Intelligent irrigation monitoring method and system

Abstract

The invention relates to the technical field of intelligent gardens and precise irrigation, and particularly discloses an intelligent irrigation monitoring method and system. According to the method, a multi-source sensor array is deployed, Kalman filtering is adopted to fuse environmental data, and a three-dimensional state vector input reinforcement learning model is constructed to generate an irrigation decision; predicting a vegetation water demand by combining a gradient lifting decision tree, forming a graded irrigation strategy and converting the graded irrigation strategy into a water pump control instruction; soil humidity feedback data are collected in real time, decision model parameters and filtering rules are dynamically adjusted, and closed-loop optimization is achieved. Through multi-source data fusion and a self-adaptive decision-making mechanism, the irrigation precision and the water resource utilization rate are remarkably improved, meanwhile, the response capacity of the system to the vegetation growth dynamic state and the environment change is enhanced, and the beneficial effects of optimization process closed loop, decision dynamic adaptation and controllable resource consumption are achieved.

Description

Technical Field

[0001] The present invention relates to the field of smart gardening and precision irrigation technology, and in particular to an intelligent irrigation monitoring method and system. Background Art

[0002] Precision irrigation is the core of efficient water resource utilization in modern smart gardens. Soil moisture exhibits significant heterogeneity in spatial distribution and temporal evolution, influenced by the complex coupling of multiple physical factors such as meteorology, soil type, and crop root distribution. Traditional irrigation decision-making relies primarily on fixed threshold triggers, empirical rules, or predictive control based on a single soil moisture model, making it difficult to dynamically adapt to this complex spatiotemporal variation. In addition, the movement of irrigation water in the soil strictly follows the laws of fluid dynamics (as described by the Navier-Stokes equations). Existing irrigation decision-making methods based on simple rules or non-physical constraint data often fail to effectively couple to such physical constraints, resulting in decision-making results that deviate from the actual laws of water and fertilizer movement, leading to uneven irrigation, deep seepage, or localized water shortages.

[0003] While existing intelligent irrigation systems attempt to incorporate data-driven methods (such as reinforcement learning) for decision optimization, they face challenges in dealing with strongly nonlinear, multi-constrained systems. First, their constructed representations of environmental states fail to fully integrate the coupled effects of multiple physical fields (water, heat, solutes, etc.), limiting the completeness and accuracy of the state description. Second, the decision-making models lack explicit embedding and constraints on the Navier-Stokes physical laws inherent in the flow of irrigation water through soil pores. This can lead to learned strategies that violate fundamental principles of fluid motion and lack reliability. Furthermore, the moisture thresholds used to trigger irrigation decisions are typically static or empirically set, unable to be adaptively and dynamically adjusted in a continuous manifold space based on real-time changes in soil structure, crop water demand, and climatic conditions, resulting in a sluggish response to spatiotemporal heterogeneity.

[0004] In terms of energy efficiency optimization, existing methods mostly focus on a single goal (such as water conservation), or separate energy efficiency goals from moisture response goals. There is a lack of a systematic mechanism for unified modeling and coordinated optimization of irrigation system dynamics (such as pipeline hydraulics and pump station energy consumption) and the moisture response dynamics of the soil-crop-atmosphere system. It is difficult to achieve optimal energy efficiency of the overall operation of the irrigation system while ensuring the moisture needs of crops. Summary of the Invention

[0005] The present invention provides an intelligent irrigation monitoring method and system to solve the problem of how to achieve adaptive response to spatiotemporal heterogeneity of soil moisture and coordinated optimization of irrigation energy efficiency based on the construction of environmental state vectors coupled by multi-physical fields, reinforcement learning decision-making with Navier-Stokes constraints, dynamic threshold correction of Riemannian manifolds, and Hamiltonian optimal control mechanism.

[0006] In order to solve the above technical problems, the present invention provides an intelligent irrigation monitoring method, comprising: Deploy an array of capacitive humidity sensors, thermocouple temperature sensors, and photoelectric light sensors. GPS timing and a double-buffered queue are used to align multi-source data timing, and a Kalman filter is used to fuse soil moisture data, soil temperature data, and light intensity data. The capacitive humidity sensor probes are inserted into the soil profile in a vertical layered manner, converting moisture content into a voltage signal output by detecting changes in the dielectric constant. The thermocouple temperature sensor, which uses a composite structure of thermocouples and thermistors, is deployed in the critical temperature-variable layer from the soil surface to a depth of 40 centimeters, outputting millivolt-level electrical signals based on the resistance-temperature linear relationship. The photoelectric light sensor, equipped with a photodiode and a spectral filter module, is fixed one meter above the vegetation canopy to convert the solar radiation spectrum into standard lux or watts per square meter. The filtered data is used to construct a three-dimensional environmental state vector, which is then input into a Q-learning decision model to generate a four-dimensional action value evaluation vector. Based on the ε-greedy strategy, dynamic irrigation thresholds and water pump control instructions are output. The control instructions are encapsulated into hexadecimal data frames using the Modbus-RTU protocol. The frame structure includes five parts: a start character, a device address code, a function code, a data area, and a CRC checksum. The expression for outputting the dynamic irrigation threshold and the water pump control instruction includes: ; in, is the threshold boundary correction amount; is the soil porosity tensor; is the Christoffel symbol, describing the curvature of soil heterogeneity; is the spatial gradient of moisture content; is the time step; Optimal control of water pump energy: ; in, is the Hamiltonian function of the water pump system; is the water flow momentum; is the coordinate of the water flow position; For water quality; is the gravitational potential energy function; is the pipeline topology constraint function; is the constraint strength coefficient; is the norm square operator; By correlating dynamic irrigation thresholds with historical soil moisture data, the water requirements of vegetation are predicted using a gradient boosting decision tree, generating an irrigation scheduling strategy with time windows and graded water volumes. Convert water pump control instructions and irrigation scheduling strategies into TLV format instructions, use PWM to control the water pump's staged start and stop, and collect soil moisture feedback data after shutdown; Comparing soil moisture feedback data with dynamic irrigation thresholds to calculate regional deviations, adjust the weight coefficients of the Q-learning reward function, and reconstruct the four-dimensional state decision model; The vegetation health indicators are integrated with the optimized decision model parameters, the Kalman filter rules are updated through covariance analysis, and a training data set is generated and fed back to the sensor data fusion module.

[0007] Furthermore, constructing a three-dimensional environment state vector specifically includes: The soil moisture data is mapped to the moisture content coefficient in the interval [0,1], the soil temperature data is converted into the temperature change factor, and the light intensity data is derived into the light intensity index; The 3D environment state vector is used as the activation value of the input layer of the Q-learning decision model and is transmitted to the hidden layer configured with 128 ReLU activation function neurons through a fully connected neural network. The output layer sets four groups of action nodes: maintaining the current irrigation status, increasing irrigation intensity, reducing irrigation intensity, and emergency water replenishment, and generates a four-dimensional action value evaluation vector.

[0008] Furthermore, the process of outputting dynamic irrigation thresholds and water pump control instructions based on the ε-greedy strategy includes: The exploration threshold ε is initially set to 0.7 and decreases with a decay coefficient of 0.99; When the random number is higher than ε, the action corresponding to the maximum value of the action value evaluation vector is selected; when it is lower than ε, the action is randomly selected; Dynamic irrigation threshold calculation includes: lowering the lower limit of soil moisture by 3-5% when increasing irrigation intensity, raising the upper limit by 2-4% when reducing irrigation intensity, and setting a temporary threshold of 20% above the normal level for emergency water replenishment.

[0009] Furthermore, the intelligent irrigation monitoring method further includes: The execution frequency calculation module uses a sliding time window to analyze the number of irrigation triggers within 24 hours. When the frequency per unit time exceeds the warning value, the interval between adjacent actions is extended by 10%-15%; Output parameters include: dynamic threshold range of soil moisture content, response level parameters including three levels of normal / accelerated / inhibited, minimum interval time for water pump start and stop, and maximum number of operations per day.

[0010] Furthermore, the intelligent irrigation monitoring method further includes: The structured threshold control group construction includes: soil moisture lower threshold field, upper threshold field and four-digit binary response level identifier; the response mode activation strategy includes: when the last digit is 1, the 3-5 minute normal mode is enabled; when the last two digits are 10, the 6-8 minute accelerated mode is enabled; when the last two digits are 11, the 1-2 minute suppression mode is enabled.

[0011] Furthermore, the process of outputting dynamic irrigation thresholds and water pump control instructions based on the ε-greedy strategy also includes: The generation of water pump start and stop control instructions includes three levels of decision logic: The first level monitors that soil moisture remains below the lower threshold for 60 seconds; The minimum interval time for the second-level calibration water pump is greater than 30 minutes; The third level checks that the number of operations per day has not reached the maximum; When the conditions are met, a four-tuple instruction is generated, which includes the device address, opening time, water flow intensity level and emergency flag.

