole is accurate. However, within the same growth stage, the canopy structure is still slowly growing and changing, requiring slight adjustments with a slow step size to track the gradual changes in the canopy. Second, if the stage switching flag is met but the bias flag is not met, it indicates that although the canopy structure has changed rapidly, the effect matrix prediction has not yet shown systematic bias. In this case, fast step size updates are not triggered to avoid introducing unnecessary large parameter disturbances while the effect matrix remains accurate. Third, if the stage switching flag is not met but the bias flag is met, it indicates that the effect matrix has shown systematic bias, but this bias is not caused by rapid changes in the canopy structure. It may originate from sensor drift or other factors such as persistent abnormalities in external meteorological conditions. In this case, fast step size updates are also not triggered, and the bias is gradually eliminated through continuous slight adjustments with a slow step size. This dual verification mechanism ensures that fast step size updates are only initiated when both changes in canopy structure and prediction bias are confirmed, reducing the risk of false triggering.

[0072] After determining the update method, the coupling coefficients of the effect matrix are updated along the normalized gradient direction using the predicted residuals. The update formula is:

[0073] In the formula, The updated coupling coefficients, The coupling coefficients before the update. For fast or slow stride length, For environment variables The predicted residuals For actuator The adjustment amount, To adjust the square norm of the quantity vector, This is a regularization constant; updates are skipped when the square norm of the adjustment vector is zero.

[0074] Specifically, The first in the effect matrix Line number The coupling coefficient of the column in the first... The value of each control cycle characterizes the actuator. Environmental variables change with each level increase The expected amount of change generated within a control period. The first calculated in S4 The prediction residuals of environmental variables, i.e., the difference between the measured change and the predicted change on the 1st... The difference between the components. The first adjustment vector recorded in step S3 Each component, i.e., the actuator In the Gear adjustment amount within a control cycle. The square norm of the adjustment vector is equal to the sum of the squares of the adjustment amounts at each actuator position. This is a regularization constant, its function being to prevent the denominator from approaching zero when the square norm of the adjustment vector is close to zero, thus preventing abnormal amplification of the update amount. A value of 0.01 is preferred. This value is determined based on the following: each component of the adjustment vector takes the values ​​of -1, 0, or 1. When only one actuator adjusts to one level, the square norm of the adjustment vector is 1. Setting it to 0.01 represents only 1% of the square norm, and its impact on normal updates is negligible. When all executors are not adjusted, the square norm of the adjustment vector is zero, and updates are skipped directly. It is not included in the calculation.

[0075] It should be noted that when the square norm of the adjustment vector is 0, that is, all actuators have not been adjusted in this control cycle, the prediction residual of this cycle is entirely caused by external disturbances and is unrelated to the coupling coefficient of the effect matrix. Therefore, the update is skipped and no modification is made to the effect matrix.

[0076] Understandably, the mathematical structure of the above update formula is the normalized gradient descent method. Compared to the simple gradient descent method, the normalized gradient descent method introduces the square norm of the adjustment vector into the denominator, decoupling the update amount from the magnitude of the adjustment amount. In the simple gradient descent method, if multiple actuators adjust simultaneously within a certain control cycle, all components of the adjustment vector are non-zero, and the absolute values ​​of the elements of the residual outer product are large, resulting in an overly large update amount; while in a control cycle with only one actuator adjusting, the update amount is underlying. This inconsistency makes the update speed of the effect matrix dependent on exactly how many actuators adjust simultaneously in each cycle, introducing unnecessary randomness. The normalized gradient descent method standardizes the update amount to the correction amount corresponding to a unit adjustment amount by dividing by the square norm of the adjustment vector. Regardless of how many actuators adjust simultaneously in a cycle, the correction magnitude of each unit adjustment amount to the coupling coefficient remains consistent, making the convergence process of the effect matrix more stable.

