Intelligent control sprinkling irrigation system for landscaping

Abstract

The application discloses an intelligent control sprinkling irrigation system for landscaping, and relates to the technical field of landscaping irrigation, which comprises a perception layer, an execution layer, an edge control layer and a cloud intelligent layer. The perception layer is provided with frequency domain reflection type soil humidity sensors with three depths of 10 cm, 40 cm and 100 cm, and is equipped with a micro-meteorological station and a pipe network pressure sensor. The cloud intelligent layer constructs a three-dimensional root system digital twin matrix, reconstructs a global water content field by using Kriging interpolation, drives a multi-agent reinforcement learning controller to generate valve and variable frequency pump station instructions, and realizes layered and fine irrigation of trees, shrubs and lawns. An extended Kalman filter self-calibration module of the edge control layer takes meteorological soft measurement values as observation benchmarks, compensates sensor drift online, and a distributed external adaptive defense decision module executes millisecond-level fusing during heavy rain, reverses soil saturation through water hammer pulses, and adaptively corrects irrigation duration during heat island effect. The application effectively solves the above three problems.

Description

Technical Field

[0001] This invention relates to the field of irrigation technology for landscaping, specifically to an intelligent control sprinkler irrigation system for landscaping. Background Technology

[0002] Garden and greening sprinkler irrigation systems are important infrastructure for the maintenance of urban green spaces, parks and landscape plants. The level of intelligence of their control systems directly determines the efficiency of water resource utilization and the quality of plant maintenance. With the expansion of urban greening and the continuous improvement of plant diversity, the limitations of traditional sprinkler irrigation control systems are becoming increasingly prominent.

[0003] The closest technical solution to the present invention is an automatic irrigation control system based on the fusion of soil moisture sensor and meteorological data. This system usually consists of a weather station, multiple soil moisture sensors, a programmable irrigation controller, an array of solenoid valves and a fixed-speed water pump. When working, the controller compares the measured soil moisture value of each zone with a preset threshold. If it is lower than the lower threshold, the corresponding zone solenoid valve is triggered to open. If it is higher than the upper threshold, the valve is closed. The water pump supplies water to the pipe network at a constant pressure at a fixed speed. All plants receive irrigation with the same water pressure and the same duration.

[0004] The aforementioned existing technical solutions have the following three defects:

[0005] Firstly, the homogenization of irrigation leads to the coexistence of drought and flood in heterogeneous plant communities. In garden settings, trees (root depth 0.8 to 2.0 meters), shrubs (root depth 0.3 to 0.8 meters), and lawns (root depth 0.05 to 0.2 meters) coexist in the same irrigation zone. The existing control system uses the measurement value of a single-depth (10 to 20 cm surface layer) soil moisture sensor as the basis for unified irrigation decisions across the entire zone. Moreover, the water pump operates at a constant pressure and the valve operates in a single on / off mode, which cannot implement differentiated water pressure and water volume for three-dimensional coordinated irrigation based on different root depths and water requirements. When the surface moisture of shallow-rooted lawns reaches the threshold, the system stops irrigation. However, at this time, the active root layer of deep-rooted trees has not been adequately replenished with water and is in a state of chronic water deficit for a long time, leading to the degeneration or even death of trees. Conversely, if irrigation is continuously carried out based on the moisture of tree roots, surface water accumulation is very likely to form in the lawn area, inducing root rot disease.

[0006] Secondly, long-term data drift from soil moisture sensors leads to systemic inaccuracies in irrigation decisions. When soil moisture sensors are buried unattended for extended periods, the output characteristics of the sensing elements undergo irreversible decay due to factors such as soil salinity, organic matter fermentation, and electrode electrochemical corrosion, resulting in systemic data drift. Existing systems lack effective in-situ self-calibration mechanisms. Once the baseline data is distorted, threshold-triggered irrigation decisions will completely fail. When the sensor's measured value is consistently low, the system misjudges long-term soil drought and continues to over-irrigate, causing a large waste of water resources and plant roots to be in a long-term anaerobic waterlogged state. When the sensor's measured value is high, the system misjudges sufficient moisture and refuses to irrigate, resulting in actual water shortage and drought for plants. Both scenarios can accumulate losses for weeks or even months before the abnormal irrigation effect is detected. Manual excavation and re-inspection are extremely costly and not feasible in engineering.

[0007] Third, existing AI irrigation models lack robustness in the face of extreme weather events. Existing machine learning irrigation models trained based on historical meteorological patterns show significant deviations from actual water demand when encountering extreme events such as localized extreme high temperatures induced by the urban heat island effect or sudden heavy rainfall, which exceed the distribution of training data. Under the urban heat island effect, local temperatures can exceed meteorological station observations by 5 to 10 degrees Celsius, but AI models irrigate normally according to meteorological station data, severely underestimating actual water demand and causing plant heat stress and death. During sudden rainstorms, the model cannot promptly identify signals that the soil has reached saturation, and may maintain or even increase irrigation output, resulting in water waste and the risk of urban flooding. To address this, we propose an intelligent control sprinkler irrigation system for landscaping. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent control sprinkler irrigation system for landscaping, in order to solve three technical problems in the existing technology: the inability to achieve precise, layered irrigation on demand for heterogeneous plant communities, the inaccuracy of decision-making due to sensor data drift, and the insufficient robustness of artificial intelligence decision-making under extreme weather events.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an intelligent control sprinkler irrigation system for landscaping, comprising a sensing layer, an execution layer, an edge control layer, and a cloud-based intelligent layer, wherein the layers are connected via an industrial Ethernet or a wireless network;

[0010] The sensing layer includes a multi-depth soil moisture sensor array, a micro weather station, and a pipeline pressure sensor.