[0012] Furthermore, the intelligent irrigation monitoring method further includes: The control instructions are encapsulated as Modbus-RTU protocol data frames, including: Start character, device address code, function code, data area and CRC check code; The first byte of the data area stores the opening duration parameter (an integer value in 0.1 seconds), the lower 4 bits of the second byte record the water flow intensity level, and the upper 4 bits retain the emergency flag.

[0013] Furthermore, the process of inputting the Q-learning decision model to generate a four-dimensional action value evaluation vector includes: When the irrigation effect meets expectations, the corresponding action reward function increases the weight coefficient by 0.1; When over-irrigation or insufficient water replenishment occurs, the penalty function reduces the weight coefficient by 0.05; The fully connected weight matrix between the input layer and the hidden layer is updated through the temporal difference algorithm.

[0014] Furthermore, the intelligent irrigation monitoring method further includes: When predicting vegetation water demand, the gradient boosting decision tree integrates the dynamic irrigation threshold and the time series characteristics of historical soil moisture data to generate an irrigation scheduling strategy that includes time window parameters and graded water parameters. The time window parameters are accurate to 15-minute time periods, and the graded water parameters are divided into three flow levels according to vegetation type.

[0015] Furthermore, an intelligent irrigation monitoring system is applied to any of the above methods, comprising: Multi-source sensor array module, including capacitive humidity sensor, thermocouple temperature sensor and photoelectric light sensor; Reinforcement learning decision module, configured with Q-learning decision model and ε-greedy strategy selector; A water demand prediction module that stores historical soil moisture data and runs a gradient boosting decision tree; Hierarchical execution module, including TLV instruction converter and PWM water pump controller; Parameter optimization module, configures the area deviation calculator and reward function weight regulator; The rule iteration module includes a covariance analysis engine and a Kalman filter rule generator.

[0016] The key innovations of the present invention include: (1) Construct a multi-physics field coupled environmental state vector generation mechanism, integrate the water and heat conduction, root suction and micro-meteorological fluctuation characteristics of the soil-atmosphere-plant system in three-dimensional space, and form a high-dimensional spatiotemporal heterogeneity dynamic representation system.

[0017] (2) Design a reinforcement learning decision architecture based on the Navier-Stokes equation constraints, embed the fluid motion continuity equation into the strategy network gradient update process, and realize the essential coupling between irrigation action generation and the motion laws of the physical field.

[0018] (3) A dynamic threshold correction mechanism for Riemannian manifolds is established to calibrate the spatiotemporal correlation threshold of soil moisture sensor data in real time through manifold curvature analysis, thereby solving the threshold drift problem caused by traditional Euclidean space measurement.

[0019] The following are its main beneficial effects: (1) This invention breaks through the dependence of traditional irrigation models on static environmental parameters. By integrating the dynamic coupling data of multiple physical fields such as atmospheric turbulence, soil water potential gradient and plant transpiration, it constructs an environmental state vector with spatiotemporal evolution characteristics, significantly improving the perception accuracy of water migration characteristics in complex farmland environments. In particular, in highly heterogeneous areas such as slopes and sandy loam, the model has stronger adaptability to working conditions than traditional homogenized models.

[0020] (2) The proposed reinforcement learning decision architecture transforms Navier-Stokes fluid dynamics constraints into physical regularization terms of the policy network, so that the irrigation strategy automatically satisfies the principles of mass conservation and momentum transfer, fundamentally avoiding the common physical law conflicts in traditional artificial intelligence decision-making, and greatly improving the feasibility and stability of irrigation actions in real fluid environments.

[0021] (3) The threshold correction mechanism based on Riemannian manifolds realizes curvature-driven dynamic calibration of soil moisture anomalies by establishing a differential geometry correlation model of sensor data, effectively suppressing false triggering of irrigation caused by sensor drift and spatial interpolation errors, and reducing the risk of redundant water consumption in the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flow chart of an intelligent irrigation monitoring method provided in an embodiment of the present application; Figure 2 This is a structural block diagram of an intelligent irrigation monitoring system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0023] Example 1: Reference Figure 1 , is a flow chart of an intelligent irrigation monitoring method provided by an embodiment of the present invention. The flow chart may include at least steps S100-S600: S100 deploys capacitive humidity sensors, thermocouple temperature sensors, and photoelectric light sensor arrays, aligns multi-source data timing through GPS timing and double-buffered queues, and uses Kalman filtering to fuse soil moisture data, soil temperature data, and light intensity data.

[0024] S200: Construct a three-dimensional environment state vector from the filtered data, input it into the Q-learning decision model to generate a four-dimensional action value evaluation vector, and output a dynamic irrigation threshold and water pump control instruction based on the ε-greedy strategy.

[0025] S300, associating dynamic irrigation thresholds with historical soil moisture data, predicting vegetation water requirements through a gradient boosting decision tree, and generating an irrigation scheduling strategy with time windows and graded water volumes.

[0026] S400: Convert the water pump control instructions and irrigation scheduling strategy into TLV format instructions, use PWM to control the water pump to start and stop in stages, and collect soil moisture feedback data after stopping.

[0027] S500: Compare the soil moisture feedback data with the dynamic irrigation threshold to calculate the regional deviation, adjust the weight coefficient of the Q-learning reward function, and reconstruct the four-dimensional state decision model.

[0028] S600, integrates vegetation health indicators with optimized decision model parameters, updates Kalman filter rules through covariance analysis, generates training data sets and feeds them back to the sensor data fusion module.

[0029] Step S100 at least includes steps S110-S130: S110: Obtain original monitoring data from the soil moisture sensor, temperature sensor, and light sensor, and perform data synchronization and alignment.

[0030] In the construction of smart gardens, a distributed multi-source sensor array is deployed within the target vegetation area. This array consists of three types of IoT sensors: Soil moisture sensors utilize capacitive or frequency domain reflectometry technology. Their probes are inserted vertically into the soil profile in layers, detecting changes in the dielectric constant and converting moisture content into a voltage output signal. Temperature sensors utilize a composite structure of thermocouples and thermistors and are deployed in the critical temperature-variable layer from the soil surface to a depth of 40 centimeters. They output millivolt-level electrical signals based on the resistance-temperature linear relationship. Light sensors, equipped with photodiodes and spectral filtering modules, are fixed one meter above the vegetation canopy and convert the solar radiation spectrum into standard lux or watts per square meter. All sensor nodes establish a star topology connection to a central data acquisition unit via the LoRa wireless networking protocol, triggering a network-wide polling and collection cycle every five minutes.

[0031] During data collection, the soil moisture sensor outputs an analog voltage signal representing moisture content between 0% and 100%, the temperature sensor outputs an electrical impedance value corresponding to the Celsius temperature scale, and the light sensor outputs a digital quantization of light intensity. Due to electromagnetic interference, multipath effects, and fluctuations in node energy consumption, raw data packets experience millisecond-level timing drift and packet loss during transmission. To address this issue, the central data acquisition unit integrates a GPS timing module, attaching a microsecond-level Coordinated Universal Time (UTC) timestamp to each arriving data packet and periodically calibrating the internal clock source using the Network Time Protocol. The timing alignment process utilizes a double-buffered queue architecture: raw data packets are first stored in the input buffer. When all three sensor data packets from the same acquisition cycle arrive (with a time window tolerance of ±2 seconds), the timestamp comparison engine is activated, discarding timed-out packets and triggering the sensor retransmission mechanism. Valid data packets are then reordered and transferred to the output buffer. A cyclic redundancy check is also performed, generating retransmission instructions for packets that fail the check. The final output structured data frame contains the synchronized original values ​​of soil moisture, soil temperature, and light intensity. The data frame is written into the cache area as the input source of step S120 to ensure spatiotemporal consistency and meet the requirements of multi-source fusion.

[0032] S120: Perform Kalman filter fusion processing on the original monitoring data to reduce environmental noise interference.

[0033] After receiving the synchronized data frame from step S110, the Kalman filter initializes a three-parameter state space: the soil moisture state variable is linked to a physical model of water diffusion; the soil temperature state variable is constructed based on the heat conduction equation; and the light intensity state variable is coupled to the atmospheric attenuation coefficient. The state transition matrix is ​​derived based on the law of thermal inertia and the principle of water conservation. The predicted soil moisture value is corrected by adding the temperature gradient change to the previous moment's moisture value. The predicted temperature value calculates the energy transfer trend using a heat capacitance-thermal resistance model. The measurement noise covariance matrix is ​​dynamically configured based on the sensor calibration accuracy: the soil moisture sensor noise variance is set to 0.5%, the temperature sensor noise variance is limited to 0.2 degrees Celsius, and the light sensor noise variance is controlled at 5%.

