Precision irrigation system based on machine learning

Abstract

The application relates to the field of agricultural water conservancy and information sensing technology, and discloses a precision irrigation system based on machine learning, which comprises a multi-source heterogeneous fluid excitation module, a fluid-solid coupling response collection module, a transient signal feature extraction and decoupling module, a physical information space-time inversion module and a closed-loop decision and precision execution module. The system actively injects transient pressure waves through the excitation module, collects fluid-solid coupling response signals modulated by soil along the line, and inverses the global wave velocity distribution of the pipe network and the soil elastic modulus by using a physical information neural network after signal decoupling, and finally executes closed-loop control based on the reconstructed agronomic parameters. The application uses a flexible pipe as a sensing carrier, constrains a deep learning model by using a physical mechanism, overcomes the poor representativeness of traditional point monitoring and the ill-conditioned problem of inverse problem solving, and realizes low-cost, high-precision distributed sensing and intelligent regulation and control of soil moisture content.

Description

Technical Field

[0001] This invention relates to the fields of agricultural water conservancy engineering and information sensing technology, specifically a precision irrigation system based on machine learning. Background Technology

[0002] The core of precision irrigation systems lies in the accurate perception and decision-making regarding the spatiotemporal distribution of soil moisture in the field. Currently, mainstream farmland soil moisture monitoring mainly relies on point-contact sensors such as frequency domain reflectometers (FDR) or time domain reflectometers (TDR). However, farmland soil physical properties exhibit high spatial heterogeneity, making it difficult for sparsely arranged discrete sensors to reconstruct a representative field moisture field distribution. Increasing sensor deployment density, on the other hand, significantly increases equipment costs, cable maintenance difficulties, and hinders field operations.

[0003] To address the challenges of distributed sensing, water hammer wave-based pipeline defect diagnosis offers a potential approach: utilizing the propagation characteristics of transient pressure waves within the pipeline to invert its state. However, reusing water supply networks as sensor networks in agricultural irrigation scenarios presents significant challenges. Traditional transient flow inversion (ITA) methods primarily rely on deterministic physical models for iterative solutions. This process is a typical inverse problem, exhibiting severe ill-conditioned characteristics, extreme sensitivity to signal noise, and computational efficiency insufficient for real-time monitoring. While deep learning methods have been widely applied in fault diagnosis in recent years, purely data-driven neural network models require massive amounts of supervised labeled data for training. In real agricultural environments, obtaining comprehensive, high-density soil elastic modulus or moisture content as training labels is extremely difficult, resulting in poor generalization ability of purely data-driven models under small sample conditions, making it difficult to accurately capture complex fluid-structure interaction characteristics.

[0004] Furthermore, most existing irrigation control systems only have single flow or pressure regulation functions, with the pipeline network serving merely as a water delivery channel, lacking the ability to actively regulate the physical properties of the fluid medium. When soil exhibits abnormal physical structures such as compaction or cavitation, simple water replenishment cannot improve the root zone microenvironment. The lack of deep collaboration and closed-loop feedback based on physical mechanisms between the monitoring and execution systems limits the operational efficiency and intelligence level of irrigation systems under complex operating conditions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a precision irrigation system based on machine learning. This system solves the problems of poor spatial representativeness and high deployment and maintenance costs caused by the reliance on discrete point sensors in traditional farmland soil moisture monitoring, as well as the ill-conditioned solutions and weak generalization capabilities of existing inversion algorithms due to their reliance on a large number of hard-to-obtain soil ground truth labels.

[0006] To achieve the above objectives, the present invention provides a precision irrigation system based on machine learning, which mainly includes a multi-source heterogeneous fluid excitation module, a fluid-structure interaction response acquisition module, a transient signal feature extraction and decoupling module, a physical information spatiotemporal inversion module, and a closed-loop decision-making and precision execution module.

[0007] The multi-source heterogeneous fluid excitation module is physically connected to the head of the field pipeline system to generate variable modulus fluid and coded wave sources, thereby actively injecting transient pressure waves into the field pipeline system. The fluid-structure interaction (FSI) response acquisition module, located at the head of the field pipeline system, is responsible for acquiring the FSI transient pressure response signal formed by the modulation of the soil constraint state along the pipeline as the transient pressure wave propagates through the system. The transient signal feature extraction and decoupling module receives the FSI transient pressure response signal and performs denoising processing, thereby decoupling the independent channel response features corresponding to specific valve locations from the total signal. The physical information spatiotemporal inversion module receives the decoupled response features, constructs a physical information neural network to invert the system wave velocity distribution across the entire field pipeline system, and derives the soil elastic modulus distributed along the pipeline. The closed-loop decision-making and precise execution module receives the inverted soil elastic modulus and reconstructs it into agronomic state parameters. Based on these parameters, it generates control commands and feeds them back to the multi-source heterogeneous fluid excitation module to execute irrigation or anomaly handling tasks.