[0011] The multi-depth soil moisture sensor array is equipped with frequency domain reflectance soil moisture sensors in each irrigation zone at three depth levels: shallow, medium and deep. The shallow layer is buried at a depth of 10 cm, corresponding to the active root layer of the lawn.

[0012] The middle layer is buried at a depth of 40 centimeters, corresponding to the main distribution layer of shrub roots;

[0013] The deep burial depth is 100 centimeters, corresponding to the core water absorption layer of the tree root system;

[0014] The micro-weather station collects real-time data on air temperature, relative humidity, wind speed, net radiation, and rainfall.

[0015] The sampling frequency of the pipeline pressure sensor is not less than 1000 Hz;

[0016] The execution layer includes a variable frequency pump station, a multi-channel intelligent solenoid valve group, and a water emitter array;

[0017] The multi-channel intelligent solenoid valve group is set up in groups according to the tree area, shrub area and lawn area;

[0018] The variable frequency pump station adjusts the pump motor speed in real time through a variable frequency driver to achieve differentiated water supply for low-pressure long-term infiltration irrigation for tree areas and high-pressure short-pulse irrigation for lawn areas.

[0019] The edge control layer includes an industrial gateway, an extended Kalman filter self-calibration module, a distributed external adaptive defense decision module, and a lightweight inference engine.

[0020] The cloud-based intelligent layer includes a 3D root system digital twin modeling server, a multi-agent reinforcement learning training and policy distribution module, and a historical database.

[0021] Furthermore, the three-dimensional root system digital twin modeling server discretizes the horizontal plane of the park into a two-dimensional grid, with each grid cell corresponding to a dominant plant type;

[0022] The soil was discretized in 5 cm thick layers along the vertical direction to construct a three-dimensional mesh.

[0023] For each 3D grid node, the root water absorption weighting coefficient is calculated based on the root depth normal distribution probability density function corresponding to the plant type at that location;

[0024] Using the sensor's measured values ​​as constraints, three-dimensional kriging interpolation is used to extend the discrete sensor measurements into a global three-dimensional water content field, thus completing the real-time status update of the digital twin matrix.

[0025] The equivalent comprehensive water shortage index for each plant type is calculated. The water shortage index is the difference between 1 and the ratio of the total weighted root water content of the plant type to the total weighted field capacity of the corresponding root system. This value is used to drive the state input of the multi-agent reinforcement learning collaborative irrigation controller.

[0026] Furthermore, the multi-agent reinforcement learning training and strategy distribution module uses the valve group and pump station pressure level of each irrigation zone as an independent decision-making agent;

[0027] Each agent's observation state vector includes the current water shortage index of trees, shrubs, and lawns in the corresponding partition, the current reference crop evapotranspiration, the current rainfall, the current total flow of the pipeline network, and the current head of the pumping station;

[0028] The actions of each agent include valve switching decisions in the tree area, valve switching decisions in the shrub area, valve switching decisions in the lawn area, and the incremental adjustment of the frequency of the variable frequency pump station.

[0029] The policy network adopts a centralized training-distributed execution paradigm for offline training in the cloud. After training, the execution network weights are compressed and distributed to the edge industrial gateway, which then completes real-time inference and command distribution with a control cycle of 1 minute.

[0030] Furthermore, the extended Kalman filter self-calibration module takes the meteorological data collected by the micro-weather station as input, calculates the reference crop evapotranspiration through the reference crop evapotranspiration formula, obtains the actual evapotranspiration of each plant type after correction by the plant crop coefficient, and combines the soil water balance equation to continuously calculate the soft measurement reference moisture content.

[0031] Sensor drift is modeled as a first-order random walk process. The three-dimensional state vector is constructed with the true water content, additive bias drift, and gain drift coefficient. The soft-sensor reference water content is used as the observation. Each state component is recursively estimated through the prediction and update steps of the extended Kalman filter.

[0032] When the absolute value of the exponentially weighted moving average of innovation continues to exceed the drift detection threshold for a predetermined period of time, the estimated compensation coefficient is automatically written into the sensor data processing pipeline to complete in-situ self-calibration without manual intervention.

[0033] Furthermore, the distributed external adaptive defense decision module includes a rainstorm circuit breaker submodule and a heat island water demand correction submodule;

[0034] When the rainstorm fuse submodule detects that the rainfall exceeds the rainstorm trigger threshold, it issues an emergency shutdown command to the variable frequency pump station and closes all zone solenoid valves within 100 milliseconds.

[0035] Subsequently, by controlling the valves in designated zones to complete rapid opening and closing with a valve opening time of 50 milliseconds, a controlled water hammer pressure transient wave is generated in the pipeline network. The pressure decay curve is collected by the pipeline network pressure sensor, and the decay coefficient is extracted by fitting using the least squares method. Soil saturation is then inverted based on the pre-calibrated decay coefficient-soil saturation mapping function.

[0036] The system remains in a circuit breaker state when soil saturation exceeds the upper limit threshold, and normal irrigation decisions are allowed to resume when soil saturation drops below the saturation recovery threshold.