[0034] The processing process is implemented through a dual prediction-update loop. The prediction phase uses state transition equations to generate a priori estimates of soil parameters at the current moment, and simultaneously calculates the error covariance matrix to quantify prediction uncertainty. The update phase incorporates observations input from S110 for correction. The core mechanism is dynamic weighting of the Kalman gain: when the deviation between the observed and predicted values ​​exceeds the sensor noise threshold, the weight of the observed data is automatically increased to suppress interference such as temperature drift and circuit thermal noise. Confidence-weighted fusion is implemented to account for the characteristics of multi-source data. For example, capacitive humidity sensors are given a higher weight than thermistor temperature sensors. The multi-sensor data association algorithm measures correlation using the Mahalanobis distance and performs a covariance cross-check on soil moisture and temperature observations to eliminate data inconsistencies caused by localized uneven sunlight. Noise suppression utilizes a fifth-order iterative smoothing filter to effectively filter out transient interference, such as light intensity jumps caused by cloud cover. The final output optimized data array contains three fusion parameters: a fused soil moisture value eliminates the influence of dielectric constant fluctuations in the capacitive sensor; a fused soil temperature value integrates surface-to-depth temperature differences; and a fused light intensity value preserves effective spectral characteristics. The array is stored in the shared memory area for calling by the S130 step, and the measurement accuracy is improved by more than 30%.

[0035] S130: Output the noise-filtered soil moisture data, soil temperature data, and light intensity data.

[0036] After inputting the fused data array generated by the S120, the central processing unit performs a standardized output process. Soil moisture data is linearly scaled to map the fused values ​​to a standard range of zero to 100 percent, and a data quality tag is embedded, including the GPS timestamp and sensor node number. Soil temperature data is converted to Celsius and filtered using a dual-threshold mechanism: values ​​outside the valid range of -10°C to +50°C are marked as invalid, and cubic spline interpolation is performed between adjacent data points to compensate. Light intensity data is uniformly converted to lux, and a low-pass filter is applied to remove up to 5 percent of residual high-frequency noise.

[0037] All parameters are encapsulated into structured data packets using a format conversion engine and stored using JSON key-value pairs: soil moisture data is written to the "moisture" field with one decimal point precision, temperature data is written to the "temperature" field with a sign, and light data is written to the "illuminance" field as an integer. The data packets are transferred to non-volatile memory via a serial peripheral interface bus, using a round-robin storage strategy to retain the last 24 hours of data, ensuring zero data loss during system power outages. The output ready state is activated through an event-triggered mechanism: once the data packet completes digital signature verification, the system sets the "data ready" flag and releases an interrupt signal, allowing the reinforcement learning decision module in step S200 to access the data packet in real time via the application programming interface. The final output data set contains three key parameters: noise-filtered soil moisture data representing the baseline amount of irrigation water available, soil temperature data reflecting the thermodynamic state of root activity, and light intensity data quantifying photosynthetically active radiant energy. This data packet serves as the core input to the closed-loop irrigation decision chain, directly driving the Q-learning strategy optimization engine in substep S210.

[0038] Step S200 at least includes steps S210-S230: S210 , inputting the soil moisture data, soil temperature data, and light intensity data into a Q-learning decision model.

[0039] The Kalman filter-processed soil moisture, soil temperature, and light intensity data output from step S130 are received. Soil moisture data is quantified as a percentage of volumetric moisture content, soil temperature data is recorded in degrees Celsius, and light intensity data is collected in lux. The three data types are normalized, mapping the raw soil moisture values ​​to a moisture coefficient in the [0, 1] range, converting the raw soil temperature values ​​to a temperature variation factor, and deriving the raw light intensity values ​​into a light intensity index. Together, these data construct a three-dimensional environmental state vector. This environmental state vector serves as the activation value of the input layer of the Q-learning decision model and is transmitted to the hidden layer via a fully connected neural network. The hidden layer is configured with 128 neurons, each of which uses a Reluctant Unit (ReLU) activation function to perform a nonlinear transformation on the input value. A fully connected weight matrix is ​​established between the input and hidden layers. This matrix is ​​assigned values ​​from a random normal distribution during model initialization and is updated using a temporal difference algorithm after each decision iteration. The output layer contains four sets of action nodes: a node for maintaining the current irrigation state, a node for increasing irrigation intensity, a node for decreasing irrigation intensity, and a node for emergency water replenishment. The output value of each action node represents the estimated cumulative reward for executing the corresponding action under specific environmental conditions. After forward propagation of the environmental state vector from the input layer to the hidden layer, a four-dimensional action value vector is generated at the output layer. This vector contains the utility score for each irrigation action. The state data set output from this step fully encodes the current environmental characteristics and also generates an action value vector that serves as the basis for irrigation decisions.

[0040] S220: Predicting the irrigation threshold and calculating the execution frequency using the Q-learning decision model.

[0041] Based on the four-dimensional action value evaluation vector generated by S210, an ε-greedy strategy is used to select the optimal irrigation action. When the random number generator generates a value above the exploration threshold ε, the action node corresponding to the maximum value in the action value evaluation vector is selected as the execution instruction. When the random number generator generates a value below the exploration threshold ε, the execution instruction is randomly selected from the four action node categories. The exploration threshold ε is initially set to 0.7 and decreases with the number of decision iterations by a decay coefficient of 0.99. Dynamic irrigation thresholds are calculated based on the selected action type: when the maintain state instruction is selected, the threshold parameters of the previous cycle are inherited; when the increase irrigation intensity instruction is selected, the lower soil moisture threshold is lowered by 3 to 5 percentage points; when the decrease irrigation intensity instruction is selected, the upper soil moisture threshold is raised by 2 to 4 percentage points; and when the emergency water replenishment instruction is selected, a temporary irrigation threshold is set that exceeds the normal range by 20%. The execution frequency calculation module receives the action selection results and historical operation records and analyzes the number of irrigation action triggers within the last 24 hours using a sliding time window. If the trigger frequency per unit time exceeds a preset warning value, the interval between consecutive action executions is automatically extended by 10 to 15%. Irrigation threshold prediction outputs two types of parameters: a dynamic soil moisture threshold interval consisting of lower and upper critical values, and an irrigation response level parameter consisting of three levels: normal response mode, accelerated response mode, and suppressed response mode. The execution frequency calculation generates parameters for the minimum interval between pump starts and stops and the maximum number of operations per day. The Q-learning model weight matrix is ​​updated based on the actual irrigation results: when the irrigation effect meets expectations, the reward function for the corresponding action increases by 0.1; when over-irrigation or insufficient water replenishment occurs, the penalty function decreases by 0.05. The moisture content floating range and response level code output in this step constitute the dynamic irrigation threshold, while the time interval threshold and frequency threshold form the execution frequency control parameters.

[0042] S230: Generate dynamic irrigation threshold parameters and water pump start and stop control instructions.

[0043] The dynamic threshold ranges and response level parameters output by the S220 are integrated to construct a structured threshold control group. This control group contains three core fields: a field for recording the lower soil moisture threshold value (5% to 25%), a field for recording the upper soil moisture threshold value (15% to 40%), and a field for storing a four-bit binary-coded response level identifier. The corresponding control strategy is activated based on the identifier: when the last digit is 1, the normal response mode is enabled, setting the pump's single-time on-time to 3 to 5 minutes; when the last two digits are 10, the accelerated response mode is enabled, extending the on-time to 6 to 8 minutes; and when the last two digits are 11, the suppressed response mode is enabled, compressing the on-time to 1 to 2 minutes. The pump start / stop control command generation module receives the threshold control group and execution frequency parameters and executes a three-level decision logic through a state machine. The first level monitors real-time soil moisture data and triggers the start condition when the soil moisture content remains below the lower threshold for 60 seconds. The second level verifies the minimum interval time parameter for the pump to ensure that the interval between consecutive operations is greater than 30 minutes. The third level checks the daily operation counter and inserts a mandatory delay when the maximum value is reached. Once the three conditions are met, a four-tuple control instruction is generated: instruction element 1 specifies the physical address code of the pump device, instruction element 2 defines the on-time duration (with 0.1-second precision), instruction element 3 sets the water flow intensity level from 1 to 3, and instruction element 4 indicates the emergency operation flag. The control instruction is encapsulated as a hexadecimal data frame using the Modbus-RTU protocol. The frame structure consists of five parts: a start character, a device address code, a function code, a data area, and a CRC checksum. The first byte of the data area stores the on-time duration parameter (converted to an integer value in 0.1-second units), the lower four bits of the second byte record the water flow intensity level, and the upper four bits reserve the emergency operation flag. The final output includes a structured parameter set with a floating threshold range and a pump control instruction data packet that complies with industrial control standards.

[0044] In another embodiment: S210: Fluid dynamics modeling of the environment state vector and Q-learning input reconstruction.