[0008] In a preferred embodiment, the multi-source heterogeneous fluid excitation module comprises a gas-liquid two-phase rheological modulus control subsystem and a distributed valve-controlled inverse perturbation coding subsystem. The gas-liquid two-phase rheological modulus control subsystem adjusts the gas content in the fluid to set different detection reference wave velocities, thereby changing the interaction scale between the pressure wave and the pipe-soil system. The distributed valve-controlled inverse perturbation coding subsystem controls the valve actions on the branch pipes of the field pipe network system based on a pseudo-random binary sequence, generating orthogonal water hammer pulse signals, thereby achieving spatial differentiation of multi-source signals at a single-point receiving end.

[0009] This invention operates based on a fluid-structure interaction (FSI) sensing mechanism for flexible pipelines. Flexible pipelines in a field pipeline network undergo radial deformation under transient pressure. This radial deformation is physically constrained by the elastic modulus of the surrounding soil, which in turn determines the propagation speed of the transient pressure wave within the pipeline. The FSI response acquisition module utilizes this mechanism, based on the correlation between propagation speed and physical constraints, and employs the principle of equivalent wave velocity incorporating soil elastic modulus variables, to establish a quantitative mapping relationship between soil confining pressure, pipeline wall thickness, pipeline material modulus, and system wave velocity. This transforms changes in soil physical state into observable changes in wave velocity.

[0010] For signal processing, the transient signal feature extraction and decoupling module employs a time-frequency analysis method based on continuous wavelet transform for denoising. This method decomposes the fluid-structure interaction (FSI) transient pressure response signal into multiple scales by selecting a mother wavelet function that matches the characteristics of the FSI transient pressure response signal. It then uses wavelet domain threshold filtering to remove steady-state mechanical noise and high-frequency electronic noise, and reconstructs a clean pressure response signal that retains the transient waveform characteristics through inverse wavelet transform. When decoupling independent channel response features, the module uses a multi-input single-output signal decoupling and localization algorithm. It calculates the cross-correlation function between the total response signal and the preset coding sequences of each valve using cross-correlation analysis. By identifying peak delays, it separates the sub-channel responses at specific valve locations and corrects the global transmission delay introduced by the gas content based on the theoretical reference wave velocity.

[0011] The core of this invention at the algorithm level lies in the physical information spatiotemporal inversion module. This module constructs a physical information neural network employing a fully connected deep feedforward architecture, establishing a nonlinear mapping relationship from the spatiotemporal coordinate vector of the input layer to the fluid head field, flow field, and system wave velocity field of the output layer. The network utilizes an automatic differentiation mechanism to calculate the partial derivatives of the output layer variables with respect to the input layer variables, and uses these partial derivatives as the basis for constructing the physical constraint terms. During training, the network embeds physical constraint equations, including the continuity equation describing the conservation of transient flow mass, the momentum equation describing the conservation of momentum, and the constitutive equation describing the nonlinear mapping relationship between system wave velocity and soil elastic modulus. The goal of network training is to minimize the weighted sum of the fitting error of boundary observation data and the residuals of the global physical equations.

[0012] Furthermore, the physical information spatiotemporal inversion module employs a composite loss function and a hybrid optimization strategy. The composite loss function is composed of a weighted average of the fitting error and the physical residual, ensuring that the inversion results conform to both the observed data and physical laws. The hybrid optimization strategy uses an adaptive moment estimation optimizer for rapid convergence in the early stages of training, and switches to a finite-memory quasi-Newton optimizer for fine-tuning in the later stages of training. The weight coefficients of each loss term are dynamically adjusted based on the statistical characteristics of the gradients during training to balance the contributions of data-driven and physics-driven approaches.