[0037] The heat island water demand correction submodule compares the measured temperature of the park's micro-weather station with the benchmark temperature published by the regional weather station. When the temperature difference between the two exceeds the heat island trigger threshold, the water demand amplification multiplier is adaptively calculated based on the temperature difference using the hyperbolic tangent function, and the water demand amplification multiplier is superimposed on the irrigation duration output by the multi-agent reinforcement learning strategy.

[0038] The system recalculates the temperature difference and water demand multiplier every 10 minutes;

[0039] When the temperature difference remains below the heat island trigger threshold for an extended period, the system will automatically exit the heat island correction mode.

[0040] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0041] 1. This invention utilizes a sensor array with three standard depth levels and a three-dimensional root system digital twin matrix, combined with the differentiated pressure regulation of a variable frequency pump station, to achieve three-dimensional, on-demand, and precise irrigation for trees, shrubs, and lawns within the same pipeline network. This eliminates the problem of uneven drought and flooding in mixed planting areas and improves plant survival rate and uniform growth.

[0042] 2. This invention utilizes an extended Kalman filter in-situ self-calibration mechanism with cross-validation of hardware and software sensors to compensate for sensor drift errors in real time without manual intervention, thereby significantly extending the effective and reliable service life of the sensors and substantially reducing system maintenance costs.

[0043] 3. This invention achieves millisecond-level circuit breaker response and water hammer pulse soil saturation inversion in the event of sudden rainstorms and adaptive amplification of water demand multiplier in the event of urban heat islands through a distributed external adaptive defense decision module, ensuring the robustness of the system's decision-making under extreme climate events and effectively reducing water waste and plant losses due to disasters. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall system architecture in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of multi-agent reinforcement learning collaborative irrigation control in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the sensor online self-calibration algorithm in an embodiment of the present invention. Detailed Implementation

[0047] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0048] I. System Overall Architecture

[0049] The intelligent control sprinkler irrigation system for landscaping of the present invention adopts a three-layer collaborative architecture of cloud-edge-device, consisting of four functional modules: perception layer, execution layer, edge control layer and cloud intelligence layer. The layers are connected through industrial Ethernet or 4G / 5G wireless network. Sensor data is uploaded to the edge gateway via MQTT protocol, and control strategies are downloaded to the execution layer via Modbus protocol. The functional modules work together through standardized interfaces to form a complete closed loop of data acquisition, status perception, decision calculation, execution control and cloud optimization.

[0050] II. Perception Layer

[0051] Within each irrigation zone, frequency domain reflectance soil moisture sensors are buried at three standard depths to form a vertical profile sensing array. The shallow layer is buried at a depth of... Centimeters, corresponding to the active layer of the lawn root system, to sense the state of surface evaporation and shallow root water absorption;

[0052] Middle burial depth Centimeters, corresponding to the main distribution layer of shrub roots, to sense the dynamics of soil moisture in the middle layer;

[0053] Deep burial depth Centimeters correspond to the core water absorption layer of tree roots, sensing the infiltration and water storage status of deep soil.

[0054] The sensor nodes of each depth layer are denoted as ,in For partition numbering, Number the horizontal positions. At the corresponding depth level, each node reports the volumetric water content in real time. The unit is cubic meters per cubic meter. The frequency domain reflectance sensor utilizes the strong correlation between the soil dielectric constant and water content to measure the volumetric water content through resonant frequency drift. It has the advantages of non-contact measurement and fast response speed. Each sensor node is collected to the ground surface through the data acquisition unit and uploaded to the edge industrial gateway via wired or wireless connection.

[0055] Each irrigation area is equipped with a micro-weather station to collect the following meteorological elements in real time: temperature (Unit: degrees Celsius), relative humidity (Percentage), Wind Speed (Unit: meters per second) Net radiation (Unit: megajoules per square meter per hour) and rainfall (Unit: mm / hour) The data is uploaded to the edge control layer with a sampling period of 1 minute. The sampling frequency of the pipeline pressure sensor is no less than 1000 Hz. It is deployed at key nodes of the pipeline trunk line to collect water hammer pressure attenuation curves and real-time pipeline pressure data.

[0056] III. Execution Layer

[0057] The execution layer includes a variable frequency pump station, a multi-channel intelligent solenoid valve group, and a water emitter array. The variable frequency pump station adjusts the pump motor speed in real time through a variable frequency driver, thereby precisely controlling the outlet pressure and flow of the pipeline network.

[0058] In normal irrigation mode, the variable frequency pump station dynamically adjusts the outlet water pressure according to the frequency set value issued by the edge control layer, realizing differentiated water supply for low-pressure long-term infiltration irrigation for tree areas and high-pressure short-pulse irrigation for lawn areas. The multi-channel intelligent solenoid valve group is set according to tree areas, shrub areas and lawn areas. Each group of valves receives independent switching commands from the edge control layer. The watering array is equipped with drippers, micro-sprinklers and rotary sprinklers according to plant type. Low-flow drip irrigation heads are used in tree areas to cooperate with low-pressure infiltration irrigation, and rotary sprinklers are used in lawn areas to cooperate with high-pressure short-pulse broad-spectrum irrigation, realizing the water output mode matched with irrigation decision.

[0059] IV. Three-dimensional root system digital twin matrix modeling

[0060] The cloud-based intelligent layer constructs a three-dimensional digital twin matrix of the root system in the mixed planting area in the cloud. This matrix is ​​used to describe the water absorption and distribution characteristics of different plant types' roots in the three-dimensional space of the soil. The modeling steps are as follows:

[0061] First, geographic information data of the plant type distribution map in the park and root depth distribution parameters of each plant species were collected, including the average root depth. and standard deviation The data is derived from the Plant Physiology Database.