[0045] Receive Kalman filter data output by S130: soil moisture , soil temperature , light intensity The three-dimensional environment state vector is established through multi-physics field coupling normalization: ① Calculation of moisture content coefficient: enter After linear normalization: ; in, : Lower limit of effective soil moisture content; : Upper limit of soil saturation moisture content; is the normalized moisture content coefficient, a dimensionless moisture content index (output to the state vector); ② Temperature change factor modeling: enter and its space-time differentials: ; in, : Dynamic temperature change intensity (output to the state vector); : soil temperature time change rate; : soil temperature spatial Laplace operator; : soil thermal diffusivity; ③Light intensity index conversion: enter Corrected for canopy attenuation: ; in, : top of atmosphere illumination reference value; : average height of crop canopy; : Logarithmic index of canopy transmittance (output to state vector); Environment state vector construction: ; in, : Environmental state vector; serves as the input of the Q-learning decision model to achieve multi-source data fusion and dimensional unification.

[0046] Neural Network Modeling of Physical Constraints: ④Navier-Stokes embedded state transfer: ; in: : Nabla operator; : pressure gradient term; : viscous force term; : water density (constraint weight initialization); : hydrodynamic viscosity (controls gradient propagation); : Water penetration rate vector (mapping neuron connection weights); : soil water pressure field (corresponding to bias term correction); : gravitational acceleration vector; This equation is added to the loss function as a physical regularization term to ensure that the decision complies with the laws of fluid dynamics.

[0047] Improved activation function design: The 128 neurons in the hidden layer use compound activation: ; in: It is an improved Gaussian decay activation function; is the rectified linear unit activation function; : weighted sum of neuron inputs; : Gaussian attenuation coefficient (suppresses abnormal pulses); ⑤Q-value update mechanism of ecological constraints: ; in: State-action value function; : learning rate (controls the model update amplitude); : Future earnings discount factor; : Vegetation health weight ( corresponding to different actions); : Soil improvement benefit function after irrigation; is the environmental state vector at time t; The irrigation action selected at time t; is the action space minimum operator; Output four-dimensional action value evaluation vector To S220.

[0048] Technical Relevance: During the S210 process: Input: S130 , , ; Vegetation health weight After irrigation ; ① Moisture content coefficient: receiving , combined with the preset / , output To the state vector; ② Temperature change factor: receiving and its space-time differentials, combined with Output To the state vector; ③Light intensity index: receiving , combined with and Output To the state vector; Environment state vector: Fusion , , Constructing a 3D vector → Enter the Q-learning network; ⑤Q value update: receive S630 and soil improvement benefits , output Vector → to S220 decision module; Technical Relevance: The environmental state vector of S210 directly supports the decision input of S220, where Participate in the threshold correction calculation in S220.

[0049] S220: Dynamic threshold prediction and execution frequency optimization based on Riemannian manifold.

[0050] Action selection strategy: ; in: : initial exploration rate (decay coefficient 0.99 per step); : Uniformly distributed random numbers; : Decision action (input from S210 vector); ⑥Non-uniform soil moisture threshold correction: Input S210 and soil structure parameters: ; in: : Soil porosity tensor ( corresponding to three-dimensional coordinates); : Christoffel symbol (describing soil heterogeneity curvature); : spatial gradient of moisture content; : Threshold boundary correction; is the time step; Dynamic threshold update rules:

[0051] ⑦ Execution frequency random optimization: Enter the number of historical operations and the time series: ; in: : Quadratic potential energy function ( is the operating frequency); : diffusion coefficient (controls the intensity of random fluctuations); : Wiener process increment (simulating environmental disturbance); When within 24 hours (Default ) triggers adjustment: ; in: is the optimized new operation time interval; is the current operation time interval; is the critical operation number threshold; Output parameters to S230: : dynamic threshold interval; is the lower limit of the dynamic threshold; : Dynamic threshold upper limit; : 2-bit response level (high bit indicates action intensity, low bit indicates urgency); : Minimum operation interval (in minutes); : Maximum number of operations per day (adaptive according to season); Technical Relationship: During S200, input: S210 Vector Sum ; Distributed sensor ; Historical operation data; Action selection: Based on S210 Vector, output decision action ;⑥Threshold correction: Receive S210 and soil structure parameters, output →Update / ;⑦Frequency optimization: receiving historical operation times , output and ; Output parameters: =[ , ]→To S230 control group; Response level → to S230 control group; Minimum interval → to S230 state machine; Daily limit → go to S230 state machine; Technical Relevance: Dynamic Threshold of S220 Output and response level Direct drive S230's Hamiltonian control system.

[0052] S230: Instruction generation and protocol encapsulation of Hamiltonian control system.

[0053] Input control group construction: ; in: is the input control group of the Hamiltonian control system; From S220 , From S220 response level; ⑧Optimal control of water pump energy: ; in: is the Hamiltonian function of the water pump system; : Water flow momentum ( ); : Water flow position coordinates; :Water quality( is the pipe volume); : gravitational potential energy function; : pipeline topology constraint function; : constraint strength coefficient; is the norm square operator; Response strategy implementation:

[0054] Technical Relevance: During the S230 process: Input: S220 and ;real time ;S220 / ; Control group construction: =⟨ , , > (all from S220); ⑧Water pump control: based on Value determination Scope and energy efficiency targets State machine verification: Trigger conditions: < (From S220 ); Interval calibration: > (from S220); Number of times check: < (from S220); Instruction package: ← Lowest bit (direct mapping); according to The value range is determined; Output =⟨ , , , > in: : opening duration (s); : velocity level {1,2,3}; : device address (such as 0xD2F3); : Emergency sign; : Control instruction quadruple; Technology closed loop: The instructions generated by S230 actually change the soil state, new Feedback is sent to S130 to start the next round of decision-making cycle.

[0055] S210 technical effect: Multi-physics field fusion: Water content ( ),temperature( ),illumination( ) is unified into a dimensionless state vector; physical constraint embedding: Navier-Stokes equations ensure that decisions conform to the laws of fluid dynamics; ecological optimization: vegetation weight Make Q-learning take into account crop health needs; S220 Technical Effect: Dynamic Threshold Adaptation: Riemannian Manifold Correction Responding to soil spatial heterogeneity; Intelligent frequency control: Langevin equation optimizes operating frequency (>8 times in 24 hours) Increased by 34%); Response level: 2 bits Values ​​encode action intensity and urgency; S230 technical effect: Energy optimization: Hamiltonian system according to different The three-level fault-tolerance mechanism: triple verification of threshold value / interval / number of times reduces the false trigger rate to <0.1%; industrial-grade protocol: improved Modbus frame supports 0.1s precision control.

[0056] Step S300 at least includes steps S310-S330: S310: Obtain historical soil moisture data and historical irrigation records during the vegetation growth cycle.

[0057] In the intelligent irrigation monitoring system, the historical data acquisition module first accesses the cloud database and calls the historical monitoring data set within the complete growth cycle of the specified vegetation. This data set contains a historical soil moisture data matrix indexed by precise timestamps, whose dimensions are expressed as the number of monitoring points multiplied by the number of historical time points. At the same time, a historical irrigation record table is associated and stored, which records the start timestamp, duration, power level of the water pump used, and actual water consumption parameters of each irrigation operation. The historical soil moisture data is directly derived from the noise-filtered soil moisture data output by step S130 and processed by Kalman filtering. This data is supported by a time series database formed by a long-term storage mechanism. Its original acquisition frequency is fixed to a single sampling every fifteen minutes. A single data record clearly contains a unique sensor number, corresponding geographic coordinate information, a specific soil layer depth identifier, and a calibrated moisture percentage value.

[0058] The system preprocesses the raw historical dataset using a time window partitioning unit. This preprocessing process divides the data into independent blocks based on natural calendar days, and rigorous data integrity verification is performed on each block. If the percentage of missing historical soil moisture data within a single calendar day exceeds a preset threshold of 5%, a data interpolation mechanism for adjacent blocks is automatically activated, reconstructing the missing data points using a time series autoregressive algorithm. A synchronous outlier cleaning module calculates the internal standard deviation of a sliding window with a fixed window size of 24 hours, removing anomalous data points that deviate from the window mean by a factor of three. Ultimately, the system outputs a structured historical dataset categorized by physiological growth stages of vegetation, including budding, rapid growth, and maturity. Each data category contains a complete time-series soil moisture curve for that stage, along with corresponding irrigation operation logs. The output interface of this structured historical dataset establishes a direct physical connection to the input port of the machine learning model in step S320, ensuring smooth data transmission.

[0059] S320. Predicting vegetation water demand based on machine learning regression analysis and associating the dynamic irrigation threshold parameter.

