Multi-objective optimization intelligent control system for water and fertilizer integration based on artificial intelligence

Abstract

The present invention relates to the field of automatic control technology, and more specifically, to an artificial intelligence-driven, integrated water and fertilizer multi-objective optimization intelligent control system. The system comprises: a water and fertilizer load analysis unit and a water and fertilizer combined control criterion generation unit; the water and fertilizer load analysis unit is configured to obtain meteorological data from a meteorological monitoring system; simultaneously receive data from soil sensors and leaf sensors to obtain crop growth status data and assess the water and fertilizer requirements of the crop per unit area at each time; and the water and fertilizer combined control criterion generation unit is configured to use an AI dynamic resource optimization engine to dynamically calculate the amount of water and fertilizer applied to the crop per unit area at each time based on the soil state and the crop's root absorption capacity. The present invention effectively improves water and nutrient utilization efficiency, reduces resource waste, and enhances the intelligence, adaptability, and robustness of the control system.

Description

Technical Field

[0001] The present invention belongs to the field of automatic control technology, and specifically relates to an artificial intelligence-driven water-fertilizer integrated multi-objective optimization intelligent control system. Background Art

[0002] In modern agricultural production systems, crop water and nutrient management is a key factor influencing yield and resource utilization efficiency. Especially in the context of the rapid development of precision agriculture, traditional timed and quantitative irrigation and conventional proportional fertilization technologies have become unable to meet the multiple requirements of high yield, water conservation, environmental protection, and intelligent management. Consequently, a large number of researchers and agricultural engineering application units have conducted technical explorations focusing on water-fertilizer integration, automated irrigation control, and precise nutrient supply, and have made considerable progress. However, while these existing technologies have laid the foundation for agricultural informatization and automation, a series of unresolved issues remain in areas such as coordinated control, dynamic response capabilities, adaptive regulation, and closed-loop optimization, which restrict their universal adaptability and long-term stability under large-scale, multi-climate, and heterogeneous plot conditions.

[0003] Currently, widely used water and fertilizer regulation systems often employ static formulation strategies based on empirical models. For example, many regions establish empirical "crop-period-dosage" databases based on years of crop cultivation data, and then formulate irrigation and fertilization plans based on weather forecasts and manual field inspections. While this approach performs acceptably in normal seasons and on stable plots, it lacks real-time adaptability to factors such as abnormal climate (such as sudden high temperatures and high evapotranspiration days), complex terrain conditions (such as sloping land and varying soil bulk density distributions), and individual crop differences, making it difficult to ensure the timeliness of regulation strategies and their alignment with actual crop needs.

[0004] To overcome the limitations of static models, some studies have introduced real-time monitoring platforms based on sensor data, using parameters such as soil moisture, electrical conductivity, temperature and humidity to perform rule-triggered irrigation and fertilization control. This type of system generally uses a threshold mechanism. For example, when the soil moisture content falls below a preset value, valve control is automatically triggered to perform irrigation operations, or the appropriate fertilizer concentration is determined based on changes in EC values. Although this type of method introduces preliminary feedback control logic, its essence is still a "single factor control" or "preset threshold trigger" logic structure, and it has not formed a system-level dynamic collaborative optimization mechanism. More importantly, this method cannot identify the physiological feedback of crops, and only formulates operations based on soil or meteorological signals, ignoring the need for comprehensive regulation of crop absorption capacity, nutrient transport rate and growth stage characteristics. Summary of the Invention

[0005] The main purpose of the present invention is to provide an integrated water and fertilizer multi-objective optimization intelligent control system driven by artificial intelligence. By constructing a water and fertilizer load analysis unit, a water and fertilizer joint control criterion generation unit and a water and fertilizer feedback control unit, real-time evaluation, dynamic optimization calculation and closed-loop correction control of the water and fertilizer requirements of crops are achieved. The system integrates an AI dynamic optimization engine for water and fertilizer application, and integrates a historical residual compensation mechanism. It can accurately output control quantities driven by multi-source environmental data and crop physiological parameters, ensuring that water and fertilizer control not only responds to the actual needs of crops, but also has long-term stability and environmental adaptability. The system effectively improves the efficiency of water and nutrient utilization, reduces resource waste, and enhances the intelligence, adaptability and robustness of control. It is suitable for smart agricultural production in multi-plot and multi-crop scenarios.

[0006] In order to solve the above problems, the technical solution of the present invention is achieved as follows:

[0007] An artificial intelligence-driven integrated water and fertilizer multi-objective optimization intelligent control system, the system includes: a water and fertilizer load analysis unit and a water and fertilizer joint control criterion generation unit; the water and fertilizer load analysis unit is used to obtain meteorological data from the meteorological monitoring system; at the same time, it receives data from soil sensors and leaf sensors to obtain crop growth status data, and evaluates the water and fertilizer requirements of crops per unit area at each time; the water and fertilizer joint control criterion generation unit is used to use an AI dynamic resource optimization engine to dynamically calculate the amount of water and fertilizer applied to crops per unit area at each time based on soil conditions and crop root absorption capacity; the AI ​​dynamic resource optimization engine includes: an AI dynamic optimization engine for water application and an AI dynamic optimization engine for fertilizer application; the AI ​​dynamic optimization engine for water application is obtained by adding an AI dynamic optimization model for water application to the accumulated part of AI historical residuals for water application; the AI ​​dynamic optimization engine for fertilizer application is obtained by adding an AI dynamic optimization model for fertilizer application to the accumulated part of AI historical residuals for fertilizer application.

[0008] Furthermore, the execution process of the AI ​​dynamic optimization engine for water application includes: calling the residual sequence of the most recent complete growth cycle in the historical residual accumulation part of the AI ​​water application, setting the negative definite convergence domain based on the linear quadratic Lyapunov function, and writing into the residual accumulation register and the safety threshold register; real-time collection of soil status and crop root absorption capacity, together with the irrigation records of the previous period, writing into the state observation buffer, and forming a state vector after unified scaling; the AI ​​dynamic optimization model for water application calls the linear matrix inequality solver to output the original water application control quantity that meets the Lyapunov descent constraint; at the same time, the adaptive gain modulator dynamically updates the matrix weight according to the convergence rate; the AI ​​historical residual accumulation part of water application incorporates the latest error into the residual accumulation register, generates the same-dimensional compensation quantity through the variable gain compensator, and buffers rapid disturbances; the original water application control quantity and the compensation quantity are added dimension by dimension, and are limited to the execution threshold range by the saturation protector to obtain the target water application quantity, and are written into the execution instruction queue.

[0009] Furthermore, the residual sequence of the most recent complete reproductive cycle in the accumulated part of the AI ​​historical residual of water application is called, and the process of setting the negative definite convergence region based on the linear quadratic Lyapunov function includes:

[0010] Step A1: Perform integrity check on the residual sequence of the accumulated part of the AI ​​historical residual of water application, and linearly interpolate the missing segments in chronological order to obtain the verified sequence;

[0011] Step A2: Write the verified sequence into the residual buffer stack and use a double-weight exponential decay filter to suppress high-frequency noise to obtain a filtered sequence;

[0012] Step A3: Detect the extreme value of the filter sequence. If it exceeds the safety threshold register threshold, trigger the abnormal backtracking logic and return to step A1;

[0013] Step B1: dynamically generate a positive definite weighting matrix based on the variance distribution of the filter sequence and write it into the matrix register;

[0014] Step B2: Construct a linear quadratic Lyapunov function from the positive definite weight matrix and the residual state vector according to the quadratic form, and verify the strict descent of all samples through the gradient iterator;

[0015] Step B3: If the drop rate of any sample is non-negative, adaptively shrink the positive definite weight matrix and return to step B2 until all samples meet the conditions;

[0016] Step C1: define the boundary of the negative definite convergence region with the maximum acceptable energy level of the linear quadratic Lyapunov function and write it into the negative definite convergence region register;

[0017] Step C2: Synchronize the boundary parameters to the safety locking logic and the water application AI dynamic optimization engine to complete the setting of the negative definite convergence region.

[0018] Furthermore, the execution process of the AI ​​dynamic optimization engine for fertilizer application includes: calling out the residual sequence of the most recent complete growth cycle from the accumulated part of the AI ​​historical residuals for fertilizer application, writing it into the residual buffer stack, establishing the initial value of the sliding surface and recording the zero-bias baseline; collecting soil conditions and crop root absorption capacity, and reading the fertilization records of the previous period, writing them into the state observation buffer after unified scaling to form a state vector; the AI ​​dynamic optimization model for fertilizer application is based on the difference between the sliding surface and the state vector, and uses the Darwin switching law to generate the original fertilizer control quantity; the adaptive gain modulator in the same period updates the switching rate according to the approximation speed; the accumulated part of the AI ​​historical residuals for fertilizer application is superimposed on the nutritional deviation of this period, and is converted into an equal-dimensional compensation quantity by a variable gain compensator to suppress high-frequency disturbances; the original fertilizer control quantity and the compensation quantity are added dimension by dimension, limited to the safety threshold by the saturation protector, and the fertilizer quantity is obtained, and written into the execution instruction queue.