[0062] Secondly, the horizontal plane of the park is discretized into A two-dimensional grid, each grid cell It corresponds to a dominant plant type, including three categories: trees, shrubs, and lawns.

[0063] Next, the soil is discretized along the vertical direction. Layer, layer thickness Centimeters, constructing a three-dimensional mesh .

[0064] For each 3D mesh node Based on the normal distribution probability density function of root depth for the corresponding plant type at that location, calculate the root water absorption weighting coefficient:

[0065] ;

[0066] in, For grid nodes The root water absorption weight is dimensionless.

[0067] The standard deviation of root depth distribution for the corresponding plant type is expressed in meters.

[0068] This represents the soil depth corresponding to this grid layer, in meters.

[0069] This represents the average root depth distribution for the corresponding plant type, in meters, with weighting coefficients. The values ​​are larger at soil depths where roots are concentrated and smaller in areas where roots are sparse, objectively reflecting the different absorption capacities of different plant roots for water at different depths.

[0070] Based on sensor measured values To constrain this, three-dimensional kriging interpolation is used to extend discrete sensor measurements into a global three-dimensional water content field. The system completes the real-time state update of the digital twin matrix. Kriging interpolation achieves unbiased optimal linear estimation from finite discrete measurement points to continuous three-dimensional space by modeling the spatial variation function of sensor nodes. It can effectively adapt to the anisotropic spatial correlation characteristics of garden soil caused by the distribution of plant roots. In the system initialization stage, the variation function model is determined by fitting the historical data of the sensor.

[0071] During operation, the variogram is updated periodically with the seasons to reflect the dynamic changes in soil properties.

[0072] After obtaining the global three-dimensional water content field Then, calculate the equivalent comprehensive water shortage index for each plant type at the current moment:

[0073] ;

[0074] in, Plant type exist The water shortage index at any given time ranges from 0 to 1. The higher the value, the more severe the water shortage. A value of 0 indicates that the soil moisture content has reached the field capacity, and a value of 1 indicates that the soil is completely dry.

[0075] It belongs to the plant type A set of three-dimensional mesh nodes;

[0076] The field water holding capacity is measured in cubic meters per cubic meter and is determined according to the soil quality corresponding to the plant type. The water shortage index serves as the core state input of the multi-agent reinforcement learning collaborative irrigation controller, directly driving irrigation decisions.

[0077] V. Multi-agent reinforcement learning collaborative irrigation control

[0078] The multi-agent reinforcement learning collaborative irrigation controller treats each irrigation execution unit (corresponding to the valve group and pump station pressure level of a sprinkler zone) as an independent decision-making agent. All agents work together to optimize the overall irrigation efficiency of the park.

[0079] Each agent exist The observed state vector at time t is defined as:

[0080] ;

[0081] in, , , Partitions Current water shortage index for inner trees, shrubs, and lawns;

[0082] Reference crop evapotranspiration at the current moment, in millimeters per hour;

[0083] This is the rainfall at the current moment, expressed in millimeters per hour.

[0084] This represents the current total flow rate of the pipeline network, expressed in cubic meters per hour.

[0085] This represents the current head of the pumping station, in meters.

[0086] Action vector for each agent Defined as:

[0087] ;

[0088] in, For valve switching decisions in the tree area, 0 indicates closed and 1 indicates open;

[0089] Decisions regarding valve operation in shrubland areas;

[0090] Decisions regarding the opening and closing of valves in the lawn area;

[0091] Hertz represents the frequency adjustment increment of the variable frequency pump station. A positive value increases the pipeline pressure (corresponding to high-pressure short-pulse irrigation for lawns), while a negative value decreases the pipeline pressure (corresponding to low-pressure long-term infiltration irrigation for trees). Differentiated water pressure control is achieved through frequency adjustment.

[0092] Global reward function It consists of three parts: irrigation benefit item, over-irrigation penalty item, and energy consumption penalty item.

[0093] ;

[0094] in, , , These are the weighting coefficients for each item;

[0095] Irrigation benefits Sum the reduction rates of water shortage index for each plant type in all zones, and encourage the reduction of water shortage index to zero;

[0096] Over-watering penalty items When the water content of any plant type in any zone exceeds the field capacity, a corresponding penalty is triggered to prevent over-irrigation and waterlogging.

[0097] Energy consumption penalty item The sum of real-time electrical power of each zone's pumping station guides the system to reduce overall energy consumption while meeting irrigation needs. Typical values ​​for the three weighting coefficients are: , , It can be adjusted through hyperparameter search according to the engineering scenario.

[0098] The policy training adopts a centralized training-decentralized execution paradigm. Each agent has an independent execution network, but shares an evaluation network that can access the global state. During training, the evaluation network can access the global state and actions of all agents, thus achieving joint training of the globally optimal policy.

[0099] When the execution network performs inference at the edge, it only relies on the local observation state vector, ensuring that the inference process is lightweight and has low latency. After training converges, the execution network weights of each agent are quantized, compressed, and then sent to the edge industrial gateway.

[0100] The edge gateway collects the current state vector with a control cycle of 1 minute, generates actions by performing network forward inference (which only requires millisecond-level calculation), converts them into valve control signals and pump station inverter frequency setpoints, and sends them to the execution layer via the Modbus protocol to complete the actual irrigation actions. The cloud relies on the historical database to periodically retrain and evaluate the network, and sends the updated execution network weights to the edge to achieve continuous optimization and iteration of the strategy.