[0060] After receiving the structured historical dataset from step S310, the core component of the water demand forecasting engine initiates the multivariate regression analysis model. During the model initialization phase, the feature engineering processing unit first performs the task of constructing the input feature vector. This task extracts key statistical features from the historical soil moisture time series data, specifically including the average soil moisture value over the last 72 hours, the rate of change of the soil moisture decline slope over the past 24 hours, and the maximum amplitude of soil moisture fluctuations during the day and night within a complete natural day. Simultaneously, the feature engineering processing unit extracts important related features from the associated historical irrigation record table, including the time interval between two adjacent independent irrigation events and the cumulative total water consumption of several irrigations prior to the current irrigation.

[0061] The target variable generation unit simultaneously calculates the theoretical water requirement of vegetation. This calculation is based on the potential evapotranspiration reference value derived from the Penman formula, combined with the specific crop coefficient corresponding to the target vegetation type, to generate an accurate baseline value for the theoretical water requirement of vegetation on a daily basis. The regression analysis model uses a gradient boosting decision tree as its basic architecture, which contains three interconnected implicit node processing layers. The first-layer processing unit receives the input feature vector generated by the feature engineering processing unit and simultaneously receives the dynamic irrigation threshold parameter set output from step S230. This parameter set includes the soil moisture trigger threshold optimized by the reinforcement learning model, the critical soil temperature threshold for irrigation operation, and the light intensity limit threshold. The second-layer processing unit performs a feature cross-fusion operation. Its core operation is to perform a weighted fusion process on the dynamic irrigation threshold parameters provided by step S230 and the historical soil moisture decline slope extracted above. The weight factors used in the fusion process are dynamically determined by the adaptive learning algorithm during the model training phase, ultimately generating a composite feature factor that represents the correlation between environmental status and historical trends. The third-layer processing unit is responsible for establishing a nonlinear mapping relationship between this composite characteristic factor and the target variable, namely the baseline value of the theoretical water demand of vegetation. By minimizing the mean square error function between the model prediction value and the actual value, it iteratively optimizes key parameters such as the split node parameters, leaf node weight values, and tree structure complexity within the decision tree.

[0062] After the model operation is completed, the strategy coordinator module associated with its output interface performs a collaborative verification function of the prediction results. This module performs a logical matching verification on the predicted water demand value output by the model and the dynamic irrigation threshold parameter generated in step S230. When the verification finds that the expected value of the soil moisture correction calculated based on the predicted water demand is lower than the currently set soil moisture trigger threshold, the threshold recalibration mechanism is automatically activated. This mechanism uses a preset default step size of 0.5 percent to gradually reduce the soil moisture trigger threshold until the corrected threshold can meet the irrigation trigger condition based on the predicted water demand. Finally, the system outputs a decision parameter group containing the optimized theoretical daily water demand value of vegetation and the recalibrated soil moisture trigger threshold. This decision parameter group is fully transmitted to the irrigation scheduling strategy generator in step S330 as the core input via the system's internal data bus.

[0063] S330: Generate an irrigation scheduling strategy including an irrigation time window and a water allocation plan.

[0064] After receiving the decision parameter group from step S320, the irrigation scheduling strategy generator generates an executable irrigation scheduling plan according to a specific logical sequence. First, the time window calculation unit determines the precise irrigation triggering time based on the optimized recalibrated soil moisture trigger threshold. The unit monitors the current soil moisture data stream output by step S130 in real time, and immediately activates the irrigation operation window when the soil moisture value is continuously monitored to be lower than the recalibration threshold. At the same time, the unit introduces the real-time light intensity data output by step S130 for time constraint control, strictly limiting the irrigation operation to start and execute only during the period when the light intensity is lower than the light intensity limit threshold set in step S230. Such period usually corresponds to the time window with weak light at night or in the early morning, so as to avoid the ineffective evaporation loss of water caused by strong sunlight conditions to the greatest extent.

[0065] The water allocation calculation unit accurately calculates the amount of water to be allocated for a single irrigation operation based on the theoretical daily water requirement of the vegetation predicted in step S320. This calculation utilizes a hierarchical progressive algorithm, scientifically dividing the total water requirement into two parts: a baseline water volume, which meets the minimum water requirement for basic plant survival, and a regulating water volume, which fine-tunes the water volume based on real-time environmental changes. The baseline water volume is calculated based on the effective root absorption depth of the target plant, quantified in millimeters of water depth. The regulating water volume is dynamically linked to the current soil temperature output in step S130, resulting in a proportional increase in the regulating water volume with increasing soil temperature. Specifically, for every degree Celsius increase in soil temperature, the regulating water volume increases by 0.8 percent above the baseline water volume.

[0066] The execution parameter conversion unit is responsible for converting the water allocation plan into control parameters that can directly drive the water pump equipment. This unit calls a pre-stored water pump flow characteristic model, which internally stores a mapping relationship table of stable water output flow corresponding to different power level settings of the water pump. Based on this mapping relationship, the conversion unit intelligently decomposes the total water demand into multiple irrigation subtasks that can be executed sequentially. For example, when the calculated total water demand is 50 liters, the unit may disassemble it into two independent 25-liter irrigation subtasks. The specific execution length of each subtask is automatically calculated based on the power level of the current water pump configuration by querying the flow mapping table and executing the time conversion formula.

[0067] The resulting irrigation scheduling strategy contains clearly structured data fields: a minute-accurate irrigation execution time window field, which defines the specific start and stop times; a water allocation sequence field, broken down by subtask execution, clearly listing the planned water supply for each subtask; and an associated control parameter set field, which includes the specified pump power level, the maximum allowed continuous run time for a single subtask, and the maximum number of retries in the event of device communication or execution failures. The entire irrigation scheduling strategy is encapsulated in a standardized JSON data format and written to a persistent strategy database for storage and management. This data is then simultaneously triggered to issue irrigation control instructions in step S410, initiating the actual irrigation process.

[0068] Step S400 at least includes steps S410-S430: S410: Send the water pump start / stop control instruction and irrigation scheduling strategy to the water pump execution module.

[0069] The system's central control unit receives the pump start / stop control instructions from step S230 and the irrigation scheduling strategy from step S330. The pump start / stop control instructions include three core parameters: a start threshold voltage signal (range: 3.0-5.0V), a stop threshold voltage signal (range: 4.2-5.5V), and a maximum continuous run duration parameter (in seconds), generated based on a Q-learning decision model. The irrigation scheduling strategy includes the time window start time (accurate to the minute), the window duration (in minutes), the target irrigation water allocation value (in liters), and the regional priority coefficient (range: 0.5-1.0).

[0070] The data encapsulation module first binary encodes the voltage threshold parameter with 0.1V accuracy and maps it to the execution module's ADC input range (0-10V range). The time parameter conversion unit converts the start time in the scheduling policy into a 32-bit Unix timestamp and the duration into a 16-bit integer in seconds. The water allocation value is converted to the corresponding PWM duty cycle value (range 15%-100%) using the preset pump flow characteristic curve (stored in the central control unit's FLASH memory), and the priority coefficient is converted into a 4-bit binary code. The encapsulated data packet uses a TLV (Type-Length-Value) structure: the type field identifies the parameter category (0xA1 for start / stop commands, 0xB2 for scheduling policy), the length field specifies the data length, and the value field stores the specific parameter value.

[0071] The central control unit transmits the encapsulated data packets to the target water pump actuator module via the RS-485 physical layer communication interface (baud rate 19200 bps, parity enabled). After receiving the data packets via the UART interface, the actuator module's built-in STM32F407 microcontroller initiates the data parsing engine: it verifies the CRC16 redundancy code to confirm data integrity, extracts parameter values ​​according to the TLV structure, stores the voltage threshold in the ADC threshold register, writes the time parameters to the RTC real-time clock module, loads the PWM duty cycle value into the timer compare register, and stores the priority coefficient in the EEPROM non-volatile memory. After configuration is complete, the microcontroller returns an ACK frame, completing the closed-loop communication verification.

[0072] S420. Control the water pump to perform graded irrigation operations according to the scheduling strategy.

[0073] The microcontroller of the water pump execution module collects three input signals in real time: the first one obtains the raw voltage (0-3.3V) of the local soil moisture sensor through a 12-bit ADC channel. This sensor is the same model as the sensor array deployed in step S110 (FDR frequency domain reflectometry, measurement accuracy ±2%); the second one reads the internal RTC clock (error ±1ppm); and the third one monitors the water pump operating status feedback signal (overcurrent / overheating flag).

[0074] The control logic unit performs a three-step process: irrigation is triggered when the soil moisture value converted by the ADC (linearly converted to volumetric water content) is below the start threshold specified in the start / stop command, the RTC time is within the time window defined by the irrigation scheduling strategy (with a ±2-minute tolerance), and no fault flag is activated. The microcontroller outputs a high-level signal to an optocoupler (model PC817) via a GPIO, turning on a MOSFET power transistor (model IRF540N), powering the water pump motor (24V DC / 350W).

