Salt-tolerant rice breeding salinity intelligent regulation method based on digital twin model

Abstract

This invention discloses an intelligent salinity control method for salt-tolerant rice breeding based on a digital twin model, belonging to the field of salinity control technology in rice breeding. The method includes the following steps: constructing a digital twin model corresponding to a physical breeding salt pond; collecting data in real time and generating standardized feature vectors to drive the digital twin model; inputting the standardized feature vectors into the digital twin model to achieve state synchronization between the digital twin and the physical entity; outputting a predicted salinity value for a future time step based on the updated digital twin model and environmental data; calculating a control command for adjusting salinity based on a preset salinity target value, system state and total disturbance, and the predicted salinity value; and sending the control command to the physical actuator. This invention provides a quantifiable, predictable, and reproducible intelligent environmental management solution for salt-tolerant rice breeding, improving the accuracy and efficiency of breeding experiments.

Description

Technical Field

[0001] This invention relates to the field of salinity regulation technology in rice breeding, specifically to an intelligent salinity regulation method for salt-tolerant rice breeding based on a digital twin model. Background Technology

[0002] In the breeding of salt-tolerant rice, precise control of soil and irrigation water salinity is a key technical challenge, directly affecting the accuracy and efficiency of breeding material screening. Traditional breeding methods rely heavily on manual experience for salinity management, generally suffering from three prominent problems: monitoring lag, extensive regulation, and poor repeatability. Specifically: salinity monitoring often involves periodic field sampling and laboratory chemical analysis, which is time-consuming, labor-intensive, and costly, making it difficult to obtain real-time and continuous information on the spatiotemporal dynamic distribution of salt concentration within salt ponds; the regulatory response lags significantly behind environmental dynamic changes, often causing salinity to deviate from the set threshold within a short period, directly affecting the comparability of growth environments among different breeding materials and the accuracy of screening results; and the regulation precision is generally low, failing to meet the stringent requirements of modern molecular breeding and phenomics for highly uniform and controllable experimental environments. These technical bottlenecks severely restrict the screening efficiency of superior salt-tolerant rice germplasm and the acceleration of the breeding process.

[0003] Therefore, there is an urgent need to construct a full-stack control system covering the physical layer, transmission layer, data layer, model layer, and application layer, creating a high-fidelity digital twin that corresponds one-to-one with the physical salt pond and environmental elements, deeply integrating mechanistic models and data-driven approaches. This twin will enable real-time monitoring of the spatiotemporal distribution of salinity in three dimensions, intelligent prediction of short-term trends, and simulation and effect evaluation of different control strategies. Furthermore, it will integrate multi-objective optimization algorithms and intelligent decision-making modules to form the optimal control scheme, which will then be precisely executed through IoT-based automatic control terminals. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a method for intelligent control of salinity in salt-tolerant rice breeding based on a digital twin model.

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

[0006] A method for intelligent salinity control in salt-tolerant rice breeding based on digital twin models includes the following steps:

[0007] Step S1: Construct a digital twin model corresponding to the physical breeding salt pond, wherein the digital twin model includes a mechanism kernel and a data-driven correction module;

[0008] Step S2: Data is collected in real time through a sensor network deployed in the physical breeding salt pond, and the data is processed for anomaly detection, dynamic temperature compensation and redundancy fusion to generate a standardized feature vector for driving the digital twin model.

[0009] Step S3: Input the standardized feature vector into the digital twin model, use the extended state observer to estimate the system state and total disturbance in real time, realize the state synchronization of the digital twin with the physical entity, and adaptively update the parameters of the mechanism kernel based on the total disturbance.

[0010] Step S4: Based on the updated digital twin model and environmental data, drive the hybrid model in the digital twin model to run and output the salinity prediction value for the future time step;

[0011] Step S5: Based on the preset salinity target setpoint, the system state and total disturbance estimated by the extended state observer, and the salinity prediction value, the active disturbance rejection control algorithm including the tracking differentiator and the disturbance compensation link is used to calculate the control command for adjusting the salinity.

[0012] Step S6: Send the control command to the physical actuator to perform water replenishment or salt adjustment operations on the salt pond, thereby achieving closed-loop intelligent control of salinity.