[0101] VI. Online self-calibration of soil moisture sensor based on multi-source data fusion

[0102] The extended Kalman filter self-calibration module achieves in-situ sensor self-calibration through cross-validation between soft and physical sensors, without the need for manual intervention.

[0103] The soft sensor construction method is as follows: The edge gateway calls the reference crop evapotranspiration formula hourly, using the air temperature, relative humidity, wind speed, and net radiation collected by the micro-weather station as inputs to calculate the reference crop evapotranspiration. The FAO-recommended Penman-Monteith formula is used:

[0104] ;

[0105] in, For reference crop evapotranspiration, the unit is millimeters per day;

[0106] The slope of the saturated vapor pressure-temperature curve is expressed in kilopascals per degree Celsius.

[0107] This is net radiation, expressed in megajoules per square meter per day.

[0108] Soil heat flux is expressed in megajoules per square meter per day.

[0109] This is the hygrometer constant, measured in kilopascals per degree Celsius.

[0110] The average daily temperature at a height of 2 meters, in degrees Celsius;

[0111] The wind speed at a height of 2 meters is measured in meters per second.

[0112] This is the saturated vapor pressure, measured in kilopascals (kPa).

[0113] This is the actual water vapor pressure, expressed in kilopascals.

[0114] Introducing the plant crop coefficient (For trees, use 1.0 to 1.2; for shrubs, use 0.7 to 0.9; for lawns, use 0.6 to 0.8, selected according to FAO-56 standard). Calculate the actual evapotranspiration for each plant type:

[0115] ;

[0116] in, Plant type exist The actual evapotranspiration at any given moment, expressed in millimeters per hour;

[0117] Plant type The crop coefficient is dimensionless and reflects the differences in canopy transpiration characteristics between different plants and a reference grassland.

[0118] Based on the soil water balance equation, the water content at the previous calibration time... Using the initial value, the soft-sensor reference moisture content is calculated on a rolling basis:

[0119] ;

[0120] in, Soil volumetric water content calculated by soft sensors, in cubic meters per cubic meter;

[0121] The initial water content at the last calibration reference time, in cubic meters per cubic meter;

[0122] for The amount of irrigation water injected at any given time is converted to the water depth per unit area, in millimeters per hour;

[0123] for The amount of deep seepage at any given time is estimated by the soil permeability coefficient and is expressed in millimeters per hour.

[0124] The equivalent root active layer depth is measured in millimeters, with 1000 mm for trees, 400 mm for shrubs, and 100 mm for lawns.

[0125] The integral is discretized into a cumulative summation with a step size of 1 hour.

[0126] The extended Kalman filter drift compensation method is as follows: Sensor drift is modeled as a first-order random walk process, and an additive bias drift is introduced. With gain drift coefficient The relationship between the actual sensor reading and the physical measurement reading is as follows:

[0127] ;

[0128] in, The sensor outputs the corresponding true water content (estimated by extended Kalman filter), in cubic meters per cubic meter.

[0129] This is the gain drift factor, with a nominal value of 1.0, which drifts slowly over time.

[0130] This is the raw output of the physical sensor, in cubic meters per cubic meter.

[0131] This is the additive bias drift, expressed in cubic meters per cubic meter.

[0132] The system state vector is defined as follows: Soft measurement reference moisture content As an observation, each state component is recursively estimated through the prediction and update steps of the extended Kalman filter.

[0133] The prediction step calculates the prior state estimate and the prior covariance:

[0134] ;

[0135] ;

[0136] in, for Prior state estimation vector at time step;

[0137] for Prior covariance matrix at time step;

[0138] for The posterior state estimation vector at time step;

[0139] for The posterior covariance matrix at time step;

[0140] The state transition matrix is ​​in the form of a third-order identity matrix;

[0141] The process noise covariance matrix is ​​given by the following diagonal elements: , , The process noise variance of the three state components is initialized based on engineering experience.

[0142] The steps for calculating innovation are as follows:

[0143] ;

[0144] in, The innovation quantity (observation residual) is expressed in cubic meters per cubic meter.

[0145] In the prior state vector The predicted value of the component, the innovation quantity, reflects the deviation between the observation value of the soft sensor and the prior prediction value of the physical sensor, and is the core indicator for judging whether the sensor has drifted.

[0146] After Kalman gain calculation and state update steps, the posterior state estimate is obtained. Its first component This is the calibrated estimate of the true water content, the second component. For bias estimation The third component For gain estimation posterior covariance matrix It reflects the uncertainty of the current state estimate, which converges continuously as observation data accumulates.

[0147] The system continuously monitors the index-weighted moving average of innovation volume:

[0148] ;

[0149] in, The index-weighted moving average of innovation volume, expressed in cubic meters per cubic meter;

[0150] The smoothing coefficient is set to 0.95.

[0151] For the current innovation volume, the exponentially weighted moving average can effectively filter out random noise interference and accurately identify monotonic drift trends. If the drift detection threshold (0.03 cubic meters per cubic meter) is exceeded for 6 consecutive hours, the system will automatically adjust the compensation coefficients estimated by the extended Kalman filter. and The data is written into the sensor data processing pipeline to correct the raw readings of the physical sensors in real time according to the drift correction model, completing in-situ self-calibration without manual intervention. After calibration, the system resets the moving average of the innovation quantity and continues to enter the next monitoring cycle.