[0019] Furthermore, the AI ​​dynamic optimization model of fertilizer application amount uses the Daltonian switching law to generate the original fertilizer control amount based on the difference between the sliding surface and the state vector. The process includes: establishing a sliding surface register in the AI ​​dynamic optimization model of fertilizer application amount, mapping the crop nutrient deviation and its time accumulation into a sliding surface vector; reading the state vector difference in real time, forming a multi-dimensional deviation mark through a difference comparator and writing it into a deviation latch; the output of the deviation latch is sent to the Daltonian switching law generator to generate a symbol matrix; the Daltonian switching law generator matches the symbol matrix dimension by dimension according to the pre-stored switching gain table to generate an initial switching pulse sequence; the adaptive gain modulator increases or decreases the switching gain according to the real-time threshold of the deviation amplitude and the accumulated part of the historical residual of the AI ​​fertilizer application amount, and outputs a gain modulation matrix; the initial switching pulse sequence is multiplied by the gain modulation matrix to obtain an open-loop control sequence, which is double-shaped by a rate limiter and a boundary layer filter to suppress high-frequency jitter;.

[0020] Furthermore, the system also includes: a water and fertilizer feedback control unit, which is used to measure the crop moisture status index, combine the water application amount, calculate the dynamic correction water application amount of the crop per unit area at each time, and perform closed-loop feedback regulation of crop water application; and measure the crop nutrient adequacy index, combine the fertilizer application amount, calculate the dynamic correction fertilizer application amount of the crop per unit area at each time, and perform closed-loop feedback regulation of crop fertilization.

[0021] Furthermore, the execution process of the water and fertilizer feedback control unit includes: reading the dynamic correction water application amount and the dynamic correction fertilizer application amount stored in the previous operation cycle, writing them into the water feedback register and the fertilizer feedback register, establishing a dual-channel error latch and clearing them to zero; real-time acquisition of the measured crop water content status index and the measured crop nutrient adequacy index, and writing them into the state observation buffer after suppressing high-frequency noise through a double-window index filter; calling the current water application amount and fertilizer application amount, forming a water error vector and a nutrient error vector in the error latch respectively, and recording the change rate mark; performing dual-loop proportional-integral modulation, including: the water error vector enters the water integral modulator to generate a water integral; the nutrient error vector enters the fertilizer integral modulator to generate Fertilizer integral; the two channels simultaneously perform proportional-integral superposition and output the first correction value; the first correction value is limited to the crop safety threshold by the limiter saturator, and the second correction value is generated and written into the dynamic correction register; the second correction value in the dynamic correction register is added to the original water application amount and the original fertilizer application amount one by one to obtain the dynamic correction water application amount and the dynamic correction fertilizer application amount of the crop per unit area at the current time, and pushed to the execution instruction queue at the same time; after irrigation and fertilization are completed, the measured crop water status index and the measured crop nutrient adequacy index are re-collected; if the absolute value of any error vector does not decrease for two cycles, the proportional-integral gain is increased and the dual-loop proportional-integral modulation is returned, otherwise the current gain is maintained to enter the next normal cycle.

[0022] Furthermore, meteorological data include: crop evapotranspiration , in mm / day; rainfall , in millimeters per day; the proportion of rainfall that is not absorbed by the soil and flows directly into the runoff is expressed as surface runoff rate To express; crop growth status data include: soil water potential , unit is kPa; crop root suction , unit is kPa; nitrogen demand of crops per unit area , in g / m² / day; the phosphorus requirement of crops per unit area , in g / m² / day; the potassium requirement of crops per unit area Unit is g / m² / day; soil adsorption rate ; Leaf chlorophyll concentration ; For time.

[0023] Furthermore, the estimated water requirement of crops per unit area at each time is:

[0024] ;

[0025] in, for The water requirement of crops per unit area at each time; the fertilizer requirement of crops at each time is estimated to be:

[0026] ;

[0027] in, for The amount of fertilizer required by crops per unit area over time; for The theoretical chlorophyll concentration of the crop at the current growth stage.

[0028] The present invention's integrated water and fertilizer multi-objective optimization intelligent control system driven by artificial intelligence has the following beneficial effects: by constructing a complete control system integrating data perception, collaborative decision-making, dynamic optimization and closed-loop feedback, it realizes the intelligent integration and real-time precise management of crop irrigation and fertilization. Compared with traditional single-channel, quantitative or threshold-triggered water and fertilizer control technologies, the present invention introduces a dual-channel AI dynamic optimization engine architecture at the control layer for the first time, and couples the AI ​​dynamic optimization model of water application and the AI ​​dynamic optimization model of fertilizer application with the historical residual compensation mechanism, effectively achieving fine fitting and continuous correction of the dynamic demand for crop water and fertilizer. At the execution level, the water and fertilizer feedback control unit monitors the crop moisture status index and nutrient adequacy index in real time, realizing closed-loop tracking and dynamic correction of the actual control results, thereby improving the adaptability and stability of the control system. In terms of scheduling strategy, the system adopts a small time difference execution logic of water application priority and fertilization delay, which significantly reduces the risk of rhizosphere loss. To ensure system stability, this invention utilizes linear quadratic Lyapunov control and sliding mode control techniques to ensure asymptotic convergence of the system control solution throughout the entire cycle, preventing long-term error drift from damaging the crop root zone. Overall, this invention effectively improves water and nutrient utilization efficiency, reduces agricultural input waste, and enhances the system's robustness to environmental perturbations. It is suitable for large-scale automated deployment across multiple crop types, climate zones, and complex terrains, demonstrating its potential for widespread application and promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of the system structure of the artificial intelligence-driven integrated water and fertilizer multi-objective optimization intelligent control system provided in an embodiment of the present invention;

[0030] Figure 2 A schematic diagram of the sliding mode control effect of the AI ​​dynamic optimization engine for fertilizer application provided in an embodiment of the present invention;

[0031] Figure 3 Schematic diagram of the synergistic effect between the water and fertilizer load analysis unit and the AI ​​dynamic resource optimization engine provided by an embodiment of the present invention;

[0032] Figure 4A schematic diagram comparing the control effects of the AI ​​dynamic optimization engine for water application provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0034] refer to Figure 1 : An integrated water and fertilizer multi-objective optimization intelligent control system driven by artificial intelligence, the system includes: a water and fertilizer load analysis unit and a water and fertilizer joint control criterion generation unit; the water and fertilizer load analysis unit is used to obtain meteorological data from the meteorological monitoring system; at the same time, it receives data from soil sensors and leaf sensors to obtain crop growth status data, and evaluates the water and fertilizer requirements of crops per unit area at each time; the water and fertilizer joint control criterion generation unit is used to use the AI ​​dynamic resource optimization engine to dynamically calculate the amount of water and fertilizer applied to crops per unit area at each time according to the soil status and the root absorption capacity of the crops; the AI ​​dynamic resource optimization engine includes: an AI dynamic optimization engine for water application and an AI dynamic optimization engine for fertilizer application; the AI ​​dynamic optimization engine for water application is obtained by adding the AI ​​dynamic optimization model for water application to the accumulated part of the AI ​​historical residual of water application; the AI ​​dynamic optimization engine for fertilizer application is obtained by adding the AI ​​dynamic optimization model for fertilizer application to the accumulated part of the AI ​​historical residual of fertilizer application.

[0035] Farmland is considered an open thermodynamic system that continuously evolves over time: radiation, rainfall, and evapotranspiration from the atmospheric boundary layer result in a continuous exchange of water energy; while ion adsorption and desorption, as well as microbial mineralization in the rhizosphere, determine the real-time transfer of nutrient potential energy. The water and fertilizer load analysis unit simultaneously monitors these two energy pathways to construct a dynamic causal chain centered on the "state-driver-response" principle. This causal chain first normalizes the scale of the exogenous driver from the meteorological monitoring system. It then aligns the time series with the water status data from soil sensors and the photosynthetic pigment concentration data from leaf sensors, and then maps these three types of information into the same multidimensional phase space. This mapping process employs an end-to-end temporal embedding strategy, ensuring that each sample occupies a unique coordinate in the phase space, forming a continuous trajectory. The tangent vector of the trajectory in space describes the instantaneous rate of change of water and fertilizer requirements, while the normal vector reflects the crop's vulnerability to external perturbations. The water and fertilizer load analysis unit uses this "cut-method" coupling relationship to convert complex and changeable physiological processes into smooth curves that can be processed instantly by computers, and then obtain the water and fertilizer requirements per unit area at each time point.

[0036] The unit for generating criteria for combined water and fertilizer regulation resides in the system's "decision-making layer." Its principle can be understood as finding a flexible, updatable balance point between the time-varying demand curve, the limited supply threshold, and environmental constraints. The AI ​​dynamic resource optimization engine is the computational core of this unit. It simultaneously deploys an AI dynamic optimization engine for water application and an AI dynamic optimization engine for fertilization. These two pathways adhere to the same control philosophy in principle, but maintain differentiation in implementation details to avoid the explosion of solution dimensions caused by large-scale coupling. The AI ​​dynamic optimization model for water application within the AI ​​dynamic optimization engine considers "water dissipation through the soil pores via dual channels of seepage and transpiration" as the primary control objective, constructing an asymptotically stable domain with the aid of an energy function. The accumulated AI historical residuals for water application record the deviations between the model's predictions and the actual feedback, injecting these deviations into the control law as novel disturbances to achieve adaptive gain adjustment. This parallel superposition of "model output-residual compensation" architecture allows the AI ​​dynamic optimization engine for water application to quickly correct the width and interval of irrigation pulses through residual drive even if external meteorological conditions change drastically, thereby maintaining a steady decline in the root zone water potential energy.