[0013] Furthermore, in step S1, the digital twin model includes a data-driven interface layer, a hybrid model kernel layer, and a service engine layer; the data-driven interface layer is used to receive and assimilate real-time sensing data; the hybrid model kernel layer is used to perform simulation and prediction calculations; and the service engine layer provides state visualization and simulation inference services.

[0014] Furthermore, in step S1, the hybrid model kernel layer is composed of a mechanistic model and a data-driven correction module connected in series; the mechanistic model is constructed based on the physical equation of salt mass conservation and is used to provide a basic simulation framework; the data-driven correction module takes the prediction residuals, historical states and environmental sequences of the mechanistic model as inputs, learns the unmodeled dynamics through a deep learning network, and outputs the dynamic correction amount of the prediction results of the mechanistic model.

[0015] Furthermore, in step S2, the dynamic temperature compensation specifically adopts a second-order temperature compensation model to correct the original reading of the conductivity sensor to the value at the standard temperature. The compensation model also incorporates the effective temperature calculated through a first-order inertial element to compensate for the thermal inertia of the sensor probe.

[0016] The redundancy fusion specifically employs a weighted average algorithm based on dynamic reliability weights, where the dynamic reliability weight of a single sensor is calculated based on the deviation of its reading from the median of the readings of the same group of sensors and the historical residual statistical level of that sensor.

[0017] Further, in step S3, the extended state observer is a discrete nonlinear observer, and the operation process includes: calculating the error between the current model state estimate and the actual measured value; based on the error, the previous state estimate, the estimated system control gain, and the observer gain parameter, updating the estimated values ​​of the system output, output derivative, and total disturbance through recursive calculation; wherein, a nonlinear function is used to handle the error during the update process, and the nonlinear function is calculated in a power form when the absolute value of the error is greater than a preset threshold, and in a linear form when the absolute value of the error is less than or equal to the preset threshold.

[0018] Furthermore, in step S3, the update process of the digital twin model specifically includes a dual-loop adaptive mechanism:

[0019] The inner loop is a state synchronization loop based on an extended state observer, which uses physical sensor measurements to correct the internal state estimates of the twin model in real time and observes the total disturbance.

[0020] The outer loop is a parameter self-learning loop, which periodically uses the total disturbance history sequence observed by the extended state observer to back-optimize and correct the time-varying parameters in the mechanism model through the parameter identification algorithm.

[0021] Furthermore, in step S5, the active disturbance rejection control algorithm adopts a cascade control structure, including an outer-loop salinity controller and an inner-loop water level controller; the outer-loop salinity controller uses the total disturbance estimated by the extended state observer for feedforward compensation; the inner-loop water level controller uses an independent extended state observer to estimate and compensate for secondary disturbances.

[0022] Furthermore, in step S5, the tracking differentiator is implemented by constructing a discrete system model using the fastest synthesis function. The discrete system model takes the target setpoint, the current tracking signal and its differential estimate, the velocity factor and the filter factor as inputs, and outputs the tracking signal and its differential estimate at the next moment through recursive calculation.

[0023] Furthermore, the hybrid model takes historical salinity sequences and historical environmental variable sequences as inputs and outputs predicted salinity values ​​at a specified future time through a nonlinear mapping function; the nonlinear mapping function is obtained through training, and the parameter update is achieved by minimizing the combined loss function during the training process.

[0024] The combined loss function includes a data fitting term and a physical law constraint term. The data fitting term is used to minimize the error between the predicted value and the actual value. The physical law constraint term is constructed based on the water-salt mass conservation equation and is used to ensure that the prediction result conforms to physical laws.

[0025] Furthermore, in step S4, the environmental data includes at least future weather forecast data, which is obtained from an external meteorological service system through an interface, and the parameters include future temperature, humidity, wind speed, solar radiation, and precipitation forecasts.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. By integrating the real-time perturbation estimation and compensation mechanism of ADRC with the high-fidelity simulation of digital twins, the present invention can stably control the salinity within the preset target value error band, providing highly uniform and reproducible environmental conditions for cutting-edge research such as molecular breeding.