[0075] Immediately after startup, the system enters the graded flow control phase: the target PWM duty cycle corresponding to the current time window is read from non-volatile memory, and a 16kHz PWM waveform is output through the TIM1 timer channel. The duty cycle is adjustable with an accuracy of 1%, corresponding to the pump speed levels (25% duty cycle = 1200 rpm, 50% duty cycle = 2400 rpm, 75% duty cycle = 3600 rpm). Triple stop monitoring is also activated: a built-in watchdog timer accumulates run time and forces a shutdown when the maximum continuous run threshold is reached; an ADC collects soil moisture every 10 seconds, terminating the system immediately if the stop threshold is exceeded; and an RTC automatically shuts down the pump if the time window is exceeded.

[0076] During operation, the microcontroller continuously records 10 status parameters: actual start time (RTC timestamp), stop time, real-time PWM duty cycle, motor drive voltage (collected by ADC), current (collected by ACS712 sensor), cumulative power consumption (calculated by voltage and current integration), fault codes, soil moisture profile (recorded once per minute), ambient temperature (collected by DS18B20), and vibration amplitude (collected by ADXL345 triaxial accelerometer). This data is stored in an external FRAM memory (256KB capacity) using a ring buffer structure.

[0077] S430: Real-time collection of soil moisture feedback data and water pump operation status data after irrigation.

[0078] After the pump shuts down, the execution module initiates a delay trigger mechanism: a built-in timer begins counting down (default 15 minutes, configurable from 5 to 30 minutes). When the countdown ends, the microcontroller sends a measurement command to the soil moisture sensor (model SEN0193) via the I²C bus. The sensor transmits a 100MHz electromagnetic wave and receives the reflected signal. The internal ASIC chip calculates the dielectric constant and outputs a 0-3.3V analog value. Simultaneously, the pump's operating status data stored in FRAM is read.

[0079] Data preprocessing involves three key operations: temperature compensation (compensation coefficient -0.3% / °C) of the raw soil moisture voltage values, 50Hz power frequency filtering of the water pump current values, and mean filtering of the vibration data. The preprocessed data set is uploaded to the central control unit via a LoRa wireless module (433MHz frequency band, 20dBm transmit power). The transmission frame contains the device ID, data acquisition timestamp, and a 16-byte checksum.

[0080] After receiving the data, the central control unit activates the Kalman filter fusion engine, processing the current soil moisture value using the same state-space model (state vector [soil moisture, temperature, light]) as in step S120. The filter parameters are identical to those in S120: process noise covariance Q = diag(0.02, 0.01, 0.05), and observation noise covariance R = diag(0.002, 0.005, 0.03). The fusion process uses the previous filter result as a prior estimate, and the optimized value is output through a prediction-update loop. The resulting soil moisture feedback data has an improved accuracy of ±0.8%. It is written to a MySQL database along with time-aligned pump operating status data. The storage structure consists of 12 fields: timestamp, device coordinates, soil moisture value, temperature value, light value, pump start / stop status, PWM setpoint, voltage, current, vibration amplitude, fault code, and data quality flag.

[0081] Step 500 at least includes steps S510-S530: S510: performing a difference comparison between the soil moisture feedback data and the dynamic irrigation threshold parameter.

[0082] After receiving the soil moisture feedback data packet and pump operating status data set transmitted in step S430, the system initiates the real-time difference comparison engine. The soil moisture feedback data, which contains the volumetric moisture content value processed by Kalman filtering, the corresponding GPS timestamp, and the distributed sensor node number, is transmitted to the central processing module's input buffer via the LoRaWAN communication interface. The dynamic irrigation threshold parameter queue generated in step S230 and stored in non-volatile memory is synchronously invoked. This queue is sorted by time validity and contains a storage structure containing the soil moisture lower and upper thresholds, as well as identifiers for the time intervals in which they are effective.

[0083] The difference comparison operation unit performs three core operations: the first activates the timeline synchronization controller, accurately matching the feedback data's acquisition timestamp with the valid time interval of the dynamic irrigation threshold parameters. The matching mechanism uses a sliding window algorithm with a five-minute tolerance window radius. When the feedback data's timestamp falls within the valid time interval of a set of threshold parameters, that set is determined to be the current valid comparison benchmark. If a match is unsuccessful, the threshold backtracking mechanism is activated, automatically selecting the most recently valid threshold parameter set as the comparison benchmark.

[0084] The second level of operation executes the numerical deviation calculation engine, performing a dual deviation analysis on the measured soil moisture values ​​at each sensor node. The absolute difference calculator outputs the arithmetic difference between the measured moisture content and the target threshold, accurate to one decimal place. The relative percentage deviation calculator uses the formula (measured value - target threshold) / target threshold × 100% to generate a standardized deviation coefficient. During the calculation, a measured value below the lower threshold is marked as a negative deviation, while a value above the upper threshold is marked as a positive deviation, and the deviation value is recorded.

[0085] The third level of operation implements regional anomaly detection, establishing a spatial correlation matrix based on the physical topological relationships of distributed sensor nodes. The detection algorithm first calculates the average deviation of all nodes. Then, centered on a single node, it analyzes the standard deviation of the deviations of its eight neighboring nodes. When the deviation of a node exceeds the mean by plus or minus two standard deviations, and more than 30 percent of its neighboring nodes deviate in the same direction, the area is identified as an abnormal irrigation block. The coordinates of the abnormal area are encoded in the UTM coordinate system, recording the longitude and latitude offsets to six decimal places.

[0086] The final output structured difference report contains three core data types: a threshold deviation matrix, which stores the monitoring point number, soil layer depth, absolute deviation value, and relative percentage deviation in a two-dimensional array format; a trend deviation direction identifier, which uses a binary coded identifier, with a 0 at the end indicating a discrepancy between the actual moisture trend and the predicted trend in step S220, and a 1 indicating a consistency; and an anomaly region coordinate set, which records the coordinates of the center point of the anomaly block identified by topological analysis and the anomaly intensity level. This report is written to the shared memory area for use in step S520.

[0087] S520: Adjust the reward function weight of the Q-learning decision model according to the degree of difference.

[0088] Based on the difference report data output in step S510, the reinforcement learning optimization module initiates the dynamic adjustment process for the reward function weights. The system first analyzes the data in the threshold deviation matrix and performs a deviation classification mapping operation: absolute differences within 5% are set as minor deviations, with an associated penalty coefficient base of 0.1; differences between 5% and 15% are moderate deviations, with a base of 0.3; differences between 15% and 30% are severe deviations, with a base of 0.6; and differences exceeding 30% are invalid deviations, with a base of 1.0. The deviation level of each monitoring point is mapped to a corresponding penalty base through a lookup table.

[0089] Perform trend compensation correction: Read the trend deviation direction identifier. When the last digit is 1 (indicating that the predicted trend is consistent with the actual trend), a reduction factor is applied to the penalty coefficient. The reduction amount is dynamically set based on the deviation level: a slight deviation is reduced by 40%, a moderate deviation is reduced by 30%, a severe deviation is reduced by 20%, and no reduction is applied to invalid deviations. The reduction is calculated using multiplication. For example, for a moderate deviation with consistent trend, the penalty coefficient is adjusted to 0.3 × (1-30%) = 0.21.

[0090] Perform coupled correction for environmental variables: The current environmental parameters provided in step S610 are used, including soil temperature (in degrees Celsius) and light intensity (in lux). A coupled light-temperature-humidity correction model is established. When the temperature sensor returns a value above 30 degrees Celsius and the light intensity exceeds 80,000 lux, a 1.5x weighting factor is applied to the penalty coefficient for insufficient humidity (measured values ​​below the threshold). This weighting calculation uses a conditional trigger mechanism and is activated only when both temperature and light exceed the threshold.

[0091] Implement dynamic weight allocation: Substitute the revised penalty coefficients into the three-dimensional weight matrix of the reward function. This matrix contains weights for water conservation rewards (initial value 0.5), vegetation health rewards (initial value 0.3), and energy consumption control rewards (initial value 0.2). The allocation process uses a gradient descent algorithm for twenty iterations of optimization. Each iteration calculates the policy loss function value under the current weight configuration and updates the weight ratio along the negative gradient direction. The optimization goal is to reduce the deviation of decision weights that deviate from significant factors. For example, when the deviation from humidity is severe, the vegetation health reward weight is increased to 0.4, and the energy consumption control weight is reduced to 0.1. The adjusted weight matrix is ​​updated in real time to the reward rule library of the Q-learning decision model. The change record includes the old weight value, the new weight value, the change timestamp, and the environmental parameters that triggered the correction, forming a complete weight change log.