[0037] The principle of the AI ​​dynamic optimization engine for fertilizer application is approached from the perspective of sliding mode control. Its core is to set a sliding target hyperplane for the "crop absorbable ion concentration." The AI ​​dynamic optimization model for fertilizer application calculates the sliding surface distance in real time based on the difference between the crop root absorption capacity and the soil buffering capacity, and then uses the Darwin switching law to generate a bipolar pulse sequence, so that the system state slides to the hyperplane with a high slope within a finite time. On this basis, the accumulated historical residuals of the AI ​​fertilizer application continue to track the accumulation and distribution of nutrient deviations, and convert this information into additional compensation vectors to suppress high-frequency oscillations in concentration. In principle, this "two-stage" strategy ensures a rapid response of the model level to step disturbances, while also making the residual channel immune to low-frequency drifts in the long term.

[0038] The AI ​​dynamic resource optimization engine utilizes an asynchronous parallel mechanism for overall scheduling: the AI ​​dynamic optimization engine for watering and fertilizing runs in independent threads. After each engine completes prediction, compensation, overlay, and instruction issuance, the scheduling coordinator implements serial queuing within millisecond time slots. The scheduling coordinator implements internal priority arbitration rules. When the system detects that irrigation operations may alter the migration path of rhizosphere ions, it temporarily increases the time slot allocation for fertilization instructions to ensure that water pulses precede nutrient pulses, thus adhering to the agronomic principle of "water first, fertilizer later, and fertilizer application during water flow."

[0039] The target water and fertilizer application rates output by the water-fertilizer control criterion generation unit enter the field execution layer and are distributed via a high-speed fieldbus to the solenoid valve array and variable-flow fertilizer application device. The execution layer hardware follows an event-driven protocol, initiating actions only upon receipt of a timestamped pulse packet to avoid resource interlocks caused by multitasking. After each hardware action, the execution layer immediately generates a feedback packet and transmits it back to the water-fertilizer feedback control unit. The water-fertilizer feedback control unit essentially decomposes the large system into two independent integral loops: a water loop processes the water status index, and a fertilizer loop processes the nutrient adequacy index. The dual-loop proportional-integral modulator is assigned different integral time constants to enhance the system's adaptability to diverse dynamic characteristics. A limiter and saturator are implemented after the two loops to mitigate abnormal peaks from impacting crop safety thresholds. The corrected compensation value is rewritten into the execution buffer and simultaneously written to the accumulated historical residuals of the water and fertilizer application AI algorithms, forming a truly closed "measurement-decision-execution-feedback-compensation" loop.

[0040] The system uses multi-source fusion feature mapping within the water and fertilizer load analysis unit to transform the nonlinear, strongly coupled, and time-varying characteristics of crop demand into continuously trackable curve trajectories. The water and fertilizer joint control criterion generation unit then decomposes the resource allocation problem into different dimensions through an "independent model prediction-residual compensation-asynchronous parallel scheduling" mechanism. Finally, the water and fertilizer feedback control unit achieves rapid correction and asymptotic equilibrium through dual-loop proportional-integral modulation. Each submodule within the three-layer architecture incorporates adaptive capabilities: load analysis relies on a continuous data stream to update feature mapping; the dynamic optimization model reconstructs the stability domain or sliding surface based on real-time errors; and residual accumulation uses accumulated deviations as a trigger for regulation, offsetting model assumption drift over the long term. This global-local dual-scale adaptive framework ensures that the system maintains optimal water and nutrient supply in real time, regardless of short-term meteorological shocks or long-term physiological changes. It also allows artificial intelligence in agricultural scenarios to transcend the controversial "black-box" optimization approach. Residual compensation and explicit stability criteria establish an interpretable, verifiable, and traceable control chain.

[0041] Furthermore, the execution process of the AI ​​dynamic optimization engine for watering volume includes: calling the residual sequence of the most recent complete growth cycle in the historical residual accumulation section of the AI ​​watering volume, setting the negative definite convergence domain based on the linear quadratic Lyapunov function, and writing it into the residual accumulation register and the safety threshold register; when calling the residual sequence of the most recent complete growth cycle in the historical residual accumulation section of the AI ​​watering volume, the system first retrieves the deviation trajectory stored in timestamp order in the solid-state cache and performs an integrity check on it. Any missing segments are linearly interpolated based on adjacent observation points to ensure that the deviation trajectory covers the entire moisture response characteristics of the crop from sowing to harvest. Subsequently, the deviation trajectory enters a dual-weight exponential attenuation filter, which compresses the residual energy within a continuous and smooth bandwidth by simultaneously attenuating high-frequency noise and low-frequency drift; this bandwidth directly determines the weight matrix range of the subsequent linear quadratic Lyapunov function. The system generates a symmetric positive definite weighted matrix in the temporary operation register area, and constructs a linear quadratic Lyapunov function based on the residual vector after filtering. The energy level of this function is used as a benchmark to calculate the descent rate point by point along the time axis. If a non-negative descent rate appears at any point, the system will implement fine-grained adjustments on the weight matrix through an adaptive shrinkage algorithm until all sampling points are in a strictly declining state. After the verification is completed, the system uses the maximum acceptable value of the energy level as the boundary of the negative definite convergence domain, first writes the complete residual vector into the residual accumulation register for adaptive gain compensation in the next cycle; then writes the boundary of the negative definite convergence domain into the safety threshold register to provide a hard constraint for the real-time control law. Through this process, the water-fertilizer integrated multi-objective optimization intelligent control system ensures that the water control loop can converge within the negative definite convergence domain under any external disturbance, while laying a stable foundation for subsequent residual compensation and gain self-tuning.

[0042] Soil conditions and crop root absorption capacity are collected in real time and written to the state observation buffer along with irrigation records from the previous period. After unified scaling, the state vector is formed. The system's tensiometers, conductivity probes, and variable-depth root pressure probes deployed within the field simultaneously output soil moisture trends, changes in ion exchange capacity, and rhizosphere osmotic potential with sub-second resolution. The acquisition thread synchronously writes these high-frequency streams of data into a high-speed circular buffer using a timestamp alignment mechanism. A parallel irrigation execution thread immediately writes the valve duty cycle, pulse width, and flow integral to a historical segment in the same physical address space after each valve pulse. The system aggregates these two streams of data at the on-chip bus level using dual-channel DMA to form a complete raw observation frame. Subsequently, a unified scaling pipeline is triggered within the FPGA logic. The first stage of the pipeline calls the discrete dynamic benchmark library to assign a corresponding real-time benchmark to each raw metric type. It then performs threshold clipping for sudden high-amplitude noise and sliding window drift removal for gradual trends across days. Finally, a parameter-free normalization based on quantile adaptation is used to convert all metrics to a unified dimensionless interval. The final stage of the pipeline uses a configurable concatenation table to sequentially push fields such as water gradient, absorption rate, and historical valve pulse width into a fixed-length array. This generates a "state vector" in the on-chip cache. This state vector is mapped to the input registers of the AI ​​dynamic resource optimization engine via the AXI bus. This not only ensures that each optimization iteration uses data with the same physical time tag, but also ensures the numerical stability of model weights through scale alignment, regardless of dimensional differences. This ensures that the AI ​​dynamic optimization engines for water and fertilizer application can converge to a unique solution even in high-concurrency scenarios, and deliver the judgment criteria to the execution layer within milliseconds.

[0043] Upon receiving the state vector, the AI-powered dynamic optimization model for watering quantity immediately locks onto a control law template that matches the current crop stage and instantiates a linear matrix inequality solver in the on-chip cache. Using a distributed interior-point method, the solver synchronously writes the second-order approximation of the objective function and the Lyapunov descent constraints into a reconfigurable logic array. Through an iterative process, the decision variables are continuously adjusted within minimal steps, ensuring that the residual error decays monotonically with each iteration. To avoid real-time performance degradation caused by excessive iteration depth, the solver is hardware-partitioned into a main core and parallel co-cores. The main core is solely responsible for searching the global feasible region, while the co-cores simultaneously perform local step-size tests on multiple candidate directions. Once the Lyapunov descent constraints are satisfied and the numerical condition number is within a safe range, an "early stopping" signal is triggered, directly writing the solution vector corresponding to that direction as the raw watering control variable to the output register bank. Before being written out, the raw watering control variable undergoes a pipelined saturation protection and quantization encoding, ensuring plug-and-play compatibility with the timing pulse protocol of the solenoid valve driver stage. Meanwhile, an adaptive gain modulator resides in a bypass channel of the model's computational path, monitoring the slope of the Lyapunov function along the time axis in real time. It then subtracts this slope from the system's built-in target convergence rate. This difference, after being accumulated through proportional pilot and integral summation, is mapped to a matrix weight adjustment coefficient. This coefficient directly applies to the weight matrix elements of the linear matrix inequality equation in the next iteration, achieving fine-grained compression or amplification of the control law gain. When significant external disturbances cause convergence to slow, the adaptive gain modulator increases the weights of key diagonal elements within milliseconds to rapidly expand the Lyapunov descent margin. Once the system enters a stable phase, the modulator gradually reduces the gain to prevent root hypoxia caused by excessive irrigation. This entire process forms a closed loop at the hardware level, maintaining stable, fast, and energy-efficient watering control without requiring software interruptions, thereby providing a reliable baseline trajectory for the subsequent residual compensation phase.