[0028] 2. The digital twin created by this invention is not a static model, but achieves millisecond-level state synchronization through ESO and continuously evolves using a dual-loop adaptive mechanism, enabling the virtual model to accurately track and predict the long-term dynamic changes of the physical entity, significantly improving the model's lifespan and practicality.

[0029] 3. This invention integrates the entire chain of capabilities from intelligent preprocessing of multi-source data and intelligent prediction of hybrid models to intelligent decision-making of cascaded ADRC, realizing a fundamental shift from passive response to active intervention and from experience-based decision-making to model-driven decision-making, and significantly improving the automation and intelligence level of breeding environment management.

[0030] 4. This invention effectively overcomes sensor drift, noise, and single-point failure problems through a data preprocessing pipeline that includes dynamic temperature compensation, redundancy fusion, and fault diagnosis. Attached Figure Description

[0031] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0032] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0033] Figure 2 This is a schematic diagram of the dual-ring adaptive update mechanism according to an embodiment of the present invention;

[0034] Figure 3 This is a flowchart of the intelligent salinity control algorithm according to an embodiment of the present invention;

[0035] Figure 4 This is a graph showing the results of adjusting the salt concentration in the experimental salt tank to 3000 mg / L according to an embodiment of the present invention.

[0036] Figure 5 This is a graph showing the results of adjusting the salt concentration in the experimental salt tank to 5000 mg / L, as described in an embodiment of the present invention.

[0037] Figure 6The graph shows the results of adjusting the salt concentration in the experimental salt tank to 7000 mg / L in an embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, the intelligent salinity control method for salt-tolerant rice breeding based on a digital twin model includes the following steps:

[0040] Step S1: Construct a digital twin model corresponding to the physical breeding salt pond, wherein the digital twin model includes a mechanism kernel and a data-driven correction module;

[0041] Step S2: Data is collected in real time through a sensor network deployed in the physical breeding salt pond, and the data is processed for anomaly detection, dynamic temperature compensation and redundancy fusion to generate a standardized feature vector for driving the digital twin model.

[0042] Step S3: Input the standardized feature vector into the digital twin model, use the extended state observer to estimate the system state and total disturbance in real time, realize the state synchronization of the digital twin with the physical entity, and adaptively update the parameters of the mechanism kernel based on the total disturbance.

[0043] Step S4: Based on the updated digital twin model and environmental data, drive the hybrid model in the digital twin model to run and output the salinity prediction value for the future time step;

[0044] Step S5: Based on the preset salinity target setpoint, the system state and total disturbance estimated by the extended state observer, and the salinity prediction value, the active disturbance rejection control algorithm including the tracking differentiator and the disturbance compensation link is used to calculate the control command for adjusting the salinity.

[0045] Step S6: Send the control command to the physical actuator to perform water replenishment or salt adjustment operations on the salt pond, thereby achieving closed-loop intelligent control of salinity.

[0046] In step S1, the digital twin model includes a data-driven interface layer, a hybrid model kernel layer, and a service engine layer; the data-driven interface layer is used to receive and assimilate real-time sensing data; the hybrid model kernel layer is used to perform simulation and prediction calculations; and the service engine layer provides state visualization and simulation inference services.

[0047] In step S1, the hybrid model kernel layer is composed of a mechanistic model and a data-driven correction module connected in series. The mechanistic model is constructed based on the physical equation of salt mass conservation and is used to provide a basic simulation framework. The data-driven correction module takes the prediction residuals, historical states and environmental sequences of the mechanistic model as inputs, learns the unmodeled dynamics through a deep learning network, and outputs the dynamic correction amount of the prediction results of the mechanistic model.

[0048] The anomaly detection specifically includes: identifying outliers by combining sliding window statistics and physical consistency checks on the raw sensor data, and filling out the outliers with linear interpolation or predicted values ​​based on state-space models.

[0049] In step S2, the dynamic temperature compensation specifically employs a second-order temperature compensation model to correct the original reading of the conductivity sensor to a value at the standard temperature. The compensation model incorporates an effective temperature calculated using a first-order inertial element to compensate for the thermal inertia of the sensor probe. The specific formula is as follows:

[0050]

[0051] in, This represents the conductivity value after compensation to a standard temperature of 25°C at time t. This represents the raw conductivity reading directly measured by the sensor at time t. This represents the effective temperature of the sensor at time t. This represents the first-order temperature compensation coefficient, in °C⁻¹. This represents the second-order temperature compensation coefficient, in °C⁻².