[0013] At the application level, the closed-loop decision-making and precise execution module maps the soil elastic modulus obtained by inversion to volumetric water content using the inverse function of the soil constitutive relation, and then corrects it using measured data from in-situ anchor point sensors. Subsequently, using the Kriging interpolation algorithm combined with the topology of the field pipeline system, the one-dimensional parameters distributed along the pipeline are reconstructed into a global two-dimensional humidity distribution map. Based on this distribution map, the system executes a zoning decision-making algorithm and a hydraulic constraint scheduling strategy to calculate the water demand of each zone and convert it into valve opening duration. A dynamic time-segmentation algorithm is used to divide the irrigation groups to ensure pipeline flow balance. In addition, when anomalies in modulus corresponding to soil compaction or cavities are identified, the anomaly response mechanism generates control commands to drive the multi-source heterogeneous fluid excitation module to start a pure bubble aeration mode, thereby achieving physical improvement of the soil microenvironment.

[0014] This invention provides a precision irrigation system based on machine learning. It has the following beneficial effects: 1. This invention utilizes a fluid-structure interaction response acquisition module and MISO signal decoupling technology to directly use flexible water conveyance pipelines in the field as sensors for sensing the soil environment. By capturing the wave velocity changes caused by the constraint of the external soil on transient pressure waves propagating inside the pipe, a distributed inversion of the elastic modulus and moisture content of the entire field soil is achieved. This feature eliminates the need to lay expensive point sensor arrays or complex cables in the farmland, reducing the cost and difficulty of hardware deployment and maintenance. It also solves the problems of poor spatial representativeness and difficulty in reflecting the overall soil moisture distribution of traditional single-point monitoring data.

[0015] 2. This invention constructs a physical information neural network in the physical information spatiotemporal inversion module, embedding physical mechanisms such as the one-dimensional Saint-Venant equations and fluid-structure interaction constitutive relations as strong constraints into the loss function. This design allows network training to no longer rely entirely on massive amounts of labeled data, but rather utilizes sparse boundary observation data and dense physical laws to jointly drive model convergence. This effectively overcomes the challenge of obtaining high-precision ground truth soil samples in agricultural scenarios, ensuring that the inversion results conform to physical laws, and improving the model's generalization ability and prediction accuracy in complex noisy environments.

[0016] 3. This invention achieves dual functions of detection and remediation through the collaborative operation of a multi-source heterogeneous fluid excitation module and a closed-loop decision-making and precision execution module. The system utilizes micro-nano bubbles to actively adjust the fluid modulus, generating detection wave sources with different reference wave velocities to correct measurement errors introduced by the bubbles. Simultaneously, when soil compaction or cavitation is detected, the system automatically switches to a pure bubble aeration mode, delivering oxygen-rich water to the root zone. This mechanism upgrades simple water replenishment to water-air coupled soil microenvironment improvement, enhancing operational efficiency under complex conditions and improving the quality of the crop growth environment. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the overall system structure and application scenarios of an embodiment of the present invention; Figure 2 This is a schematic diagram of the physical information spatiotemporal inversion module architecture according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the fluid-structure interaction sensing mechanism of a flexible pipeline according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See attached document Figure 1 This embodiment provides a precision irrigation system based on machine learning, which is mainly composed of three physically connected parts: an irrigation head hub unit, a field fluid-structure interaction sensing network unit, and a communication and computing control unit.

[0020] The irrigation headworks unit, serving as the control source and signal acquisition terminal for the fluid medium, is physically located within the water source or pump house. Along the water flow direction, this unit sequentially includes a water pump, a filtration system, a micro-nano bubble generator, a flow control valve assembly, and a high-frequency dynamic pressure monitoring component.

[0021] Micro-nano bubble generator: Connected to the main water supply pipe via bypass or series connection. The device includes an inlet control valve, a gas-liquid mixing pump, and dissolved gas release nozzles. The inlet control valve is electrically connected to the I / O interface of the main control computer to regulate the volumetric flow rate of the inhaled gas. The device is configured to generate microbubbles with diameters between 100 nanometers and 50 micrometers and uniformly disperse the bubbles into the irrigation water, forming a gas-liquid two-phase mixed fluid. A one-way valve is installed at the device outlet to prevent backflow.

[0022] High-frequency dynamic pressure monitoring component: Installed on the main pipe downstream of the micro-nano bubble generator, at the inlet of the field pipeline network. The core of the component is a high-frequency dynamic pressure sensor, whose sensing element is in direct contact with the fluid. The sensor's frequency response range is configured from 0Hz to 20kHz, and the sampling frequency is set to no less than 10kHz to capture millisecond-level transient characteristics of water hammer waves. The sensor output is connected to a high-speed data acquisition card (DAQ).