[0044] The AI ​​system accumulates historical residual errors in watering volume, immediately invoking real-time error calculations after each irrigation feedback loop completes. The latest error is written to the residual accumulation register at the same physical address. A sliding window is maintained within the register, ensuring that the cumulative curve grows continuously over time without slowing down response due to outdated data. When a threshold event is triggered by a sliding window update, the variable gain compensator is synchronously invoked. This compensator uses a table lookup to obtain a gain baseline that matches the current crop growth stage. It dynamically adjusts the amplification factor based on the root mean square (RMS) value of the residual within the window and applies this factor directly to the latest residual, generating a compensation value with the same dimensions as the original watering control value. At the hardware level, the compensation value is added dimensionally to the original watering control value via a pipeline adder. The resulting instantaneous vector then enters the saturation protector. The saturation protector sets upper and lower limits based on the soil structure safety threshold, the crop root pressure tolerance threshold, and the maximum flow rate threshold of the irrigation equipment. Any components exceeding this range are hard-truncated to ensure that the target watering volume neither causes soil erosion nor water shortages. The processed target watering volume is directly encoded on-chip as a protocol frame and written into the execution instruction queue. After comparison with the previous queue element, a hardware arbiter determines the insertion position, ensuring that all irrigation pulses arrive at the valve control node in a logical chronological order. This three-level coordinated implementation of residual accumulation, variable gain compensation, and saturation protection enables the system to rapidly correct the control vector through compensation when faced with rapid disturbances such as sudden high wind evaporation or rewetting after a brief rainstorm. It also maintains overall operating boundaries through the saturation protector, achieving a balance of high precision and high safety.

[0045] Furthermore, when executing step A1, the bus controller is first used to quickly map the accumulated residual sequence of the AI ​​historical residuals of water application from non-volatile storage to the on-chip dynamic cache. Each data unit has a sequentially increasing timestamp and cyclic redundancy check bit. The integrity check logic compares the timestamp interval and the check bit value one by one in a pipeline manner on the chip. If a timestamp gap is found, a marker is immediately inserted into the cache to take up the space. If a redundancy check discrepancy is detected, the corresponding unit is marked as invalid. All markers are then sent to the missing segment reconstruction module. The reconstruction module assembles a list of continuous missing segments according to the markers and calls a linear interpolation operator. The operator extracts the most recent valid data from both ends of the missing segment as the endpoint, and generates an interpolation sequence with equal step size in the on-chip multiplication and addition array. The entire process relies on hardware parallelism to ensure that the completed data maintains a smooth transition with the original sequence and does not introduce artifacts beyond the range. After the interpolation is completed, the verification module calculates the cyclic redundancy check bit for the newly added data again and updates the timestamp. Finally, the verified sequence is written back to the accumulated part of the AI ​​historical residual of the water application amount. At the same time, an interrupt is triggered to send a state synchronization signal to the adaptive gain modulator to ensure that subsequent control law iterations can run in a data domain without gaps, thereby stabilizing the timing consistency of the integrated water and fertilizer multi-objective optimization intelligent control system.

[0046] When the system writes the verified sequence to the residual buffer stack, it first calls the on-chip memory controller to allocate a contiguous address space to ensure the physical contiguity of the residual data along the timeline. This design enables subsequent pipeline processing to complete the migration of the entire frame in a single burst read / write, avoiding wait cycles caused by cross-page access. After the write is complete, the residual buffer stack automatically updates the top-of-stack pointer and generates a new data arrival interrupt signal, triggering the dual-weight exponential decay filter to operate. The dual-weight exponential decay filter utilizes a two-stage multiply-add structure: the first stage uses a larger attenuation coefficient to rapidly attenuate transient high-amplitude peaks, while the second stage uses a smaller attenuation coefficient to smooth inter-cycle low-frequency drift. These two stages operate in parallel in hardware through a shared multiplier array, ensuring that each data sample completes the dual attenuation within a single clock cycle. The filter also incorporates dynamic coefficient adjustment logic to fine-tune the two-stage attenuation ratio in real time based on the variance of the most recent samples in the residual buffer stack, ensuring optimal bandwidth even when external weather conditions change rapidly. The processed data directly overwrites the same address in the residual buffer stack, maintaining the compact structure of the time series and outputting a read-only copy for subsequent linear quadratic stability determination. At this point, the filter sequence is synchronized across both the system memory and cache storage levels. This not only provides a high-frequency noise-free error driver for the AI ​​dynamic optimization engine for water application, but also achieves millisecond-level latency through hardware-level parallelism, laying a solid foundation for the real-time and stability of the integrated multi-objective optimization and intelligent control system for water and fertilizer.

[0047] After the filter sequence is written back to the residual buffer stack, it is immediately fed into the range detection module. This module consists of two parallel pipelines: one extracts the maximum value of the current window in a single-cycle pipeline on the cache read side, and the other simultaneously extracts the minimum value. Both pipelines share the same address generator to ensure data alignment. The maximum and minimum values ​​are hardwired to subtract the range result, which is fed into a hardware comparator connected to the safety threshold register on the next clock cycle. The safety threshold register is programmed with a fixed set of upper and lower limits by the agronomic expert library during system power-up calibration. The adaptive gain modulator can fine-tune this limit during runtime based on long-term statistical drift. When the range detection value first exceeds the threshold in the safety threshold register, the comparator outputs an overload flag to the interrupt controller, which generates a high-priority exception signal, immediately freezes the residual buffer stack write pointer, locks the current filter sequence version, and issues a backtrace instruction to the exception backtrace logic. The abnormal backtracking logic directly accesses the verified sequence backup area written in step A1 through DMA, reloads the original data with the fastest bus width on the chip to overwrite the filter sequence, and then clears the attenuation weight history of the residual buffer stack and resets the internal state of the dual-weight exponential decay filter to put it in the initial gain configuration. After completing the data reload, the abnormal backtracking logic returns a reset pulse to the extreme difference detection module, cancels the freeze and releases the write pointer, and the system automatically jumps back to step A1 to re-execute the integrity check and linear interpolation process. Through the hardware closed loop formed by the extreme difference detection module, comparator, safety threshold register, interrupt controller and abnormal backtracking logic, the system can cut off the harmful data link within microseconds when abnormal spikes or continuous drifts appear in the residual sequence, and roll back to a stable data state, ensuring that the water-fertilizer integrated multi-objective optimization intelligent control system can maintain the convergence reliability of the control law and the numerical security of the hardware path under any extreme environmental disturbance;

[0048] During step B1, the system first performs a fast variance scan on the filter sequence within the current sliding window. This scan utilizes an on-chip vectorized multiply-add array to obtain the variance value for each dimension within a single cycle. The system then calls the eigenstatistics logic to map these variance values ​​into a set of diagonal weights, forming an initial draft of the positive-definite weighted matrix. Subsequently, a controlled perturbation is injected onto each diagonal line via the diagonal compensation unit to ensure that all eigenvalues ​​of the matrix are strictly positive, and the generated results are immediately written to a dedicated matrix register. When step B2 is initiated, the controller performs a quadratic operation on the newly written positive-definite weighted matrix and the residual state vector in a multi-port arithmetic logic unit, constructing the linear quadratic Lyapunov function in real time. Once constructed, the gradient iterator immediately performs a batch gradient evaluation on the same data channel. This iterator computes the rate of descent of the Lyapunov function along the time axis in parallel for all samples in the window. The rate of descent vector is fed back to the controller in the next clock cycle. If the gradient iterator detects a non-negative descent rate for any sample, step B3 is triggered. The system calls the adaptive contraction logic to proportionally compress all diagonal weights of the positive-definite weight matrix and simultaneously updates the matrix registers, then reenters step B2 for reverification. This loop runs continuously on the hardware pipeline at nanosecond intervals until all descent rates returned by the gradient iterator are negative, indicating that the linear quadratic Lyapunov function has achieved strict descent within the current energy surface. At this point, the positive-definite weight matrix is ​​locked, and the final version in the matrix registers will serve as the core weights for subsequent control law solutions. It will also be written to the log buffer for long-term tracking by the adaptive gain modulator, providing traceable stability proof for the integrated multi-objective optimization intelligent control system for water and fertilizer.

[0049] When the system enters step C1, the controller first calls the linear quadratic Lyapunov function energy curve locked by the previous round of gradient iterators, searches for its energy peak along the time axis, and marks it as the maximum acceptable energy level. This level value is immediately written to the register mapping table and written to the negative definite convergence domain register through a hardwired path. The upper and lower edge dual threshold marks are generated within the register to ensure that any subsequent state evolution will be immediately captured by the hardware limiting logic once it crosses the boundary. The system then enters step C2, copying the boundary parameters in the negative definite convergence domain register via the high-speed bus to the local register group of the safety lockout logic and the water application AI dynamic optimization engine. After obtaining the boundary parameters, the safety lockout logic automatically refreshes the internal comparator threshold, causing it to perform real-time cross-bounds detection on each upcoming frame of energy data. Once the detection result is non-zero, an emergency irrigation stop or a reduction sequence is triggered, ensuring that the field irrigation hardware does not exceed the crop tolerance range under extreme disturbances. After the AI ​​dynamic optimization engine for water application synchronizes to the same boundary parameters, it writes them into the energy constraint register of the control law solver. This ensures that the linear matrix inequality solution process has the latest boundary reference at each step size estimation, thus preventing convergence domain diffusion caused by sudden environmental changes. Synchronization is accomplished through a hardware handshake protocol, during which any clock drift is forcibly aligned by the timing corrector, ensuring that the negative definite convergence domain boundaries remain absolutely consistent across the perception, decision-making, and execution layers. This allows the integrated multi-objective optimization intelligent control system for water and fertilizer to consistently lock onto the same stable attraction domain during long-term operation, achieving precise, asymptotic, and verifiable dynamic control of irrigation resources.