[0077] In this embodiment, the fast step size is preferably 0.08, and the slow step size is preferably 0.01. The fast step size is greater than the slow step size, and both the fast and slow step sizes are constants greater than 0 and less than 1. The fast and slow step sizes are determined based on the following criteria: With slow step size updates, the effect matrix remains accurate. The slow step size needs to be sufficiently small to avoid unnecessary fluctuations in the effect matrix caused by random perturbations in a single cycle. Taking a temperature prediction residual standard deviation of 0.3 degrees Celsius as an example, with a slow step size of 0.01 and only one actuator adjusting by one level, the square norm of the adjustment vector is 1. The maximum correction to a certain coupling coefficient in the effect matrix in a single cycle is approximately 0.003. This correction is less than 1% compared to the typical value of 0.8 degrees Celsius per level for the coupling coefficient, and will not affect the stability of the effect matrix. After accumulating over 100 control cycles, slow step size updates can adjust the coupling coefficient to approximately 0.3, which is sufficient to track the coupling coefficient drift caused by slow canopy growth within the same growth stage.

[0078] Under the fast step size update, the effect matrix has shown systematic biases that need to be corrected within a short time. A fast step size of 0.08 is used, which is eight times the slow step size. Taking a 2 percentage point continuous shift in humidity prediction residuals as an example, with a fast step size of 0.08 and only one actuator adjusting by one level, the correction amount for the fan-humidity coupling coefficient in the effect matrix per cycle is approximately 0.16 percentage points per level. After 10 to 15 control cycles, the coupling coefficient can be adjusted to 1.6 to 2.4 percentage points per level, essentially eliminating the systematic bias caused by the change in reproductive stage. The fast step size should not be too large; otherwise, the effect matrix may oscillate due to over-correction. The value of 0.08 strikes a balance between correction speed and stability. This value was determined through simulation verification on historical planting data. In the simulation, the number of control cycles required for the effect matrix to recover from the inaccurate state to the stable state and the maximum overshoot during the recovery process were tested under the conditions of fast step size of 0.05, 0.08, 0.10 and 0.15 respectively. When the fast step size is 0.08, the recovery cycle is about 12 control cycles and the maximum overshoot is controlled within 5% of the true value of the coupling coefficient, which is the best overall performance.

[0079] After the effect matrix is ​​updated, it needs to be subject to reasonableness constraints to prevent the coupling coefficients from exceeding physically reasonable ranges due to cumulative updates. The reasonableness constraint method is to truncate each coupling coefficient in the updated effect matrix with upper and lower limits. Specifically, the reasonable range for the fan's coupling coefficient to temperature is -2 degrees Celsius per level to -0.1 degrees Celsius per level. This range is determined based on the fact that the fan promotes evaporative cooling by accelerating airflow, and its influence on temperature is always negative; that is, the temperature decreases when the fan speed increases. In actual greenhouses, increasing the fan speed by one level can reduce the temperature by no more than 2 degrees Celsius within a 10-minute control cycle, and at least 0.1 degrees Celsius. If the updated fan's coupling coefficient to temperature exceeds this range, it is truncated to the nearest boundary value. The reasonable ranges for other coupling coefficients are similarly determined based on the physical direction of each actuator's effect on each environmental variable and its actual adjustment capability. The reasonable range for the coupling coefficient between the irrigation pump and humidity is 0.5 percentage points to 4 percentage points per increment; the reasonable range for the coupling coefficient between the irrigation pump and soil moisture content is 0.3 percentage points to 3 percentage points per increment; and the reasonable range for the coupling coefficient between the shading net and temperature is -1.5 degrees Celsius to -0.05 degrees Celsius per increment. For actuator-environment variable pairs that do not have a physical direct coupling relationship, the reasonable range for their coupling coefficient is set to -0.5 to 0.5, allowing for weak indirect coupling effects but limiting their magnitude.

[0080] In another embodiment, after a fast step-length update is triggered, the triggering conditions for the fast step-length update may be met for multiple consecutive control cycles before the effect matrix stabilizes. To avoid effect matrix oscillations caused by continuous large updates during the fast step-length update process, a maximum number of consecutive triggers for the fast step-length update is set, preferably 20 times. When the number of consecutive triggers for the fast step-length update reaches 20, even if the stage switching flag and the bias flag are still simultaneously set, a forced switch to a slow step-length update is initiated. Normal mode judgment logic is restored only after the bias flag is eliminated. The maximum number of consecutive triggers of 20 is determined as follows: with a fast step-length of 0.08, the cumulative correction amount over 20 control cycles is sufficient to cover the typical change range of the coupling coefficient during the fertility stage switching. If the bias flag is still not eliminated after 20 cycles, it indicates that the deviation may not be caused by changes in the canopy structure, and continuing the fast step-length update may be counterproductive.