[0152] VII. Distribution-based Adaptive Defense Decision Module for Extreme Climate

[0153] The distributed external adaptive defense decision module serves as an upper-level security supervision mechanism for the multi-agent reinforcement learning master control strategy. It monitors meteorological data in real time and is specifically designed to handle two extreme climate scenarios: sudden rainstorms and urban heat islands.

[0154] Sudden Rainstorm Scenario: Soil Saturation Determination via Water Hammer Pulse Inversion

[0155] When the rain gauge detects rainfall When the rainfall exceeds the rainstorm trigger threshold (15 mm / hour, reaching the rainstorm intensity standard), the edge controller issues an emergency shutdown command to the variable frequency pumping station with a response time of less than 100 milliseconds. Simultaneously, all zone solenoid valves are closed, immediately interrupting sprinkler irrigation. After the emergency shutdown is completed, the controller waits for the pipeline pressure to stabilize, and then sends a controlled pressure pulse sequence to the designated zone: rapidly opening one test valve with an opening time of 50 milliseconds, followed by immediate closing, artificially creating controlled water hammer and generating pressure transient waves in the pipeline network. The pipeline pressure sensor collects the water hammer pressure decay curve at a sampling frequency of no less than 1000 Hz and records it as a discrete time series.

[0156] The attenuation characteristics of water hammer pressure waves at the end of the pipe network (soil infiltration interface) are strongly correlated with the degree of soil saturation: the closer the soil is to saturation, the greater the impedance of the infiltration interface, the smaller the energy of the pressure wave transmitted into the soil, and the greater the attenuation coefficient of the reflected wave within the pipe network. The smaller;

[0157] The drier the soil, the lower the interfacial impedance, the faster the attenuation, and the higher the attenuation coefficient. The larger the value, the more efficient the attenuation coefficient-saturation mapping function becomes. Inverting soil saturation:

[0158] ;

[0159] in, This is an estimate of soil saturation, ranging from 0 to 1. Indicates complete saturation;

[0160] The attenuation coefficient is extracted from the pressure attenuation curve by least squares fitting, with units of 1 / second.

[0161] The empirical mapping function, established through calibration experiments under known saturation conditions during the system installation and commissioning phase, is constructed using polynomial fitting.

[0162] When the soil saturation exceeds 0.85, the system determines that the soil is nearly saturated, maintains the sprinkler irrigation interruption state, and repeats the water hammer pulse detection every set time to continuously monitor the soil condition.

[0163] When rainfall ends and soil saturation drops below 0.70, the multi-agent reinforcement learning master control strategy is allowed to resume normal irrigation decision-making authority.

[0164] Urban heat island scenario: Adaptive amplification of water demand multiplier

[0165] The system will display the actual temperature measured by the park's micro-weather station. Compared with the benchmark temperature published by the regional meteorological station Compare and calculate the heat island intensity:

[0166] ;

[0167] in, The intensity of the local heat island is expressed in degrees Celsius.

[0168] The measured air temperature at the park's micro-weather station is in degrees Celsius.

[0169] The baseline temperature is issued for regional meteorological stations, in degrees Celsius.

[0170] When the heat island intensity exceeds the heat island trigger threshold (2 degrees Celsius), the heat island correction mode is triggered, and the system adaptively calculates the water demand multiplier based on the heat island intensity.

[0171] ;

[0172] in, This is the water demand multiplier, which is dimensionless. A value greater than 1 indicates that the local water demand is higher than the prediction value of the multi-agent reinforcement learning model.

[0173] The multiplier gain parameter is set to 0.4, which means that by using the saturation characteristics of the hyperbolic tangent function, the amplification of water demand is constrained to a range of no more than 40%, thus avoiding excessive amplification in extreme cases.

[0174] It is the hyperbolic tangent function, and its range is... This ensures that the multiplier changes smoothly within a bounded range;

[0175] The heat island trigger threshold is set to 2 degrees Celsius.

[0176] The normalized scale parameter for the heat island intensity is set to 3 degrees Celsius. This parameter is used to adjust the sensitivity of the response curve so that the multiplier increases smoothly and non-linearly with the increase of temperature difference.

[0177] The water demand multiplier is added to the irrigation duration output by the multi-agent reinforcement learning master policy:

[0178] ;

[0179] in, The actual irrigation duration is adjusted for heat island effect, and the unit is seconds.

[0180] The original irrigation duration, in seconds, is output by the multi-agent reinforcement learning strategy. The system recalculates the heat island intensity and water demand multiplier every 10 minutes and dynamically adjusts the intensity.

[0181] When the heat island intensity remains below the heat island trigger threshold (2 degrees Celsius) for 30 minutes, the system automatically exits the heat island correction mode and resumes normal multi-agent reinforcement learning decision-making.

[0182] VIII. Overall System Workflow

[0183] The overall system workflow is divided into three parallel lanes: the perception layer, the edge control layer, and the cloud intelligence layer, forming a complete data closed loop.

[0184] The sensing layer uses a multi-depth soil moisture sensor array to periodically collect volumetric water content at three depth levels in each irrigation zone. The micro-weather station collects air temperature, relative humidity, wind speed, net radiation, and rainfall every 1 minute. The pipeline pressure sensor collects pipeline pressure and flow data in real time. All data are uploaded to the edge industrial gateway via the MQTT protocol.