[0050] During the first clock tick of its startup, the AI ​​dynamic optimization engine for fertilizer application batch-loads the residual sequence for the most recent complete growth cycle from the AI ​​historical residual accumulation section via a high-speed bus. These residuals are immediately written to the residual buffer stack. The residual buffer stack uses a monolithic ring storage structure to ensure physical data continuity along the timeline, allowing all residual values ​​to be read sequentially by downstream logic without page jitter. During the same cycle, the system calls the sliding mode initialization unit, subtracting the top value of the residual buffer stack from the center value of the crop nutrient safety zone. This difference is used to calibrate the initial value of the sliding mode surface and record the zero-bias baseline in the register table, serving as a symmetric reference surface for subsequent deviation integration. The subsequent perception phase uses a multi-threaded DMA to simultaneously pull data streams on soil state and crop root absorption capacity, and reads fertilization records from the previous period at the field level. A hardware scale alignment pipeline compresses all indicators into a unified dimensionless interval, which is then concatenated into a state vector in a fixed field order in the state observation buffer. The AI-powered dynamic optimization model for fertilizer application then begins solving the sliding mode control problem. The difference between the sliding mode surface and the state vector serves as the driving variable and enters the Darwin switching law generator. This generator internally triggers a parallel table lookup on the high-dimensional symbol matrix, matching the sign bits with the preset gains dimension by dimension, and outputs the original fertilizer control variable. This output maintains an integer-ratio structure throughout the entire logic path, facilitating direct interpretation by subsequent actuators. The adaptive gain modulator monitors the sliding mode surface's convergence rate over time and dynamically compares the actual approach slope with the target approach speed. If the actual speed is slower, the switching rate is increased within the same cycle; if it is faster, the switching rate is gradually reduced to avoid overshooting ion concentration accumulation at the root zone. The AI-powered historical residual accumulation component accumulates nutrient deviations at the end of the current feedback cycle. This is then mapped to a compensation vector of the same dimension as the original fertilizer control variable through a variable gain compensator. This compensation vector is then added dimension by dimension with the original fertilizer control variable in a hardware parallel adder. The overall vector is then clipped by the upper and lower thresholds of the saturation protector. Any component exceeding the crop's salt tolerance threshold or the device's maximum flow rate threshold is immediately truncated. The trimmed fertilizer rate vector is encoded into a fieldbus-compatible command frame and written into the execution command queue. The queue scheduler sorts the pulse packets by timestamp to ensure that they arrive at the variable-rate fertilization device in sequence. The entire process, centered on on-chip parallel logic, ensures that residual drive, sliding mode solution, gain modulation, compensation superposition, and safety saturation are completed within a closed control cycle. This enables the AI ​​dynamic optimization engine for fertilizer rate to maintain a smooth transition and asymptotic stability of rhizosphere nutrient concentration even in the face of sudden temperature drops or drastic changes in soil microbial activity.

[0051] Furthermore, in the AI-powered dynamic optimization model for fertilizer application, the system first allocates a separate register cluster for the sliding surface register within the on-chip programmable logic. The bit width of this register cluster is strictly aligned with the state vector dimensions to ensure lossless mapping of crop nutrient deviations into binary fixed-point numbers. The model then calls an integrator-accumulator unit, integrated within the same logic area, to synchronously feed the real-time sampled crop nutrient deviations and the integral of the deviations within the historical window into the accumulator. The continuous-time cumulative data output by the integrator-accumulator unit is hardwired directly to the sliding surface register, where it automatically concatenates the vectors to generate a sliding surface vector with a stable topology. This vector is written to the shared cache by the DMA engine on the next clock cycle for subsequent computational stages. A second parallel data channel continuously reads the current state vector from the state observation buffer and calculates the difference between the state vector and the previous fertilization record using a dedicated subtractor. This difference is then fed into a difference comparator. The difference comparator uses parallel threshold decision logic to encode the positive and negative polarity and amplitude range of each dimension into multidimensional deviation tags, which are written to the deviation latch in the same clock cycle. The deviation latch uses a hardware grid to maintain each dimension tag at a fixed data line position, ensuring that the Darby switching law generator can read it with zero latency in subsequent cycles. After receiving the multidimensional deviation tags from the deviation latch, the Darby switching law generator retrieves the corresponding switching coefficients from its internal lookup table, converts the positive and negative polarity into symbol matrix elements, and uses pipeline logic to complete the entire matrix assembly within a single clock cycle, ultimately outputting the symbol matrix as the symbol carrier of the control law drive variable. This fully hardware-parallel, pipelined implementation eliminates the need for any software interrupts or cache writebacks during the entire process of generating the sliding mode surface vector, capturing the state vector difference, and constructing the symbol matrix. This significantly reduces round-trip bus latency and enables the AI ​​dynamic optimization model for fertilizer application to complete a full sliding mode control iteration in milliseconds, ensuring smooth transitions in crop root nutrient concentrations despite rapid climate fluctuations or sudden changes in soil ion adsorption.

[0052] The Darwin switching law generator pre-populates a switching gain table that fully corresponds to the dimensions in the on-chip read-only memory. After the symbol matrix enters the generator, the row-first control logic pairs each symbol bit with the gain coefficient in the same column. Multiplication and accumulation are performed in a single-cycle pipeline, and an initial switching pulse sequence is output. This sequence is written to the register array via a wide parallel bus, ensuring that the pulse width and duty cycle are precisely aligned with the clock. The adaptive gain modulator continuously monitors the deviation amplitude and reads the real-time threshold from the accumulated historical residual of the fertilizer application AI. After obtaining the gain modulation factor through the difference amplifier, it calls the increment-decrement microstepping algorithm to update the corresponding entries in the switching gain table by dimension to generate the gain modulation matrix. The gain modulation matrix and the initial switching pulse sequence are vector-multiplied by the on-chip multi-channel parallel multiplication unit to produce the open-loop control sequence. The open-loop control sequence first enters the rate limiter, which compresses any pulse leading edge based on the rising slope limit window to prevent mechanical shock in the valve-controlled actuator caused by overly rapid driving. The sequence then flows into the boundary layer filter, which uses a sliding average kernel to attenuate peak energy within the switching frequency band. The output pulse is compressed in both amplitude and frequency, thereby suppressing high-frequency chattering. The shaped pulse is then written to the execution register, awaiting dispatch by the scheduler to the field-side variable-flow fertilization device. This ensures that the switching law maintains the rapid approximation characteristics of the sliding mode surface while avoiding unnecessary oscillations due to discretization effects.

[0053] The water and fertilizer feedback control unit, located at the core of the closed loop between the execution and decision layers, receives field measurements from the field sensor network after each irrigation and fertilization pulse. Using multiple synchronous acquisition channels, the system converts the real-time outputs of the tensiometer, water potential probe, and leaf reflectance spectrometer into a crop moisture status index. Simultaneously, the synchronous outputs of the ion-selective electrode and fluorescence colorimetric probe are converted into a crop nutrient adequacy index. After hardware-level zero-hold processing and noise clipping, these two indices are subtracted from the water and fertilizer application rates executed during the previous control cycle in the same arithmetic logic unit to generate a moisture error vector and a nutrient error vector. The error vectors first pass through a dual-window exponential filter to eliminate transient spikes before being fed into the water and fertilizer integral modulators, respectively. Within the modulators, a short-term proportional correction and a long-term integral correction are superimposed into a single correction signal. The system then uses variable gain mapping to convert these correction signals into dynamic corrections for water and fertilizer application rates. During this conversion process, limiter and saturation logic implements bidirectional clipping of each component based on crop safety thresholds, soil mechanical thresholds, and equipment physical thresholds to prevent overshoot or undershoot. The trimmed dynamic water and fertilizer rates are written to the execution instruction buffer, where the queue scheduler allocates the next available time slot. This ensures that the dynamic water rate is triggered first in the valve control sequence and that the dynamic fertilizer rate follows with minimal delay after the water pulse ends, fully implementing closed-loop feedback regulation of crop watering and fertilization. After each adjustment cycle, the water and fertilizer feedback control unit writes the new error accumulation value back to the residual accumulation register, providing the latest edge data for the AI ​​historical residual accumulation of water and fertilizer rates. It also encapsulates the current gain configuration, threshold trimming records, and exponential filter status into log frames and synchronizes them uplink to the cloud-based model repository for long-term iterative updates of the control law templates at the decision-making layer. This entire process is completed within millisecond-level hardware channels, ensuring that the dynamic water and fertilizer rates remain closed in real time despite the dual disturbances of rapid climate fluctuations and soil buffering changes, thereby ensuring that the crop water status index and crop nutrient adequacy index return to the target range along an asymptotically stable trajectory throughout the growing cycle.