[0052] The effective temperature is calculated using a first-order inertial element model, and the specific formula is as follows:

[0053]

[0054] in, This represents the actual water temperature measured at time t. The thermal time constant of the sensor is obtained through sensor calibration or step response testing. It characterizes the temperature response speed of the probe and can avoid transient distortion of salinity readings caused by rapid fluctuations in water temperature, such as the injection of new water, so that the compensation is more in line with physical reality.

[0055] The redundancy fusion specifically employs a weighted average algorithm based on dynamic reliability weights. The dynamic reliability weight of a single sensor is calculated based on the deviation of its reading from the median of the readings of other sensors in the same group and the historical residual statistical level of that sensor. The specific formula is as follows:

[0056]

[0057] in, This represents the dynamic confidence weight of sensor i at time t, with a value ranging from 0 to 1. This represents the degree of consistency deviation of sensor i at time t. This represents the historical residual constant of sensor i up to time t. and This represents the adjustment coefficient, with values ​​ranging from [0.5, 2] to [1, 5].

[0058] The specific formulas for consistency deviation and historical residual abnormality are as follows:

[0059]

[0060]

[0061] in, This represents the temperature-compensated reading of sensor i at time t. This represents the median of all effective redundant sensor readings at time t. Indicates the length of the historical time window. Represents a time index variable. This represents the temperature-compensated reading of sensor i at time n. The optimal estimate of time n is obtained by periodically performing manual calibration measurements using laboratory-grade high-precision instruments. The measurements at these calibration times are used as the optimal estimate of time n. For non-calibration times, the optimal estimate is obtained through one of the following methods: 1. Linear interpolation of the measurements at adjacent calibration times; 2. When the time interval between the last calibration and the last calibration is less than a preset threshold (e.g., 24 hours), the last calibration value is directly used.

[0062] In step S3, the extended state observer is a discrete nonlinear observer, and the operation process includes: calculating the error between the current model state estimate and the actual measured value; based on the error, the previous state estimate, the estimated system control gain, and the observer gain parameter, updating the estimated values ​​of the system output, output derivative, and total disturbance through recursive calculation; wherein, a nonlinear function is used to handle the error during the update process, and the nonlinear function is calculated in a power form when the absolute value of the error is greater than a preset threshold, and in a linear form when the absolute value of the error is less than or equal to the preset threshold.

[0063] The extended state observer is the core computing unit of a digital twin system for enabling virtual-real interaction and model updates. It does not rely on a precise mathematical model of the system, but instead treats model parameter perturbations (such as changes in pipe resistance coefficients) and external environmental disturbances (such as rainfall and evaporation) as a unified total disturbance of the system, extending it into a new state variable for real-time observation.

[0064] For a second-order system, a discrete nonlinear extended state observer is designed as follows:

[0065]

[0066] in, This represents the sensor measurement value after preprocessing at the k-th sampling time. This represents the estimation error of the extended state observer. Let represent the estimated value of the system output and the estimated value of the system output derivative at the k-th sampling time, respectively, representing the real-time state of the digital twin model. This represents the extended state variable, i.e., the estimate of the total system disturbance at the k-th sampling time. This represents the observer gain parameter. This represents a nonlinear function, used to increase the gain in the small error range to accelerate convergence. Indicates the sampling step size. and This represents a nonlinear factor, typically 0 < , <1, used to adjust the shape of nonlinear functions. This represents the width of the linear interval, used to divide the linear and nonlinear segments. When the absolute value of the error is less than or equal to δ, the function is linear; when it is greater than δ, it is nonlinear. This represents the discrete-time index, indicating that the current sampling time is the k-th time. This represents the control input at the k-th sampling time, such as valve opening command, water supply flow setpoint, etc., and is the controller's output. The estimated value of the system control gain represents the strength of the influence of the control input on the rate of change of the system state. The nonlinear factor usually takes the same value, ranging from [0.25, 0.75], preferably 0.5. The setting of the linear interval width is related to the sensor measurement noise level, and can usually be taken as 2 to 5 times the standard deviation of the steady-state measurement error.