[0081] like Figure 2As shown in the figure, this diagram visually demonstrates the superiority of the dual-step switching mechanism of this invention in terms of mathematical convergence. The horizontal axis represents the control period, and the vertical axis represents a coupling coefficient in the effect matrix. The diagram compares the parameter update performance of the method of this invention with that of the traditional independent control method when encountering changes in the growth stage. It can be seen from the figure that a stage change event occurs near the 10th control period, i.e., the rice growth stage changes, causing a significant change in the canopy structure, and the actual coupling relationship in the physical environment changes accordingly. At this time, the parameters of the independent control method remain fixed and cannot adapt to the change in the physical environment, which will inevitably lead to coupling oscillations and a decrease in control accuracy between multiple actuators. In contrast, this invention, by detecting the stage switching flag and the bias flag, promptly triggers a fast step size update after the stage change occurs. It can be clearly seen from the figure that the solid line rises rapidly after the event occurs, causing the coupling coefficient to quickly approach the new true value in a short time, achieving rapid parameter convergence. Subsequently, when the system deviation decreases and the condition of simultaneous occurrence of both flags is no longer met, the system seamlessly switches to a slow step size for micro-tracking adjustments. During the slow step length phase, the curve remains flat, and smooth corrections are made for the gradual changes in the canopy. This ensures efficient convergence speed in response to abrupt changes while effectively avoiding excessive overshoot and oscillation of parameters.

[0082] Understandably, the adaptive update of the effect matrix forms a complete closed-loop feedback mechanism. Step S3 uses the effect matrix to predict and select the optimal action; step S4 detects the prediction residual after the optimal action is executed and calculates the residual bias rate; step S5 determines the fast or slow step size based on the stage switching flag and the bias flag to update the coupling coefficient of the effect matrix along the normalized gradient direction of the predicted residual. The updated effect matrix is ​​then used for the next round of prediction in step S3. This closed loop enables the effect matrix to continuously track the real coupling relationship between the actuator and the environment inside the greenhouse. Regardless of changes in the canopy structure, the effect matrix can automatically adjust to a state that matches the current physical environment within a limited control cycle. In the actual scenario of rice breeding, when moving from the tillering stage to the jointing stage, the canopy height increases and the number of leaves increases, the ventilation path of the fan is more obstructed, and the actual influence coefficient of the fan on temperature and humidity decreases; when moving from the jointing stage to the heading stage, the canopy becomes more dense, the water vapor generated by the irrigation pump evaporation stays in the canopy for a longer time, and the indirect influence coefficient of the irrigation pump on humidity increases. All of the above changes will be captured by the predicted residuals and reflected in the corresponding coupling coefficients of the effect matrix through normalized gradient updates, so that step S3 always selects the optimal action based on the latest coupling relationship in each control cycle.

[0083] Thus, by verifying the fast or slow step size through both the stage switching flag and the bias flag, the coupling coefficient of the effect matrix is ​​updated with the normalized gradient direction. This enables the effect matrix to quickly adapt to changes in the canopy structure during the growth stage switching, continuously track the gradual changes in the canopy during the stable growth period, maintain the accuracy and stability of multi-actuator collaborative control throughout the entire growth cycle, and eliminate multi-actuator coupling oscillations caused by changes in the canopy structure.

[0084] The present invention also provides an intelligent control system for rice breeding field environment based on the Internet of Things (IoT). The intelligent control system for rice breeding field environment based on the IoT includes a processor and a memory. The memory stores computer program instructions. When the computer program instructions are executed by the processor, the intelligent control method for rice breeding field environment based on the IoT in the above embodiments is implemented.