[0054] The execution process of the water and fertilizer feedback control unit includes: reading the dynamic correction water application amount and dynamic correction fertilizer application amount stored in the previous operation cycle, writing into the water feedback register and the fertilizer feedback register, establishing a dual-channel error latch and clearing it; real-time collection of the measured crop water content index and the measured crop nutrient adequacy index, and writing them into the state observation buffer after suppressing high-frequency noise through a double-window index filter; calling the current water application amount and fertilizer application amount, forming a water error vector and a nutrient error vector in the error latch respectively, and recording the change rate mark; performing dual-loop proportional-integral modulation, including: the water error vector enters the water integral modulator to generate a water integral; the nutrient error vector enters the fertilizer integral modulator to generate a fertilizer integral Component; the two channels simultaneously perform proportional-integral superposition and output a first correction value; the first correction value is limited to the crop safety threshold by a limiter saturator, and a second correction value is generated and written into the dynamic correction register; the second correction value in the dynamic correction register is added to the original water application amount and the original fertilizer application amount one by one to obtain the dynamic correction water application amount and the dynamic correction fertilizer application amount of the crop per unit area at the current time, and pushed to the execution instruction queue at the same time; after irrigation and fertilization are completed, the measured crop water status index and the measured crop nutrient adequacy index are re-collected; if the absolute value of any error vector does not decrease for two consecutive cycles, the proportional-integral gain is increased and the dual-loop proportional-integral modulation is returned, otherwise the current gain is maintained to enter the next normal cycle,

[0055] Furthermore, meteorological data include: crop evapotranspiration , in mm / day; rainfall , in millimeters per day; the proportion of rainfall that is not absorbed by the soil and flows directly into the runoff is expressed as surface runoff rate To express; crop growth status data include: soil water potential , unit is kPa; crop root suction , unit is kPa; nitrogen demand of crops per unit area , in g / m² / day; the phosphorus requirement of crops per unit area , in g / m² / day; the potassium requirement of crops per unit area Unit is g / m² / day; soil adsorption rate ; Leaf chlorophyll concentration ; For time.

[0056] Furthermore, the estimated water requirement of crops per unit area at each time is:

[0057] ;

[0058] This formula comprehensively considers meteorological driving factors, soil moisture dynamics and crop root water absorption capacity, and can be structurally divided into two parts. Part I: Used to describe the actual loss of soil moisture per unit time, where It indicates the evapotranspiration of crops, which is the amount of water lost by crop unit area due to evaporation and transpiration in one day, with the unit of mm / day; Indicates the rainfall on that day, also in millimeters per day. Subtracting the two gives the water deficit faced by crops. Since not all rainfall is absorbed by the soil, there is also surface runoff loss. The system uses The ratio is expressed as follows, which corrects the net input rainwater. It reflects the coupling relationship between the dynamic process of soil water potential changing with time and the water absorption capacity of crop roots. It represents the rate of change of soil water potential per unit time, in kPa / day. This factor represents the suction capacity of crop roots, also expressed in kPa. A rapid drop in soil water potential indicates intense water evaporation or migration, leading to an increased water requirement. Stronger root suction increases the crop's ability to extract water from the soil. These two factors contribute to the water requirement adjustment factor.

[0059] in, for The water requirement of crops per unit area at each time; the fertilizer requirement of crops at each time is estimated to be:

[0060] ;

[0061] in, Represents the time per unit area of ​​crops The demand intensity for nitrogen, phosphorus and potassium is The sum of the three is the theoretical total nutrient requirement; The soil adsorption ratio is a factor used to describe the effect of soil adsorption of ions in fertilizer on the available nutrients. If the soil adsorption capacity is high, more nutrients need to be applied to meet the actual absorption of crops. is the deviation correction factor between the actual chlorophyll concentration of the crop and the theoretical concentration, is the measured value, The chlorophyll concentration of crops reflects their photosynthetic efficiency and nutrient status. When the measured value is lower than the theoretical value, it indicates that the crop is currently nutrient deficient and the fertilizer requirement should be increased accordingly.

[0062] Furthermore, the accumulated residual of AI history of water application for:

[0063] ;

[0064] Accumulated part of AI historical residual of fertilizer application amount for:

[0065] ;

[0066] in, for The water requirement of crops per unit area over time; for The amount of fertilizer required by crops per unit area over time; for The amount of water applied to crops per unit area over time; for The amount of fertilizer applied to crops per unit area over time; for Time-dependent crop water status index; for Time-dependent crop water status index; for Crop nutrient adequacy index of time; for Crop nutrient adequacy index over time.

[0067] In the multi-objective optimization intelligent control system for integrated water and fertilizer management driven by artificial intelligence, the historical residual accumulation mechanism is the core link for realizing intelligent closed-loop feedback control. By comparing the difference between the system predicted output and the actual response of the crop in each control cycle, an error feedback term is formed to dynamically adjust the amount of water and fertilizer applied in the next cycle. This structure not only enhances the self-learning and adaptability of the system, but also constructs a time-series-related strategy correction framework. Specifically, the system establishes historical residual accumulation models for the watering and fertilization processes, which are used to record the control deviation in real time and quantify it as a feedback correction amount. The historical residual accumulation part of the watering amount is defined as follows:

[0068] ;

[0069] The formula has a clear structure and consists of three products, each of which reflects a different feedback mechanism. It represents the difference between the theoretical water demand and the actual water supply in the previous cycle, reflecting the static deviation between the model prediction and the actual supply; if the value is positive, it means that the actual water supply is insufficient, otherwise it means that the water supply is excessive. It represents the relative change of the crop water content index between two consecutive moments, and measures the dynamic impact of water application on the crop water content. If the water content index decreases, this item is positive, indicating that the crop water content has deteriorated and the amount of water application needs to be increased in the next round. The scaling factor is used to normalize the weight of control actions and prevent abnormal extreme values ​​from excessively interfering with the model. Overall, this structure achieves a quantitative accumulation of historical errors, dynamically responding to changes in crop status while maintaining numerical stability, thereby achieving automatic optimization of watering strategies.

[0070] In terms of fertilization control, the system also sets up a corresponding historical residual accumulation module, the structure of which is as follows:

[0071] ;

[0072] The structure of this formula is highly similar to the water application part, and its principles and mechanisms of action are basically the same, but the variables of interest are converted to crop nutrient responses. Indicates the difference between the total amount of fertilizer required calculated by the system and the actual amount of fertilizer applied; the second item The third item represents the change ratio of the crop nutrient adequacy index, reflecting whether the fertilization behavior improves the crop nutrient absorption status; It is used to adjust the overall impact weight and keep the residual range within a reasonable range.

[0073] Furthermore, the amount of water applied is dynamically corrected for:

[0074] ;

[0075] in, is the photosynthetically active radiation intensity, with the unit being μmol / m² / s.

[0076] This formula is the third-level correction mechanism in the entire intelligent watering strategy. Its principle is not based on offline correction of historical residuals, nor does it rely solely on previous water demand estimates or soil state assessments. Instead, it makes fine-grained real-time adjustments to water application based on dynamic feedback of the current physiological state of the crop, forming a terminal closed loop of regulation, so that the actual response of the crop itself can have a direct impact on water input, thereby improving the system's adaptability to nonlinear disturbances and complex growth conditions. First of all, from a structural point of view, the formula is composed of three factors multiplied together. The first term Indicates the degree of deficiency of the crop's current water status index. It is a standardized dimensionless variable, usually ranging from 0 to 1, reflecting the water saturation of crop cells or the state of water tension of leaves. When the value is 1, the crop is fully hydrated and no water adjustment is required. When the value is far less than 1, the crop is in a state of water stress, and the value of this factor will increase, and the dynamic correction strength will increase accordingly. Therefore, this factor reflects the positive correlation between the current degree of crop water shortage and the intensity of water regulation.

[0077] Item 2 It is a water application-photosynthesis ratio regulating factor. Indicates the amount of water applied per unit area in the current period, which is the control base value calculated by the AI ​​dynamic optimization engine and the accumulated historical residuals; is the photosynthetically active radiation intensity, in , which measures a crop's ability to receive sunlight and is one of the direct drivers of transpiration and photosynthesis. This ratio effectively establishes a relationship between water application and the intensity of light energy driving it: when light intensity is high, crop transpiration is high, and water demand increases accordingly; when light intensity is low, water application should be reduced to avoid excess water, which can lead to root hypoxia or disease. Therefore, this factor provides a physiological-environmental synergistic regulation mechanism, standardizing water application efficiency under energy conditions.

[0078] Item 3 This is the first-order derivative of the water status index with respect to time, indicating the changing trend of crop water status. A positive derivative indicates that crop water status is improving; even if water is currently insufficient, the system can temporarily hold back on increasing watering. A negative derivative indicates that water status is deteriorating, necessitating rapid water replenishment to prevent further physiological stress. Therefore, this factor represents a "rate-regulating channel" in regulation, considering not only the state value itself but also its rate of evolution, reflecting the system's ability to respond to dynamic trends.

[0079] Multiply the three terms together to form This is the dynamic emergency water supply compensation at that moment. In actual control, this compensation does not directly replace the main control value, but is superimposed on the original water supply, that is, the system execution amount is:

[0080] ;

[0081] in is the output of the AI ​​engine and the residual module, and It is the correction increment driven by crop status, and the two together determine the actual implementation water volume.