[0067] By extending the state observer, the digital twin can capture the dynamic deviations of the physical entity in real time and use this to correct the boundary conditions of the virtual model, thus achieving high-fidelity model updates.

[0068] like Figure 2 As shown, the update process of the digital twin model in step S3 specifically includes a dual-loop adaptive mechanism:

[0069] The inner loop is a state synchronization loop based on an extended state observer, which uses physical sensor measurements to correct the internal state estimates of the twin model in real time and observes the total disturbance.

[0070] The outer loop is a parameter self-learning loop, which periodically uses the total disturbance history sequence observed by the extended state observer to back-optimize and correct the time-varying parameters in the mechanism model through the parameter identification algorithm.

[0071] The parameter identification algorithm adopts recursive least squares method or recursive least squares method with forgetting factor. The parameters include, but are not limited to, the equivalent leakage coefficient characterizing the leakage characteristics of salt ponds, and the crop water absorption function parameters reflecting the salt absorption capacity of rice at different growth stages.

[0072] In step S5, the active disturbance rejection control algorithm adopts a cascade control structure, which includes an outer-loop salinity controller and an inner-loop water level controller. The outer-loop salinity controller uses the total disturbance estimated by the extended state observer for feedforward compensation. The inner-loop water level controller uses an independent extended state observer to estimate and compensate for secondary disturbances.

[0073] The Active Disturbance Rejection Control (ADRC) algorithm adopts a cascade control structure, with the outer loop salinity controller as the main loop and the inner loop water level or flow controller as the secondary loop. Its working principle includes:

[0074] The outer ring is responsible for calculating the ideal water replenishment flow rate setpoint required to maintain the salinity based on the deviation between the target salinity and the current salinity. The inner ring is responsible for rapid execution and precise control of valves or pumps to ensure that the actual flow rate or water level quickly tracks the setpoint.

[0075] The extended state observer built into the outer loop controller estimates and outputs the total disturbance acting on the salinity process in real time. When generating output commands, the controller performs feedforward compensation on the disturbance, thereby actively canceling it out before environmental disturbances affect salinity.

[0076] The inner loop controller has an independent extended state observer specifically designed to estimate and compensate for secondary disturbances occurring during execution that are imperceptible to the outer loop (such as valve dead zones and pipeline resistance changes). The inner loop controller also performs feedforward compensation for these disturbances to ensure that the flow commands from the outer loop are executed faithfully and quickly.

[0077] Through precise division of labor and coordination between the inner and outer loops, and the real-time observation and active compensation of disturbances by their respective ESOs, this cascade structure effectively decouples the slow salinity control process from the fast flow control process. This enables the system not only to respond quickly to changes in commands but also to strongly resist various types of disturbances across the entire chain, from the field environment to the actuators, ultimately achieving high-precision, highly robust closed-loop intelligent control of salinity.

[0078] In step S5, the tracking differentiator is implemented by constructing a discrete system model using the fastest synthesis function. The discrete system model takes the target setpoint, the current tracking signal and its differential estimate, the velocity factor and the filter factor as inputs, and outputs the tracking signal and its differential estimate at the next moment through recursive calculation.

[0079] To resolve the conflict between system response speed and overshoot, and to prevent system oscillations caused by abrupt changes in the target salinity value, a discrete-time tracking differentiator is introduced. By arranging a transient process, the discontinuous target signal is transformed into a smooth tracking signal and its derivative signal. The fastest synthesis function is employed. The discrete system model is constructed as follows:

[0080]

[0081] in, Represents the fastest control synthesis function The calculated output value at the current moment. This represents the target salinity setpoint at the k-th sampling time. This represents the output of the target at the k-th sampling time. The tracking signal This represents the output of the target at the k-th sampling time. The differential estimate, i.e., the rate of change of the tracking signal, The velocity factor represents the tracking rate of the system. Indicates the sampling step size. This represents the filter factor, which is usually taken as an integer multiple of the sampling step size h. This represents the fastest synthesis function. The inputs are two state variables and parameters, and the output is a control variable. It is used to arrange the transient process. The speed factor determines the speed of the transient process. Its value is related to the system's requirement for tracking the setpoint and the capability of the actuator. It is usually tuned through simulation or on-site debugging.