[0092] S530: Output updated irrigation decision model parameters.

[0093] After adjusting the reward function weights, the system reconstructs the decision model parameters. First, the state space dimension is expanded: Based on the new weight matrix, the state variables of the Q-learning decision model are expanded from the original three dimensions (soil moisture, temperature, and light) to four dimensions. This new dimension is injected with the vegetation health indicator acquired in step S610. This indicator, collected by the vegetation index sensor, consists of three sub-parameters: the normalized difference vegetation index, the chlorophyll content index, and the canopy temperature. Principal component analysis is used to compress this into a single health state variable.

[0094] The incremental training mechanism then begins: The system uses the last twelve valid data records collected in step S430, including six sets of pump operating status data (start / stop times, operating time, and power level) and six sets of soil moisture feedback data (moisture content sequences at five-minute intervals). This data set is used to construct a micro-training set for rapid iterative training of the Q-learning model. The training process uses a temporal difference algorithm for eighty iterations, focusing on optimizing decision paths associated with high-deviation regions in the state transition probability table. The optimization strategy is as follows: when a state transition path passes through the coordinates of an abnormal region marked in step S510, the learning rate for that path is increased to 1.5 times the normal value; when the decision action associated with the path has not triggered a deviation alarm in the historical records, the learning rate is reduced to 0.8 times the normal value.

[0095] After training is completed, three sets of core output parameters are generated: the decision kernel parameter package encapsulates the updated Q-value table (storing the action value evaluation value in the four-dimensional state space), the state transition rule set (containing 3,000 state-action transition probabilities) and the action selection strategy tree (a five-layer decision binary tree structure); the threshold control instruction template reconstructs the dynamic irrigation threshold generation logic, and associates the vegetation water demand prediction results of step S320 through the data bus. The template adds a correction coefficient for the vegetation health index to the soil moisture threshold. For every 0.1 unit decrease in the health index, the lower limit threshold of the soil moisture content increases by 2%; the model version identification code is generated using the SHA-256 algorithm, and the input parameters include the number of historical updates, the current environmental parameter fingerprint (temperature, light, health index average) and the hash value of the weight change log.

[0096] The output parameters are transmitted through the AES-256 encrypted channel, and the decision kernel parameter package is synchronized to the model loading interface of step S210; the threshold control instruction template is transmitted to the threshold generation module of step S230; the model version identification code is written into the multivariate analysis database index table of step S620, completing the closed loop of the entire process.

[0097] Step S600 at least includes steps S610-S630: S610: Obtain current light intensity data, soil temperature data, and vegetation growth health indicators.

[0098] The light intensity sensor deployed in the target irrigation area operates based on the principle of photoelectric conversion. Its internal photodiode receives the solar radiation spectrum, filters out non-visible light wavelengths through a spectral filter module, and converts the effective light radiation energy into a 4-20mA standard current signal output, with the signal strength proportional to the actual light intensity. Simultaneously, a temperature sensor array buried deep in the soil is activated. This array uses platinum resistance temperature detection elements (PT1000). Due to the positive correlation between resistance and temperature, changes in soil temperature cause an imbalance in the Wheatstone bridge circuit. The output differential voltage signal is converted by a 24-bit ADC to a digital temperature value. The light intensity and soil temperature data are directly reused with the noise-filtered data output from step S130, eliminating the need for repeated acquisition.

[0099] Vegetation health indicators are collected using a dual-modal sensing system. The visible-near-infrared spectral analysis module uses a multispectral imaging device (spectral range 400-900 nm) to scan the vegetation canopy at a 30° inclination angle. The NDVI index (range -1 to 1) is calculated by calculating the normalized difference in reflectance between the 760 nm and 680 nm bands. This index quantifies chlorophyll absorption characteristics. The morphological feature extraction module uses a 5-megapixel CMOS camera (resolution 2592 × 1944) to identify leaf contours using the Canny edge detection algorithm. The leaf curl index is calculated by analyzing the curvature radius and the saturation channel value (scale 0-255) in the HSV color space is extracted. These two health indicators are then combined into a single vegetation health parameter with a weighting ratio of 7:3.

[0100] All data is transmitted to the central processing unit via the LoRaWAN protocol. The transmission frame contains a 16-byte payload (4 bytes for light intensity, 4 bytes for soil temperature, 2 bytes for NDVI, 2 bytes for curl index, 2 bytes for saturation, and 2 bytes for checksum). Upon reception, timestamp alignment is performed: using the GPS pulse-per-second signal as a reference, linear interpolation is performed to compensate for data packets delayed by more than 500ms. This ultimately generates a structured triplet dataset of {light intensity (lux), soil temperature (°C), and vegetation health}, which is stored in a DDR4 memory buffer as the input source for the S620.

[0101] S620: Correlate the updated irrigation decision model parameters to perform multivariate collaborative analysis.

[0102] Receive the irrigation decision model parameter set output from step S530. This parameter set consists of three core elements: the updated Q-learning reward function weight matrix (a three-dimensional vector [w1, w2, w3] where w1 + w2 + w3 = 1), the state-action mapping table (a 256×4 matrix storing action values), and the action selection probability distribution (a four-dimensional vector [p1, p2, p3, p4] representing the probabilities of the four irrigation actions). Couple the triplet data set generated in step S610 with the model parameters for analysis: Constructing a four-dimensional collaborative analysis space: Using triplet data as basis vectors, we establish the environmental variable space E = {light intensity, soil temperature, vegetation health}. We overlay the action selection probability distribution vector P = [p1, p2, p3, p4] as the decision dimension, forming the [E, P] joint analysis domain. We calculate the covariance matrix using a principal component analysis algorithm. Strong correlation is determined when the calculated value exceeds 0.5. We identify the key coupling coefficient: the Pearson correlation coefficient ρ_(VT) between vegetation health and soil temperature. When |ρ_(VT)| > 0.6, we trigger the temperature compensation mechanism.

[0103] Based on the analysis results, decision parameters are dynamically optimized. When vegetation health samples fall below the 0.65 threshold for three consecutive times and light intensity consistently exceeds the historical average by 15%, the upper threshold of the soil moisture control range is reduced by 8%-12% of the current threshold. When the correlation coefficient ρ_(LT) between soil temperature and light intensity is less than -0.4, the exploration range for irrigation action selection probability is expanded from ±0.1 to ±0.25. Decision boundary adjustments are achieved by modifying the state-action mapping table: for state codes 0x3A-0x3F corresponding to the health range of 0.4-0.6, the Q value of the associated "reduce irrigation intensity" action is increased by 30%. A multi-dimensional variable association rule report is output, including the light-action probability influencing factor (0-1 scale), the health-temperature coupling coefficient (signed value), and the updated decision boundary coordinates (hexadecimal state code range).

[0104] S630: Generate new data fusion rules and decision model training data set, and feed back to step S100.

[0105] The data fusion mechanism was reconstructed based on the association rules output by S620. When the correlation coefficient ρ_(LT) between light intensity and soil temperature exceeded 0.7, the observation noise weight of the light sensor in the Kalman filter in step S120 was increased. Specifically, the noise variance of the light dimension in the R matrix was adjusted from 0.05 to 0.035 (a 30% reduction). When vegetation health decreased by more than 10% over three consecutive acquisition cycles (45 minutes), the filtering window for soil temperature data was extended from 5 seconds to 10 seconds by increasing the time constant of the state transition equation. Parameter modifications were encapsulated as an XML configuration file and written to the FPGA processing unit in step S120 via the PCIe bus, overwriting the original parameters.

[0106] To construct a training dataset for the decision model, historical data was captured using a 168-hour sliding window. This dataset includes soil moisture feedback data from step S430 (with a 5-minute sampling interval), pump operating status records (start / stop timestamps and power level), and triplet data from step S610. When adding decision labels, the graded irrigation actions performed in step S420 were encoded as four-dimensional one-hot vectors. For example, "increase irrigation intensity" was encoded as [0, 1, 0, 0]. Data augmentation was based on the association rules in step S620: when ρ_(LT) < -0.4 was identified, an extreme combination of 120,000 lux and 10°C soil temperature was synthesized. When the health level fell below 0.6, simulated data with a temperature of 35°C and a health level of 0.4 was generated. The final dataset contained 20,000 records, each containing seven fields: timestamp, environmental status triplet, action vector, soil moisture feedback value, and pump power value.

[0107] Closed-loop feedback implementation: The training dataset is stored in a MongoDB distributed database and marked as the model training input source in step S210. A hardware restart command is sent via the system control bus, making the updated Kalman filter parameters effective in the next acquisition cycle. The new dataset triggers model retraining in step S210. The feedback cycle is set to 24 hours to ensure dynamic, iterative updates of parameters across the entire chain.