[0023] Flow monitoring component: An electromagnetic flow meter or ultrasonic flow meter is installed near the pressure sensor to measure instantaneous flow velocity and cumulative flow, with a data refresh rate of not less than 1Hz.

[0024] The field fluid-solid coupling sensing network unit serves as both a water transport channel and a physical sensing carrier, and is laid in the farmland area to be irrigated.

[0025] Flexible fluid-structure interaction (FSI) piping: The irrigation network consists of main pipes, branch pipes, and capillary pipes (drip tape or seepage pipes). The capillary pipes are made of low-modulus polymers, including linear low-density polyethylene (LLDPE) or polyvinyl chloride (PVC). The capillary pipe wall thickness ranges from 0.2 mm to 1.0 mm. The piping is laid underground at a depth of 20 cm to 50 cm below the surface, ensuring direct contact between the pipe's outer wall and the root zone soil, forming a pipe-soil coupling physical interface. The radial stiffness of the piping is designed to allow for slight radial expansion deformation under internal pressure pulses.

[0026] Distributed solenoid valve array: Solenoid valves are installed at the inlet of each branch pipe or specific capillary group. The solenoid valves are direct-acting or pilot-operated high-speed solenoid valves with high-frequency switching capability and a response time of less than 50 milliseconds. All solenoid valves are spatially distributed in a matrix or tree topology, and each solenoid valve has an independent physical address code. The solenoid valves not only serve as the on / off actuators for irrigation water but also as signal sources for generating reverse pressure disturbances.

[0027] The communication and control link is responsible for connecting the head hub, field actuators, and the main control computing platform.

[0028] Main control computer: Deployed in the main control room or cloud server, with built-in Physical Information Neural Network (PINN) inversion algorithm module and irrigation decision module. The computer connects to a high-speed data acquisition card via industrial bus (such as RS485, Modbus) or industrial Ethernet to read time-series pressure data and flow meter data collected by high-frequency dynamic pressure sensors.

[0029] Valve control network: The main control computer connects to the distributed solenoid valve array in the field via a programmable logic controller (PLC) or an IoT gateway. The control network supports concurrent command transmission and can send pseudo-random binary sequence (PRBS) control signals to designated solenoid valves to drive them to perform high-frequency micro-twitch actions.

[0030] Gas-liquid control link: The main control computer is connected to the air intake control valve and mixing pump frequency converter of the micro-nano bubble generator, and the gas content is adjusted through analog output signals (420mA or 010V).

[0031] The flexible fluid-structure interaction pipe is surrounded by soil medium. The inside of the pipe is filled with a gas-liquid two-phase mixture.

[0032] Fluid domain: contains liquid water and discretely distributed microbubbles, whose equivalent bulk modulus is controlled by the micro / nano bubble generator.

[0033] Solid domain (pipe wall): has a fixed material elastic modulus and geometric dimensions.

[0034] Environmental domain (soil): The soil acts as a continuous medium, constraining the radial displacement of the pipe wall. The radial constraint stiffness of the soil on the pipe wall is determined by the soil's elastic modulus. At the physical contact surface, the normal stress on the outer surface of the pipe wall is balanced by the reaction force provided by the soil.

[0035] The entire system does not include electronic humidity sensors, wireless sensing nodes, or battery-powered devices buried in the field soil. All electronic sensing elements are centralized in the high-frequency dynamic pressure monitoring component at the head hub, and all active field devices are simply distributed solenoid valves connected to the pipeline power supply. The multi-source heterogeneous fluid excitation module serves as the signal source for the entire system. Its function is to generate a fluid medium with controllable physical properties and to generate coded pressure waves using distributed actuators, providing multi-dimensional input signals for subsequent sensing and inversion. Physically, it consists of two parts: a gas-liquid two-phase rheological modulus control subsystem integrated into the irrigation headworks, and a distributed valve-controlled inverse disturbance coding subsystem deployed in the field pipeline network.

[0036] The gas-liquid two-phase rheological modulus control subsystem is used to actively change the physical properties of the fluid medium in the pipeline, specifically by changing the equivalent density and equivalent bulk modulus of the mixed fluid. The gas-liquid two-phase rheological modulus control subsystem mainly consists of a micro-nano bubble generator, an air inlet flow control valve, and a fluid control logic unit in the main control computer. Its physical location is on the main water delivery pipeline of the irrigation head.