[0082] The hybrid model takes historical salinity and historical environmental variable sequences as inputs and outputs a predicted salinity value at a specified future time through a nonlinear mapping function. The nonlinear mapping function is obtained through training, and the parameter update is achieved by minimizing the combined loss function during the training process.

[0083] The combined loss function includes a data fitting term and a physical law constraint term. The data fitting term is used to minimize the error between the predicted value and the actual value. The physical law constraint term is constructed based on the water-salt mass conservation equation and is used to ensure that the prediction result conforms to physical laws.

[0084] To address the highly nonlinear, time-varying, and coupled characteristics of salinity dynamics, an intelligent prediction framework integrating physical mechanism models and data-driven models was constructed. This framework is based on the water-salt transport equations of the physical field and uses real-time sensor data to correct parameters, ensuring the model maintains high consistency with the real system. Furthermore, a deep learning network LSTM is introduced to extract features from salinity time-series data, capturing long-term dependencies and short-term disturbance patterns. A multivariate prediction model is then constructed using multi-source inputs (meteorological data, water level, evaporation, etc.) to improve the responsiveness to abrupt environmental changes.

[0085] Mathematically, the prediction model can be formalized as:

[0086]

[0087] in, This represents the future salinity value output by the predictive model. This represents the standardized salinity sequence from time tc to time t. This represents the standardized multi-source environmental variable sequence from time tc to time t. This represents the nonlinear mapping obtained through training. Indicates the prediction step size. This represents the window size, i.e., the time step from time t−c to time t. Let f represent the set of all trainable parameters in the hybrid prediction model f. Using an LSTM network, then... It includes the weight matrix and bias vector of the input gate, forget gate, and output gate, as well as the parameters of the fully connected layer. The initial value of θ is usually randomly initialized. The gradient of the combined loss function with respect to θ is calculated through the backpropagation algorithm, and the optimizer is used to iteratively update θ to minimize the combined loss function.

[0088] The specific formula for the combined loss function is as follows:

[0089]

[0090] in, Represents the combination loss function. Represents the data fitting term. Represents the physical law constraint terms. and These represent the corresponding weight coefficients, and are set accordingly. Using 1.0 as the baseline, The value range is [0.1, 0.5], and the optimal value is determined by performing grid search or cross-validation on the validation set;

[0091] The formulas for calculating the data fitting term and the physical law constraint term are as follows:

[0092]

[0093]

[0094] in, This indicates the number of samples in the training batch. This represents the predicted future salinity value for the j-th sample. This represents the true future salinity value corresponding to the j-th sample. This represents the square of the L2 norm, i.e., the mean square error. This represents the rate of change of the salinity sequence over time as predicted by the model. This indicates that, based on the water-salinity conservation equation, the predicted salinity and input environmental characteristics are used... The calculated theoretical rate of change in salinity, This indicates the time step of the calculation.

[0095] like Figure 3 As shown, the sensor data (water level, conductivity) of the physical entity are used as system observations. The internal state and total external disturbance of the system are estimated in real time through ESO, so as to realize the millisecond-level state synchronization and parameter correction of the physical entity by the virtual model, and then solve the optimal control quantity to drive the actuator (solenoid valve) to act.

[0096] In step S4, the environmental data includes at least future weather forecast data, which is obtained from an external meteorological service system through an interface. The parameters include future temperature, humidity, wind speed, solar radiation, and precipitation forecasts.

[0097] like Figure 4 , Figure 5 , Figure 6 As shown, three experimental salt ponds (Aa-01, Aa-20, and Aa-75) with different salinity targets were selected at the national base for a 130-day continuous monitoring experiment. The salinity target values ​​were set at 3000 mg / L, 5000 mg / L, and 7000 mg / L, respectively, to simulate three typical salt-tolerant rice breeding stress environments: low, medium, and high. During the experiment, the system simultaneously recorded the "intelligent control salinity" command curve output by the digital twin intelligent control system, as well as the "measured salinity" data obtained daily through manual measurement using a handheld high-precision salinity meter. By comparing and analyzing the relationship between these two data points and the "salinity target value" baseline, the accuracy and stability of the control algorithm, as well as the reliability of the entire digital twin system, can be comprehensively evaluated from two dimensions: different salinity levels and different monitoring methods.