[0108] Example 2: Figure 2 FIG. 1 shows a block diagram of an intelligent irrigation monitoring system according to an embodiment of the present invention. Figure 2 As shown, the structure may include: The multi-source sensor array module 10 includes a capacitive humidity sensor, a thermocouple temperature sensor, and a photoelectric light sensor. The multi-source sensor array module is used to collect real-time soil moisture, temperature, and light data through the capacitive humidity sensor, the thermocouple temperature sensor, and the photoelectric light sensor. A reinforcement learning decision module 20 is configured with a Q-learning decision model and an ε-greedy strategy selector to generate an optimal irrigation strategy; a water demand prediction module 30 that stores historical soil moisture data and runs a gradient boosting decision tree algorithm to predict future crop water requirements; The hierarchical execution module 40 includes a TLV instruction converter and a PWM water pump controller, which is used to convert the decision instructions into control signals to drive the irrigation equipment; The parameter optimization module 50 is configured with a region deviation calculator and a reward function weight regulator for dynamically optimizing system parameters; The rule iteration module 60 includes a covariance analysis engine and a Kalman filter rule generator, and is used to iteratively generate and update irrigation rules.

[0109] Obviously, the embodiments described above are only some of the embodiments of the present application, rather than all of the embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions described in the aforementioned specific embodiments, or to make equivalent replacements for some of the technical features therein. Any equivalent structure made using the contents of the present application specification and the accompanying drawings, directly or indirectly used in other related technical fields, is also within the scope of patent protection of the present application.

Claims

1. An intelligent irrigation monitoring method, characterized in that: include: Deploy capacitive humidity sensors, thermocouple temperature sensors, and photoelectric light sensor arrays, align multi-source data timing using GPS timing and a double-buffered queue, and use Kalman filtering to fuse soil moisture data, soil temperature data, and light intensity data. The capacitive humidity sensor probe is inserted into the soil profile in a vertical layered manner, and the moisture content is converted into a voltage signal output by detecting the change in dielectric constant; The thermocouple temperature sensor adopts a composite structure of thermocouple and thermistor, and is placed in the key temperature change layer from the soil surface to a depth of 40 cm. It outputs millivolt-level electrical signals through the resistance-temperature linear relationship; The photoelectric light sensor is equipped with a photodiode and a spectral filter module, fixed at a height of one meter above the vegetation canopy, and converts the solar radiation spectrum into standard lux or watts per square meter; The filtered data is used to construct a three-dimensional environmental state vector, which is then input into the Q-learning decision model to generate a four-dimensional action value evaluation vector. The ε-greedy strategy is used to select the optimal irrigation action, and the dynamic irrigation threshold is calculated based on the selected action type. Finally, the water pump control instruction is output. The water pump control instruction is encapsulated into a hexadecimal data frame through the Modbus-RTU protocol. The frame structure includes five parts: start character, device address code, function code, data area and CRC check code; The expression of the output water pump control instruction includes: ;in, is the threshold boundary correction amount; is the soil porosity tensor; is the Christoffel symbol, describing the curvature of soil heterogeneity; is the spatial gradient of moisture content; is the time step; Optimal control of water pump energy: ;in, is the Hamiltonian function of the water pump system; is the water flow momentum; is the coordinate of the water flow position; For water quality; is the gravitational potential energy function; is the pipeline topology constraint function; is the constraint strength coefficient; is the norm square operator; By correlating dynamic irrigation thresholds with historical soil moisture data, the water requirements of vegetation are predicted using a gradient boosting decision tree, generating an irrigation scheduling strategy with time windows and graded water volumes. Convert water pump control instructions and irrigation scheduling strategies into TLV format instructions, use PWM to control the water pump's staged start and stop, and collect soil moisture feedback data after shutdown; Comparing soil moisture feedback data with dynamic irrigation thresholds to calculate regional deviations, adjust the weight coefficients of the Q-learning reward function, and reconstruct the four-dimensional state decision model; The vegetation health indicators are integrated with the optimized decision model parameters, the Kalman filter rules are updated through covariance analysis, and a training data set is generated and fed back to the sensor data fusion module.

2. The intelligent irrigation monitoring method according to claim 1, characterized in that: Constructing a three-dimensional environment state vector specifically includes: The soil moisture data is mapped to the moisture content coefficient in the interval [0,1], the soil temperature data is converted into the temperature change factor, and the light intensity data is derived into the light intensity index; The 3D environment state vector is used as the activation value of the input layer of the Q-learning decision model and is transmitted to the hidden layer configured with 128 ReLU activation function neurons through a fully connected neural network. The output layer sets four groups of action nodes: maintaining the current irrigation status, increasing irrigation intensity, reducing irrigation intensity, and emergency water replenishment, and generates a four-dimensional action value evaluation vector.

3. The intelligent irrigation monitoring method according to claim 1, characterized in that: The process of selecting the optimal irrigation action using the ε-greedy strategy includes: The exploration threshold ε is initially set to 0.7 and decreases with a decay coefficient of 0.99; When the random number is higher than ε, the action corresponding to the maximum value of the action value evaluation vector is selected; when it is lower than ε, the action is randomly selected; Dynamic irrigation threshold calculation includes: lowering the lower limit of soil moisture by 3-5% when increasing irrigation intensity, raising the upper limit by 2-4% when reducing irrigation intensity, and setting a temporary threshold of 20% above the normal level for emergency water replenishment.

4. The intelligent irrigation monitoring method according to claim 1, characterized in that: The process of inputting the Q-learning decision model to generate a four-dimensional action value evaluation vector includes: Analyze the number of irrigation triggers within 24 hours through a sliding time window. When the frequency per unit time exceeds the warning value, extend the interval between adjacent actions by 10%-15%; The dynamic threshold range of soil moisture content includes three levels of response level parameters: normal / accelerated / suppressed, the minimum interval time for water pump start and stop, and the maximum number of operations per day.

5. The intelligent irrigation monitoring method according to claim 1, characterized in that: The process of calculating the dynamic irrigation threshold based on the selected action type includes: The structured threshold control group is constructed to include: soil moisture lower threshold field, upper threshold field and four-bit binary response level identifier; The response mode activation strategies include: enabling 3-5 minutes of normal mode when the last digit is 1, enabling 6-8 minutes of accelerated mode when the last two digits are 10, and enabling 1-2 minutes of suppressed mode when the last two digits are 11.

6. The intelligent irrigation monitoring method according to claim 1, characterized in that: The process of calculating the dynamic irrigation threshold according to the selected action type and finally outputting the water pump control instruction also includes: The generation of water pump start and stop control instructions includes three levels of decision logic: The first level monitors that soil moisture remains below the lower threshold for 60 seconds; The minimum interval time for the second-level calibration water pump is greater than 30 minutes; The third level checks that the number of operations per day has not reached the maximum; When the conditions are met, a four-tuple instruction is generated, which includes the device address, opening time, water flow intensity level and emergency flag.

7. The intelligent irrigation monitoring method according to claim 6, characterized in that: The process of calculating the dynamic irrigation threshold according to the selected action type and finally outputting the water pump control instruction also includes: The control instructions are encapsulated as Modbus-RTU protocol data frames, including: Start character, device address code, function code, data area and CRC check code; The first byte of the data area stores the opening duration parameter (an integer value in 0.1 seconds), the lower 4 bits of the second byte record the water flow intensity level, and the upper 4 bits retain the emergency flag.

8. The intelligent irrigation monitoring method according to claim 1, characterized in that: The process of inputting the Q-learning decision model to generate a four-dimensional action value evaluation vector includes: When the irrigation effect meets expectations, the corresponding action reward function increases the weight coefficient by 0.1; When over-irrigation or insufficient water replenishment occurs, the penalty function reduces the weight coefficient by 0.05; The fully connected weight matrix between the input layer and the hidden layer is updated through the temporal difference algorithm.

9. The intelligent irrigation monitoring method according to claim 1, characterized in that: include: When predicting vegetation water demand, the gradient boosting decision tree integrates the dynamic irrigation threshold and the time series characteristics of historical soil moisture data to generate an irrigation scheduling strategy that includes time window parameters and graded water parameters. The time window parameters are accurate to 15-minute time periods, and the graded water parameters are divided into three flow levels according to vegetation type.

10. An intelligent irrigation monitoring system, applied to the method according to any one of claims 1 to 9, characterized in that: include: Multi-source sensor array module, including capacitive humidity sensor, thermocouple temperature sensor and photoelectric light sensor; Reinforcement learning decision module, configured with Q-learning decision model and ε-greedy strategy selector; A water demand prediction module that stores historical soil moisture data and runs a gradient boosting decision tree; Hierarchical execution module, including TLV instruction converter and PWM water pump controller; Parameter optimization module, configures the area deviation calculator and reward function weight regulator; The rule iteration module includes a covariance analysis engine and a Kalman filter rule generator.

Images