[0082] To illustrate the practical application logic of this mechanism, consider the following example: In a certain control cycle , the crop monitoring system measures the crop moisture status index , which is lower than the target value of 0.70, indicating that the current crop water shortage; the trend of the water content index over time is , indicating that its condition is slowly deteriorating; the current AI engine calculates the water application amount to be mm; photosynthetically active radiation is . Substitute into the formula to calculate:

[0083] ;

[0084] Since the derivative is negative, it indicates that the state is deteriorating. However, the negative result of this term indicates that the water content has not decreased significantly under strong light conditions, so there is no need to significantly increase water input. Conversely, if the derivative decreases significantly or the light is very strong, this term may become positive and there will be a significant increase.

[0085] The greatest advantage of this feedback compensation mechanism is that it incorporates the dynamic physiological response signals of the crop itself as the ultimate control factor for watering behavior, achieving a closed-loop evolution from "environment-prediction-model" to "crop-response-regulation." This is a typical mathematical implementation of crop-responsive intervention logic in smart agriculture. This formula is highly practical at the control execution level, effectively mitigating resource mismatches caused by non-model errors such as light fluctuations and water diffusion hysteresis. It is an indispensable dynamic correction mechanism in precise control structures.

[0086] Furthermore, the amount of fertilizer applied can be dynamically modified. for:

[0087] ;

[0088] in, It is the plant nutrient conversion rate, in g / m² / day.

[0089] First, from a structural point of view, the formula consists of three factors. It is the degree of deviation of the current nutrient adequacy index of the crop and a static judgment of the nutritional status of the crop. It is derived from the integration of crop tissue nitrogen concentration, chlorophyll concentration, stem color, chlorophyll index and other indicators, ranging from 0 to 1, and is a standardized variable. When the index is low, the factor approaches 1, indicating that the crop has sufficient nutrition and does not need additional fertilizer. When the index is low, the factor approaches 1, indicating that the crop is lacking nutrition and needs to increase fertilizer application. This directly constitutes the intensity benchmark of the dynamic correction amount.

[0090] Item 2 It is the ratio factor of fertilization and nutrient conversion efficiency. Indicates the amount of fertilizer applied per unit area in the current period. It is the value obtained by superimposing historical residuals on the basis of the AI ​​dynamic optimization engine, and has taken into account the impact of theoretical demand and deviation from the previous period. Indicates the plant nutrient conversion rate, in units of , which is the ability of plants to assimilate or convert applied nutrients per unit area into organic matter (such as proteins, enzymes, and chlorophyll). This is a dynamic variable determined by crop growth models, metabolic rates, and current climatic conditions. It is typically estimated through a comprehensive evaluation of photosynthesis rate, nitrogen use efficiency, and nutrient redistribution efficiency. A higher ratio indicates a lower efficiency of plant conversion per unit fertilizer application, leading the system to increase supplementation intensity during correction. If efficiency is high, this parameter is automatically reduced to prevent over-fertilization.

[0091] Item 3 This is the first-order derivative of the nutrient adequacy index with respect to time, representing the trend of changes in the crop's nutritional status. A positive value indicates that nutritional status is improving and immediate fertilization may not be necessary; a negative value indicates that nutritional status is deteriorating and that rapid system intervention is necessary. Combined with the previous two terms, this derivative term forms a sensitive gating factor in the dynamic correction formula, enabling the system to perceive the "trend status" of crop nutrition. This mechanism allows the system to determine not only whether the crop's current nutritional status is adequate, but also whether its nutritional status is on a "recovery" or "deterioration" path, thereby determining whether intervention is necessary.

[0092] Combining these three, the system gets This is the current dynamic emergency fertilization compensation amount, which will eventually form the corrected output of the total fertilization amount together with the previous fertilization plan:

[0093] ;

[0094] in It is the combined result of AI optimization and historical correction, which is the static control output. This is the result of dynamic intervention driven by crop physiology. This mechanism significantly improves the system's response speed and physiological matching accuracy, and is a key technical path for shifting the fertilization process from an "environment-centric" to a "crop-centric" approach.

[0095] Figure 2This figure illustrates the sliding mode control effect of the AI ​​dynamic optimization engine for fertilizer application rate. It details the application of the Darby switching law in fertilizer application rate control, particularly highlighting the significant contribution of the adaptive gain modulator to control performance. The horizontal axis represents the 30-day control period, and the vertical axis represents the deviation in crop nutrient sufficiency. Negative values ​​indicate nutrient deficiency, positive values ​​indicate nutrient excess, and zero represents an ideal nutrient balance. The dashed line in the figure indicates the position of the sliding mode surface, which is set at zero deviation, the ideal control target for the system design. The long dashed line represents the Darby switching law control effect without the adaptive gain modulator. This curve shows that starting from an initial deviation of -0.4, while the system gradually approaches the sliding mode surface, the convergence rate is relatively slow, with a significant slowdown as it approaches the sliding mode surface. Ultimately, zero deviation has not yet been fully reached on the 30th day. The solid line represents the Darby switching law control curve with the adaptive gain modulator, demonstrating significantly improved control performance. This curve not only has a faster initial convergence rate, but more importantly, it reaches near the sliding mode surface around the 15th day and remains stable thereafter. The rectangular area marked with thin dotted lines in the figure represents the range of action of the boundary layer filter, which is located within the ±0.05 deviation range near the sliding mode surface. Within this range, the boundary layer filter effectively suppresses high-frequency chattering, making the control signal smoother and avoiding the common chattering problem in traditional sliding mode control. By comparing the two curves, it can be clearly seen that the adaptive gain modulator dynamically adjusts the switching gain based on the deviation amplitude and the real-time threshold of the accumulated part of the AI ​​historical residual of the fertilizer application amount, significantly improving the system's response speed and control accuracy, and verifying the technical advantages of the variable gain compensator conversion equal-dimensional compensation method.

[0096] Figure 3This diagram illustrates the synergistic effect of the water and fertilizer load analysis unit and the AI ​​dynamic resource optimization engine. This diagram comprehensively compares the resource utilization efficiency of different control strategies over a complete growing cycle. The horizontal axis represents the 180-day growing cycle, and the vertical axis represents the resource utilization efficiency index. Higher values ​​indicate better utilization of water and fertilizer resources. The long dashed line in the figure represents the performance curve of the traditional water and fertilizer management method. This curve maintains a relatively gentle upward trend throughout the growing cycle, gradually increasing from an initial 0.2 to 0.26. The overall improvement is limited, and the rate of increase remains largely constant, lacking the ability to adapt to changing demand at different growing stages. The medium dashed line represents the control effect of the water and fertilizer load analysis unit alone, demonstrating significant improvement over the traditional method. Resource utilization efficiency increases from 0.3 to 0.185, demonstrating the technical advantages of load analysis based on meteorological data, soil sensors, and leaf sensors. The thick solid line represents the final result of the synergistic effect of the water and fertilizer load analysis unit and the AI ​​dynamic resource optimization engine, demonstrating the most superior control performance. This curve not only has the fastest initial improvement speed, but more importantly, it reaches a high efficiency level of more than 1.0 after the 90th day, and finally stabilizes at around 1.4, achieving nearly 5 times the performance improvement compared to the traditional method. The solid dots in the figure indicate the time nodes where the historical residual accumulation plays a key role, which are located at the 60th, 90th, 120th and 150th days respectively. These nodes correspond to the key transition periods of different growth stages of crops. The rectangular area marked with a fine dotted line box represents the main action range of the variable gain compensator, which is between the 60th and 120th days. In this range, the compensator effectively improves the adaptive adjustment ability of the system by accumulating the deviation between the predicted state and the measured state. The legend clearly distinguishes the three different control strategies and the key points of historical residual accumulation, verifying the significant technical advantages of the AI ​​dynamic resource optimization engine in the coordinated control of water and fertilizer.

[0097] Figure 4This figure shows a comparison of the control effects of the AI ​​dynamic optimization engine for water application. It clearly demonstrates the significant advantages of the AI ​​dynamic optimization engine for water application in this invention over traditional control methods in controlling crop moisture status errors. The horizontal axis represents the time axis, measured in days, covering the full 60-day observation period; the vertical axis represents the numerical change in crop moisture status error, reflecting the degree of deviation between the actual moisture status and the ideal moisture status. The dashed line in the figure represents the control curve of the traditional control method. It can be observed that this curve exhibits significant fluctuations throughout the entire time period, with the error value consistently remaining between 0.26 and 0.29 and exhibiting repeated oscillations, indicating that the traditional method has difficulty achieving precise moisture status control and poor system stability. In contrast, the control curve of the AI ​​dynamic optimization engine for water application, represented by the solid line, exhibits significantly different characteristics. Starting from an initial error value of 0.30, the curve shows a clear downward trend and stabilizes within a low error range below 0.08 after the 30th day. The dotted line in the figure indicates the boundary of the negative definite convergence region set based on the linear quadratic Lyapunov function, which is located at an error value of 0.12. It can be clearly seen from the figure that the control curve of the water application AI dynamic optimization engine successfully entered the convergence region around the 25th day and remained stable within the region in the following period, fully verifying the effectiveness of the negative definite convergence region setting method of the Lyapunov function. This phenomenon shows that by calling the residual sequence of the most recent complete reproductive cycle in the accumulated part of the historical residual of the water application AI, combined with the synergy of the dual-weight exponential decay filter and the gradient iterator, the system can achieve asymptotically stable control and effectively suppress the control oscillation problem commonly seen in traditional methods.