[0098] In-depth analysis of the curves reveals that the proposed intelligent control method exhibits excellent control performance at different salinity levels. Specifically, throughout the complete 130-day rice growth cycle, the "intelligent control salinity" curve (typically a model-predicted control command or expected value) and the "measured salinity" curve (representing the actual state of the physical world) consistently and closely adhere to the preset salinity target line, maintaining a very narrow fluctuation bandwidth. Quantitative statistics show that the measured salinity values ​​of all salt ponds remained stable within ±100 mg / L of the target value throughout the experiment. For example, for a salt pond with a target of 3000 mg / L, salinity fluctuations were effectively limited to 2900-3100 mg / L; for a salt pond with a target of 5000 mg / L, salinity was maintained at 4900-5100 mg / L; and even with a control target as high as 7000 mg / L, the system still stabilized the actual salinity within the range of 6900-7100 mg / L. This result first verifies the effectiveness of the regulation itself: the intelligent regulation system can overcome time-varying disturbances such as evaporation, rainfall, irrigation, and crop water absorption, and dynamically adjust water replenishment or salt removal strategies, so that the key environmental parameter of salinity is precisely "locked" near the set point required for the breeding experiment, meeting the stringent requirements of modern breeding for environmental uniformity and repeatability.

[0099] Further analysis of the comparative relationship between the "intelligently controlled salinity" and "measured salinity" curves reveals the deeper effectiveness of the digital twin system. The high degree of overlap between the two curves is not a simple static approximation, but rather a synchronization and consistency demonstrated during 130 days of dynamic changes. This strongly proves the high fidelity of the digital twin model and the advanced nature of the predictive control algorithm. The digital twin continuously and proactively simulates the water-salinity dynamics of the salt pond in its virtual space by assimilating sensor data (water level, initial salinity, etc.) in real time and estimating the "total disturbance" caused by environmental disturbances. Based on this simulation prediction, the ADRC algorithm can generate optimized control commands in advance. When these commands are applied to the physical salt pond, their effects are reflected in the "measured salinity," which highly matches the predicted value (reflected in the expected trajectory of "intelligently controlled salinity"). This indicates that the entire closed-loop system operates smoothly, from simulation and decision-making in the "virtual space" to execution and feedback in the "physical space." The mapping of the digital twin to the physical entity is accurate, and its model-based predictions are reliable. Therefore, prediction-based decision-making is effective. Particularly noteworthy is that the system exhibited similar precise control capabilities in experiments at three different salinity levels. This demonstrates that the model integrating the mechanism and data, along with the ADRC algorithm, constructed in this paper possesses good generalization and robustness, adapting to control requirements across different concentration ranges without experiencing performance degradation due to increasing absolute salinity.

[0100] The examples described herein are merely preferred embodiments of the invention and are not intended to limit the concept and scope of the invention. Any modifications and improvements made by those skilled in the art to the technical solutions of the invention without departing from the design concept of the invention should fall within the protection scope of the invention.