[0037] When the gas-liquid two-phase rheological modulus control subsystem is in detection mode, the main control computer controls the opening of the air intake flow control valve through analog signals, introducing a specific volume flow rate of gas (such as air or pure oxygen) into the water flow. After the high-speed shearing action of the gas-liquid mixing pump, a gas-liquid mixed fluid containing uniform microbubbles is formed.

[0038] The physical essence of this process is the active adjustment of the compressibility of the fluid medium within the pipe. The density of the mixed fluid... Following the two-phase flow mixing law, based on the liquid phase density gas phase density And the gas phase volume fraction (i.e., gas content). Decision. Among them, gas content Defined as the ratio of the gas phase volume to the total volume of the gas-liquid mixture. The relationship for calculating the mixture density is: ; Because the density of gas is much smaller than that of liquid, and the gas content is... In this embodiment, the amount is controlled within a minute range (e.g., 0 < 0). The density of the mixed fluid changes relatively little (<0.05). However, the bulk modulus of the mixed fluid... It exhibits high sensitivity to changes in gas content. Because gases are much more compressible than liquids, the introduction of even small amounts of air bubbles significantly reduces the stiffness of the fluid mixture. The speed of sound (pressure wave propagation) in the fluid mixture... It depends on the ratio of the bulk modulus to the density of the mixed fluid. According to Wood's equation, the theoretical wave velocity of a gas-liquid two-phase mixture is described as follows: ; In this formula, The bulk modulus of the liquid phase (water). This refers to the bulk modulus of the gas phase. During an adiabatic process, the bulk modulus of the gas phase... Approximately equal to the adiabatic index With absolute pressure The product of, i.e. .because (Approximately 2.2 GPa) much greater than (approximately 0.14 MPa), even Only 0.01 (i.e. 1%), in the denominator This item will also take the lead, resulting in The speed of sound in pure water (approximately 1480 m / s) drops drastically to hundreds of meters per second or even lower.

[0039] Based on the aforementioned physical mechanism, this subsystem performs a variable medium scanning operation. Instead of sending a single, fixed gas content command, the main control computer generates a gas content control sequence that varies over time. For example, control sequence settings. The concentration is switched sequentially between 0.1%, 0.5%, 1.0%, and 2.096, corresponding to each set gas content. The fluid medium within the pipeline network obtains a specific wave velocity reference. .

[0040] Hardware transmission frequency Changing wave speed while keeping it constant or under constraints Directly changed the wavelength of the pressure wave Their relationship is By adjusting the gas content, the system physically scales the wavelength of the detection signal, altering the interaction scale between the pressure wave and the pipeline-soil coupling system. This variable modulus scanning mechanism enables the system to excite and detect soil-pipeline coupling resonance modes that cannot be excited in a single pure water medium, thereby acquiring multi-dimensional physical response data without increasing the number of sensors. The generated mixed fluid then enters the field pipeline network, serving as the physical carrier and sensing medium for the detection signal.

[0041] The distributed valve-controlled reverse disturbance coding subsystem utilizes the existing branch pipe solenoid valve array in the field pipeline network as a distributed signal transmission source. In conventional irrigation mode, the solenoid valves are only used to control the water flow; in the global sensing mode, the solenoid valves function as hydraulic acoustic exciters.

[0042] The main control computer sends data to devices located in different geographical locations. Each solenoid valve unit sends independent control commands. These commands are not continuous open or close signals, but rather time-sequential pulses generated based on a pseudo-random binary sequence. The solenoid valve unit responds to these pulse signals by performing local frequency micro-jumping actions. Micro-jumping is defined as the valve core performing a small-amplitude opening and closing cycle within a short time (half a second), or rapidly jittering within a small opening range, rather than a full opening or closing. This mechanical valve core jittering causes a transient change in the local fluid velocity across the valve cross-section. According to water hammer theory, this velocity change generates transient pressure waves in incompressible or weakly compressible fluid media. These transient pressure waves propagate upstream against the flow in the pipeline network, carrying physical path information from the signal source (valve position) to the receiving end (head sensor).

[0043] To effectively distinguish multiple signal sources from different spatial locations at a single-point receiver and prevent information loss due to aliasing of multiple source signals, the system employs orthogonal coding or non-correlated coding strategies. Define the... The action state function of each electromagnetic valve is: .