[0085] The IoT-based intelligent control system for rice breeding field environment also includes other components well-known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent control of field environment in rice breeding based on the Internet of Things, characterized in that, Includes the following steps: The temperature, humidity, and soil moisture content inside the greenhouse are obtained to generate an environmental state vector, and the light intensity at the top and bottom of the canopy is also obtained. The canopy transmittance is calculated based on the light intensity at the top of the canopy and the light intensity at the bottom of the canopy. The canopy transmittance is the proportion of light intercepted by the canopy. The difference between the canopy transmittance and the canopy transmittance before the preset time interval is used to obtain the transmittance jump amount. The absolute value of the transmittance jump amount is compared with the stage switching threshold to obtain the stage switching flag. The effect matrix is ​​used to predict the predicted environment state vector after each candidate action is executed. The effect matrix represents the coupling coefficient of each actuator gear change to each environmental variable. The weighted deviation between each predicted environment state vector and the target state vector is calculated. The candidate action with the smallest weighted deviation is selected as the optimal action and executed. The difference between the measured change after executing the optimal action and the predicted change corresponding to the effect matrix is ​​calculated to obtain the prediction residual; the proportion of the predicted residuals with the same sign within the sliding window is statistically analyzed to obtain the residual bias rate; the residual bias rate is compared with the bias threshold to obtain the bias flag. The residual bias rate is calculated as follows: Within the sliding window, the number of positive periods with a prediction residual greater than zero and the number of negative periods with a prediction residual less than zero are counted respectively. The residual bias rate is obtained by taking the larger of the positive cycle number and the negative cycle number and the total number of cycles of the sliding window. When the stage switching flag and the bias flag are both true, a fast step size is used; otherwise, a slow step size is used to predict the coupling coefficient of the effect matrix by updating the residual along the normalized gradient direction. The coupling coefficients for updating the effect matrix using the predicted residuals along the normalized gradient direction include updating them using the following formula: In the formula, The updated coupling coefficients, The coupling coefficients before the update. For fast or slow stride length, For environment variables The predicted residuals For actuator The adjustment amount, To adjust the square norm of the quantity vector, This is a regularization constant; updates are skipped when the square norm of the adjustment vector is zero. The effect matrix is ​​a 3x3 matrix, where the row indices correspond to three environment variables and the column indices correspond to three types of executors; the effect matrix contains the following information: Line number The elements of the column are coupling coefficients, representing the first... Each time the actuator is increased by one level, the first... The expected change of an environmental variable within a control period.

2. The intelligent control method for rice breeding field environment based on the Internet of Things according to claim 1, characterized in that, The formula for calculating the canopy transmittance is as follows: In the formula, The canopy transmittance is the light transmittance of the canopy. The light intensity at the top of the canopy. The light intensity at the bottom of the canopy is denoted as .

3. The intelligent control method for rice breeding field environment based on the Internet of Things according to claim 1, characterized in that, The stage switching threshold is determined in the following manner: Collect the transmittance jump during each control cycle in the stable growth period when no reproductive stage transition has occurred, and calculate the mean and standard deviation of the collected transmittance jump. The sum of the mean and the standard deviation by a preset multiple is determined as the stage switching threshold.

4. The intelligent control method for rice breeding field environment based on the Internet of Things according to claim 1, characterized in that, The weighted bias is calculated as follows: For each environmental variable, calculate the difference between the components of the predicted environmental state vector and the corresponding components in the target state vector; The weighted bias is obtained by multiplying the squares of each difference by their respective tolerance weights and then summing the results. The tolerance weight is the reciprocal of the square of the tolerance of the corresponding environmental variable, and the tolerance is the maximum deviation of the corresponding environmental variable from the target state vector.

5. The intelligent control method for rice breeding field environment based on the Internet of Things according to claim 1, characterized in that, The method of predicting the predicted environment state vector after each candidate action is executed using the effect matrix includes: The adjustment vector is obtained by subtracting the current actuator gear from the current actuator gear of each candidate action. The effect matrix is ​​multiplied by the adjustment vector and then added to the environmental state vector to obtain the predicted environmental state vector corresponding to the candidate action.

6. The intelligent control method for rice breeding field environment based on the Internet of Things according to claim 1, characterized in that, The method for determining the bias flag is as follows: Calculate the residual bias rates for temperature and humidity separately; When the residual bias rate corresponding to temperature reaches the bias threshold, or the residual bias rate corresponding to humidity reaches the bias threshold, the bias flag is determined to be valid.

7. The intelligent control method for rice breeding field environment based on the Internet of Things according to claim 1, characterized in that, The fast step size is greater than the slow step size, and both the fast step size and the slow step size are constants greater than 0 and less than 1.

8. An intelligent field environment control system for rice breeding based on the Internet of Things, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent control method for rice breeding field environment based on the Internet of Things as described in any one of claims 1-7 is implemented.

Images