Claims

1. A method for intelligent salinity control in salt-tolerant rice breeding based on a digital twin model, characterized in that, Includes the following steps: Step S1: Construct a digital twin model corresponding to the physical breeding salt pond, wherein the digital twin model includes a mechanism kernel and a data-driven correction module; In step S1, the digital twin model includes a data-driven interface layer, a hybrid model kernel layer, and a service engine layer; the data-driven interface layer is used to receive and assimilate real-time sensor data; the hybrid model kernel layer is used to perform simulation and prediction calculations; and the service engine layer provides state visualization and simulation inference services. The hybrid model kernel layer consists of a mechanistic model and a data-driven correction module connected in series. The mechanistic model is constructed based on the physical equation of salt mass conservation and is used to provide a basic simulation framework. The data-driven correction module takes the prediction residuals, historical states and environmental sequences of the mechanistic model as input, learns the unmodeled dynamics through a deep learning network, and outputs the dynamic correction amount of the prediction results of the mechanistic model. Step S2: Data is collected in real time through a sensor network deployed in the physical breeding salt pond, and the data is processed for anomaly detection, dynamic temperature compensation and redundancy fusion to generate a standardized feature vector for driving the digital twin model. Step S3: Input the standardized feature vector into the digital twin model, use the extended state observer to estimate the system state and total disturbance in real time, realize the state synchronization of the digital twin with the physical entity, and adaptively update the parameters of the mechanism kernel based on the total disturbance. Specifically, the update process of the digital twin model in step S3 includes a dual-loop adaptive mechanism: The inner loop is a state synchronization loop based on an extended state observer, which uses physical sensor measurements to correct the internal state estimates of the twin model in real time and observes the total disturbance. The outer loop is a parameter self-learning loop, which periodically uses the total disturbance history sequence observed by the extended state observer to back-optimize and correct the time-varying parameters in the mechanism model through the parameter identification algorithm; Step S4: Based on the updated digital twin model and environmental data, drive the hybrid model in the digital twin model to run and output the salinity prediction value for the future time step; Step S5: Based on the preset salinity target setpoint, the system state and total disturbance estimated by the extended state observer, and the salinity prediction value, the active disturbance rejection control algorithm including the tracking differentiator and the disturbance compensation link is used to calculate the control command for adjusting the salinity. Step S6: Send the control command to the physical actuator to perform water replenishment or salt adjustment operations on the salt pond, thereby achieving closed-loop intelligent control of salinity.

2. The method according to claim 1, characterized in that, In step S2, the dynamic temperature compensation specifically adopts a second-order temperature compensation model to correct the original reading of the conductivity sensor to the value at the standard temperature. The compensation model also incorporates the effective temperature calculated through a first-order inertial element to compensate for the thermal inertia of the sensor probe. The redundancy fusion specifically employs a weighted average algorithm based on dynamic reliability weights, where the dynamic reliability weight of a single sensor is calculated based on the deviation of its reading from the median of the readings of the same group of sensors and the historical residual statistical level of that sensor.

3. The method according to claim 2, characterized in that, In step S3, the extended state observer is a discrete nonlinear observer, and the operation process includes: calculating the error between the current model state estimate and the actual measured value; based on the error, the previous state estimate, the estimated system control gain, and the observer gain parameter, updating the estimated values ​​of the system output, output derivative, and total disturbance through recursive calculation; wherein, a nonlinear function is used to handle the error during the update process, and the nonlinear function is calculated in a power form when the absolute value of the error is greater than a preset threshold, and in a linear form when the absolute value of the error is less than or equal to the preset threshold.

4. The method according to claim 3, characterized in that, In step S5, the active disturbance rejection control algorithm adopts a cascade control structure, which includes an outer-loop salinity controller and an inner-loop water level controller. The outer-loop salinity controller uses the total disturbance estimated by the extended state observer for feedforward compensation. The inner-loop water level controller uses an independent extended state observer to estimate and compensate for secondary disturbances.

5. The method according to claim 4, characterized in that, The tracking differentiator is implemented by constructing a discrete system model using the fastest synthesis function. The discrete system model takes the target setpoint, the current tracking signal and its differential estimate, the velocity factor and the filter factor as inputs, and outputs the tracking signal and its differential estimate at the next moment through recursive calculation.

6. The method according to claim 5, characterized in that, The hybrid model takes historical salinity and historical environmental variable sequences as inputs and outputs a predicted salinity value at a specified future time through a nonlinear mapping function. The nonlinear mapping function is obtained through training, and the parameter update is achieved by minimizing the combined loss function during the training process. The combined loss function includes a data fitting term and a physical law constraint term. The data fitting term is used to minimize the error between the predicted value and the actual value. The physical law constraint term is constructed based on the water-salt mass conservation equation and is used to ensure that the prediction result conforms to physical laws.

7. The method according to claim 6, characterized in that, In step S4, the environmental data includes at least future weather forecast data, which is obtained from an external meteorological service system through an interface. The parameters include future temperature, humidity, wind speed, solar radiation, and precipitation forecasts.

Images