[0098] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-driven, integrated water and fertilizer multi-objective optimization intelligent control system, characterized by: The system includes: a water and fertilizer load analysis unit and a water and fertilizer joint control criterion generation unit; the water and fertilizer load analysis unit is used to obtain meteorological data from a meteorological monitoring system; simultaneously receive data from soil sensors and leaf sensors to obtain crop growth status data, and evaluate the water and fertilizer requirements of the crops per unit area at each time; the water and fertilizer joint control criterion generation unit is used to use an AI dynamic resource optimization engine to dynamically calculate the amount of water and fertilizer applied to the crops per unit area at each time based on the soil status and the root absorption capacity of the crops; the AI ​​dynamic resource optimization engine includes: an AI dynamic optimization engine for water application and an AI dynamic optimization engine for fertilizer application; the AI ​​dynamic optimization engine for water application is obtained by adding an AI dynamic optimization model for water application to the accumulated part of AI historical residuals for water application; the AI ​​dynamic optimization engine for fertilizer application is obtained by adding an AI dynamic optimization model for fertilizer application to the accumulated part of AI historical residuals for fertilizer application; The execution process of the AI ​​dynamic optimization engine for watering volume includes: calling the residual sequence of the most recent complete growth cycle in the historical residual accumulation part of the AI ​​watering volume, setting the negative definite convergence domain based on the linear quadratic Lyapunov function, and writing it into the residual accumulation register and the safety threshold register; real-time collection of soil status and crop root absorption capacity, together with the irrigation records of the previous period, writing them into the state observation buffer, and forming a state vector after unified scaling; the AI ​​dynamic optimization model for watering volume calls the linear matrix inequality solver to output the original watering control quantity that meets the Lyapunov descent constraint; at the same time, the adaptive gain modulator dynamically updates the matrix weight according to the convergence rate; the AI ​​historical residual accumulation part of watering volume incorporates the latest error into the residual accumulation register, generates the same-dimensional compensation quantity through the variable gain compensator, and buffers rapid disturbances; the original watering control quantity and the compensation quantity are added dimension by dimension, and the saturation protector is limited to the execution threshold range to obtain the target watering quantity, and writes it into the execution instruction queue; The process of calling the residual sequence of the most recent complete reproductive cycle in the accumulated part of the historical residual of the water application AI and setting the negative definite convergence region based on the linear quadratic Lyapunov function includes: Step A1: Perform integrity check on the residual sequence of the accumulated part of the AI ​​historical residual of water application, and linearly interpolate the missing segments in chronological order to obtain the verified sequence; Step A2: Write the verified sequence into the residual buffer stack and use a double-weight exponential decay filter to suppress high-frequency noise to obtain a filtered sequence; Step A3: Detect the extreme value of the filter sequence. If it exceeds the safety threshold register threshold, trigger the abnormal backtracking logic and return to step A1; Step B1: Dynamically generate a positive definite weighting matrix based on the variance distribution of the filter sequence and write it into the matrix register; Step B2: Construct a linear quadratic Lyapunov function from the positive definite weight matrix and the residual state vector according to the quadratic form, and verify the strict descent of all samples through the gradient iterator; Step B3: If the drop rate of any sample is non-negative, adaptively shrink the positive definite weight matrix and return to step B2 until all samples meet the conditions; Step C1: define the boundary of the negative definite convergence region with the maximum acceptable energy level of the linear quadratic Lyapunov function and write it into the negative definite convergence region register; Step C2: Synchronize the boundary parameters to the safety locking logic and the water application AI dynamic optimization engine to complete the setting of the negative definite convergence region.

2. The water-fertilizer integrated multi-objective optimization intelligent control system based on artificial intelligence drive according to claim 1 is characterized in that: The execution process of the AI ​​dynamic optimization engine for fertilizer application includes: calling out the residual sequence of the most recent complete growth cycle from the accumulated part of the AI ​​historical residuals for fertilizer application, writing it into the residual buffer stack, establishing the initial value of the sliding surface and recording the zero-bias baseline; collecting soil conditions and crop root absorption capacity, reading the fertilization records of the previous period, and writing them into the state observation buffer after unified scaling to form a state vector; the AI ​​dynamic optimization model for fertilizer application is based on the difference between the sliding surface and the state vector, and uses the Darwin switching law to generate the original fertilizer control quantity; the adaptive gain modulator in the same period updates the switching rate according to the approximation speed; the accumulated part of the AI ​​historical residuals for fertilizer application is superimposed on the nutritional deviation of this period, and is converted into an equal-dimensional compensation quantity by a variable gain compensator to suppress high-frequency disturbances; the original fertilizer control quantity and the compensation quantity are added dimension by dimension, limited to the safety threshold by the saturation protector, and the fertilizer quantity is obtained, and written into the execution instruction queue.

3. The water-fertilizer integrated multi-objective optimization intelligent control system based on artificial intelligence drive according to claim 2 is characterized in that: The AI-powered dynamic optimization model for fertilizer application rate uses the Darwin switching law to generate the original fertilizer control variable based on the difference between the sliding surface and the state vector. The process includes: establishing a sliding surface register in the AI-powered dynamic optimization model for fertilizer application rate, mapping the crop nutrient deviation and its time accumulation into a sliding surface vector; reading the state vector difference in real time, forming a multidimensional deviation tag through a difference comparator, and writing it into a deviation latch; sending the output of the deviation latch to a Darwin switching law generator to generate a symbol matrix; the Darwin switching law generator matches the symbol matrix dimension by dimension according to a pre-stored switching gain table to generate an initial switching pulse sequence; an adaptive gain modulator incrementally or decrements the switching gain based on the real-time threshold of the deviation amplitude and the accumulated portion of the AI-powered historical residual, and outputs a gain modulation matrix; multiplying the initial switching pulse sequence by the gain modulation matrix to generate an open-loop control sequence, which is then subjected to a two-stage shaping process of a rate limiter and a boundary layer filter to suppress high-frequency chattering; and multiplying the initial switching pulse sequence by the gain modulation matrix to generate an open-loop control sequence, which is then subjected to a two-stage shaping process of a rate limiter and a boundary layer filter to suppress high-frequency chattering.

4. The artificial intelligence-driven water-fertilizer integrated multi-objective optimization intelligent control system according to claim 3 is characterized in that: The system also includes: a water and fertilizer feedback control unit, which is used to measure the crop moisture status index, combine it with the water application amount, calculate the dynamic correction water application amount of the crop per unit area at each time, and perform closed-loop feedback regulation of crop water application; and measure the crop nutrient adequacy index, combine it with the fertilizer application amount, calculate the dynamic correction fertilizer application amount of the crop per unit area at each time, and perform closed-loop feedback regulation of crop fertilization.

5. The water-fertilizer integrated multi-objective optimization intelligent control system based on artificial intelligence drive according to claim 4 is characterized in that: The execution process of the water and fertilizer feedback control unit includes: reading the dynamic correction water application amount and dynamic correction fertilizer application amount stored in the previous operation cycle, writing into the water feedback register and the fertilizer feedback register, establishing a dual-channel error latch and clearing it; real-time collection of the measured crop water content index and the measured crop nutrient adequacy index, and writing them into the state observation buffer after suppressing high-frequency noise through a double-window index filter; calling the current water application amount and fertilizer application amount, forming a water error vector and a nutrient error vector in the error latch respectively, and recording the change rate mark; performing dual-loop proportional-integral modulation, including: the water error vector enters the water integral modulator to generate a water integral; the nutrient error vector enters the fertilizer integral modulator to generate a fertilizer integral Component; the two channels simultaneously perform proportional-integral superposition and output a first correction value; the first correction value is limited to the crop safety threshold by a limiter saturator, and a second correction value is generated and written into the dynamic correction register; the second correction value in the dynamic correction register is added to the original water application amount and the original fertilizer application amount one by one to obtain the dynamic correction water application amount and the dynamic correction fertilizer application amount of the crop per unit area at the current time, and pushed to the execution instruction queue at the same time; after irrigation and fertilization are completed, the measured crop water status index and the measured crop nutrient adequacy index are re-collected; if the absolute value of any error vector does not decrease for two consecutive cycles, the proportional-integral gain is increased and the dual-loop proportional-integral modulation is returned, otherwise the current gain is maintained to enter the next normal cycle, 6. The artificial intelligence-driven water-fertilizer integrated multi-objective optimization intelligent control system according to claim 5 is characterized in that: Meteorological data include: crop evapotranspiration , in mm / day; rainfall , in millimeters per day; the proportion of rainfall that is not absorbed by the soil and flows directly into the runoff is expressed as surface runoff rate To express; crop growth status data include: soil water potential , unit is kPa; crop root suction , unit is kPa; nitrogen demand of crops per unit area , in g / m² / day; the phosphorus requirement of crops per unit area , in g / m² / day; the potassium requirement of crops per unit area Unit is g / m² / day; soil adsorption rate ; Leaf chlorophyll concentration ; For time.

7. The artificial intelligence-driven water-fertilizer integrated multi-objective optimization intelligent control system according to claim 6 is characterized in that: The estimated water requirement of crops per unit area at each time is: ； in, for The water requirement of crops per unit area at each time; the fertilizer requirement of crops at each time is estimated to be: ； in, for The amount of fertilizer required by crops per unit area over time; for The theoretical chlorophyll concentration of the crop at the current growth stage.

Images