[0185] After receiving the sensor data, the edge control layer extends the Kalman filter self-calibration module to calculate the soft measurement reference water content, performs cross-validation, recursively estimates the sensor drift compensation coefficient, and outputs the calibrated true water content. At the same time, the distributed external adaptive defense decision module continuously monitors meteorological data: if the rainfall exceeds the rainstorm trigger threshold, it executes the millisecond-level circuit breaker process and water hammer pulse soil saturation inversion.

[0186] If the heat island intensity exceeds the heat island triggering threshold, the water demand multiplier is calculated and the irrigation duration output by the multi-agent reinforcement learning strategy is corrected.

[0187] If there are no extreme weather events, the lightweight inference engine is invoked to run the execution network of each agent, generate valve switching decisions and frequency adjustment commands for the variable frequency pump station, and send them to the execution layer via the Modbus protocol. The edge control layer uploads the processed data to the cloud every hour.

[0188] The cloud-based intelligent layer receives data uploaded from the edge and updates the global three-dimensional water content field through Kriging interpolation, calculating the water shortage index for each plant type.

[0189] The evaluation network of the multi-agent reinforcement learning network is retrained regularly (weekly) with the latest historical data, and the updated execution network weights are distributed to the edge.

[0190] The real-time irrigation status and historical data trends of the park are displayed through a visualization platform.

[0191] Based on control commands, the execution layer uses variable frequency pumps to drive the intelligent solenoid valves of the corresponding zones to open as needed, and completes three-dimensional, layered, and precise irrigation of trees, shrubs, and lawns through the irrigation array. After the execution is completed, the sensor data flows back to the edge control layer to drive the next decision cycle and realize the continuous closed-loop operation of the system.

[0192] IX. Alternative Implementation Methods

[0193] In the first alternative implementation method, the active water hammer pulse excitation module is cancelled and replaced by deploying high-sensitivity acoustic emission sensors at key nodes of the pipeline network to passively monitor the low-frequency acoustic emission signals (frequency range of 1 kHz to 50 kHz) generated by soil moisture seepage. The system uses fast Fourier transform to extract the characteristic spectrum of the passive acoustic signal and uses a classifier to identify the soil saturation state, replacing the water hammer attenuation coefficient inversion method. This alternative method does not require artificially creating water hammer disturbances and is suitable for application scenarios where the pipeline materials have strict limitations on water hammer impact, such as old pipeline networks using low-pressure polyethylene hoses.

[0194] In the second alternative implementation method, the three-dimensional Kriging interpolation method is cancelled. Instead, the sensor network of the irrigation zone is modeled as a spatial graph structure, with sensor nodes as graph nodes (carrying location coordinates and measured water content attributes). Weighted graph edges are established between adjacent nodes according to spatial distance. A graph convolutional network is used to aggregate information on the graph structure. The water content value at any spatial location is predicted using sensor data from multi-hop neighborhoods as input, realizing the reconstruction of the three-dimensional water content field of the entire region. This alternative method can automatically learn the anisotropic characteristics of soil spatial correlation from historical data without assuming the form of the variation function in advance. The reconstruction accuracy is higher than that of Kriging interpolation. It is suitable for systems that have accumulated sufficient historical sensor data (usually 3 to 6 months). In the initial deployment stage, Kriging interpolation can be used as a benchmark. After the graph convolutional network training converges, the alternative method can be switched. The two schemes can be deployed in parallel in the system for comparison and verification.

[0195] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An intelligent control sprinkler irrigation system for landscaping, characterized in that, It includes a perception layer, an execution layer, an edge control layer, and a cloud intelligence layer, with each layer connected via industrial Ethernet or a wireless network; The sensing layer includes a multi-depth soil moisture sensor array, a micro-weather station, and a pipeline pressure sensor. The multi-depth soil moisture sensor array is equipped with frequency domain reflectance soil moisture sensors at three depths within each irrigation zone: shallow (10 cm), medium (40 cm), and deep (100 cm). The micro-weather station collects real-time data on air temperature, relative humidity, wind speed, net radiation, and rainfall. The execution layer includes a variable frequency pump station, a multi-channel intelligent solenoid valve group, and an irrigation device array; the multi-channel intelligent solenoid valve group is grouped into tree areas, shrub areas, and lawn areas; the variable frequency pump station adjusts the pump motor speed in real time through a variable frequency driver. The edge control layer includes an industrial gateway, an extended Kalman filter self-calibration module, a distributed external adaptive defense decision module, and a lightweight inference engine. The cloud-based intelligent layer includes a 3D root system digital twin modeling server, a multi-agent reinforcement learning training and policy distribution module, and a historical database.

2. The intelligent control sprinkler irrigation system for landscaping according to claim 1, characterized in that, The three-dimensional root system digital twin modeling server discretizes the horizontal plane of the park into a two-dimensional grid, with each grid unit corresponding to a dominant plant type; The soil was discretized in 5 cm thick layers along the vertical direction to construct a three-dimensional mesh. For each 3D grid node, the root water absorption weighting coefficient is calculated based on the root depth normal distribution probability density function corresponding to the plant type at that location; Using the sensor's measured values ​​as constraints, three-dimensional kriging interpolation is used to extend the discrete sensor measurements into a global three-dimensional water content field; The equivalent comprehensive water shortage index for each plant type is calculated. The water shortage index is the difference between 1 and the ratio of the total weighted root water content of the plant type to the total weighted field capacity of the corresponding root system. The value ranges from 0 to 1, and any value exceeding this range is treated as 0. This value is used as the state input to drive the multi-agent reinforcement learning collaborative irrigation controller.

3. The intelligent control sprinkler irrigation system for landscaping according to claim 2, characterized in that, The multi-agent reinforcement learning training and strategy distribution module uses the valve group and pump station pressure level of each irrigation zone as an independent decision-making agent. Each agent's observation state vector includes the current water shortage index of trees, shrubs, and lawns in the corresponding partition, the current reference crop evapotranspiration, the current rainfall, the current total flow of the pipeline network, and the current head of the pumping station; The actions of each agent include valve switching decisions in the tree area, valve switching decisions in the shrub area, valve switching decisions in the lawn area, and the incremental adjustment of the frequency of the variable frequency pump station. The policy network adopts a centralized training-distributed execution paradigm for offline training in the cloud. After training, the execution network weights are compressed and distributed to the edge industrial gateway, which then completes real-time inference and command distribution with a control cycle of 1 minute.

4. The intelligent control sprinkler irrigation system for landscaping according to claim 3, characterized in that, The variable frequency pump station adjusts the increment based on the variable frequency pump station frequency during the actions of each intelligent agent, and achieves differentiated water supply for low-pressure long-term infiltration irrigation for tree areas and high-pressure short-pulse irrigation for lawn areas by adjusting the speed of the water pump motor.

5. The intelligent control sprinkler irrigation system for landscaping according to claim 1, characterized in that, The extended Kalman filter self-calibration module takes meteorological data collected by micro-weather stations as input, calculates reference crop evapotranspiration using the reference crop evapotranspiration formula, obtains the actual evapotranspiration of each plant type after correction by the plant crop coefficient, and calculates the soft measurement reference moisture content by combining the soil water balance equation. A three-dimensional state vector is constructed using the true water content, additive bias drift, and gain drift coefficient. The soft-sensor reference water content is used as the observation. Each state component is recursively estimated through the prediction and update steps of the extended Kalman filter. When the absolute value of the exponentially weighted moving average of innovation quantity continuously exceeds 0.03 cubic meters per cubic meter for 6 hours, the estimated additive bias drift and gain drift coefficient are written into the sensor data processing pipeline to complete the in-situ self-calibration of the sensor.

6. The intelligent control sprinkler irrigation system for landscaping according to claim 1, characterized in that, The distributed adaptive defense decision module includes a rainstorm circuit breaker submodule; When the rainstorm fuse submodule detects that the rainfall exceeds 15 mm per hour, it issues an emergency shutdown command to the variable frequency pump station and closes all zone solenoid valves within 100 milliseconds. Subsequently, the valves in the designated zones are controlled to open and close rapidly in a 50-millisecond opening time, generating a controlled water hammer pressure transient wave in the pipeline network. The pressure decay curve is collected by the pipeline network pressure sensor, and the decay coefficient is extracted by fitting using the least squares method. Soil saturation is then inverted based on the pre-calibrated decay coefficient-soil saturation mapping function. When soil saturation exceeds 0.85, sprinkler irrigation is suspended; when soil saturation drops below 0.70, normal irrigation can be resumed.

7. The intelligent control sprinkler irrigation system for landscaping according to claim 1, characterized in that, The distributed external adaptive defense decision module also includes a heat island water demand correction submodule; The heat island water demand correction submodule uses the difference between the measured temperature at the park's micro-meteorological station and the benchmark temperature published by the regional meteorological station as the heat island intensity. When the heat island intensity exceeds 2 degrees Celsius, the water demand amplification multiplier is adaptively calculated based on the heat island intensity using the hyperbolic tangent function, and the water demand amplification does not exceed 40% of the irrigation duration output by the original multi-agent reinforcement learning strategy. The water demand multiplier is added to the irrigation duration output by the multi-agent reinforcement learning strategy to form the corrected irrigation duration; The urban heat island water demand correction submodule recalculates the urban heat island intensity and water demand amplification multiplier every 10 minutes. When the heat island intensity remains below 2 degrees Celsius for 30 minutes, the heat island correction mode is exited, and normal multi-agent reinforcement learning decision-making is resumed.

8. A control method for an intelligent control sprinkler irrigation system for landscaping as described in any one of claims 1 to 7, characterized in that, The process includes the following: A multi-depth soil moisture sensor array collects the volumetric water content of shallow, middle and deep layers in each irrigation zone in real time; a micro-weather station collects air temperature, relative humidity, wind speed, net radiation and rainfall in real time; and a pipeline pressure sensor collects pipeline pressure data in real time. The extended Kalman filter self-calibration module of the edge control layer calculates the soft measurement reference water content based on meteorological data, cross-validates the soft measurement reference water content with the sensor's measured value, recursively estimates the sensor drift compensation coefficient, and outputs the calibrated true water content. The distributed adaptive defense decision module monitors rainfall and heat island intensity in real time. When the rainfall exceeds the rainstorm trigger threshold, it performs millisecond-level circuit breaking, controlled water hammer pulse excitation and soil saturation inversion decision. When the heat island intensity exceeds the heat island trigger threshold, it performs water demand multiplier calculation and irrigation duration correction. Based on the calibrated real water content, the cloud-based intelligent layer constructs a global three-dimensional water content field through three-dimensional kriging interpolation, calculates the water shortage index of each plant type, and drives the multi-agent reinforcement learning controller to generate valve switching decisions and frequency adjustment commands for each zone and variable frequency pump station. Based on control commands, the execution layer uses variable frequency pumps to drive the intelligent solenoid valves of the corresponding zones to open as needed, thereby completing the layered and precise irrigation of trees, shrubs, and lawns.

Images