[0044] The pressure disturbances generated by all valve units are physically superimposed in the pipeline network, ultimately forming an induced flow rate at the main pipeline location. Described as the sum of contributions from each dispersed wave source: ; In the formula, Represents the total number of solenoid valves involved in the action; For the first The flow coefficient of a valve, which characterizes the influence weight of the valve's actuation amplitude on the flow rate, depends on the valve's physical diameter and the amplitude of the vibration. For the first The time delay parameters of the encoded sequence of each valve relative to the system reference clock are used to further distinguish different signal sources on the time axis; This represents the baseline disturbance flow rate for a single valve unit under nominal operating conditions. Through a distributed coding excitation mechanism, the system constructs a multi-input, single-output reverse sonar detection field, enabling the sensor at the head to accurately locate physical state changes at the end of the pipeline network or in specific branches via signal demodulation algorithms. This overcomes the physical limitation of insufficient end-point detection sensitivity due to signal attenuation during long-distance transmission from a single wave source.

[0045] See attached document Figure 3 The fluid-structure interaction response acquisition module is the core component of the system for obtaining physical environment feedback. It mainly includes two parts: flexible pipe fluid-structure interaction sensing mechanism and sensitivity analysis, and physical response characteristic analysis.

[0046] The flexible pipeline fluid-structure interaction sensing mechanism utilizes a flexible irrigation network buried in the field as a distributed physical sensing array. It achieves sensing by capturing the response of the fluid pressure wave propagation characteristics within the pipe to the external environment. This mechanism physically relies on the mechanical coupling interface formed between the flexible pipeline structure and the surrounding soil medium, as well as the fluid dynamic monitoring components installed at the head of the pipeline.

[0047] At the physical mechanism level, buried flexible pipelines (such as polyethylene drip irrigation tape or seepage pipes) exhibit significant fluid-structure interaction effects. When the fluid inside the pipe transmits transient pressure pulses, the pipe wall experiences dynamic circumferential stress and undergoes slight radial expansion or contraction deformation. For pipelines placed in free space, the deformation is limited only by the stiffness of the pipe material itself; however, for irrigation pipelines buried underground, the radial deformation of the pipe wall is physically constrained by the surrounding soil medium (confining pressure).

[0048] The constraint effect of the soil medium on the pipe wall is mechanically equivalent to a distributed spring system acting on the outer wall of the pipe. The radial constraint stiffness of the soil on the pipe is not a constant value, but directly depends on the physical state parameters of the soil, especially the soil compaction and moisture content. When a pressure wave passes through the pipe, the pipe cross-section is actually a complex dynamic system composed of fluid, pipe wall, and soil.

[0049] According to the theory of transient fluid flow and water hammer, the propagation speed of pressure waves in a pipeline system (i.e., the system wave velocity) is... The wave velocity depends not only on the compressibility of the fluid medium inside the pipe, but also strictly on the combined elastic properties of the pipe cross-section. For buried flexible pipelines constrained by soil, the equivalent wave velocity model considering fluid compressibility and the combined stiffness of the pipe and soil is described as follows: ; In the formula, The bulk modulus of the aforementioned gas-liquid mixture is... The density of the mixed fluid. Represents the inner diameter of the pipe. These two parameters, representing the pipe wall thickness, define the pipe's geometric characteristics. This is the inherent elastic modulus of the pipe material (such as PE, PVC), and is a known material constant. Location along the pipeline The soil elastic modulus is a physical variable to be measured that varies with spatial location and soil moisture content. This is the soil-pipe interaction coefficient, used to correct the geometry of the pipe-soil interface and the stress transfer efficiency.

[0050] The physical model described above shows the propagation speed of the pressure wave inside the pipe. Directly affected by the soil modulus outside the pipe Modulation of soil modulus. It serves as a variable parameter in the wave velocity equation. When the soil environment changes (e.g., irrigation leading to increased humidity), the matrix suction between soil particles and the pore water pressure change, resulting in an alteration in the macroscopic soil modulus. Drift occurs, altering the effective radial constraint stiffness of the pipe cross-section, ultimately causing a shift in the physical value of the pressure wave velocity within the pipe. Therefore, the system does not need to directly contact the soil; it only needs to monitor the wave velocity through the fluid medium. By observing the spatiotemporal distribution characteristics, a physical correlation can be established with the mechanical state of the soil outside the pipe.

[0051] Sensitivity analysis and physical response characteristic analysis elucidate how changes in soil physical state are quantitatively mapped to observable pressure wave signal characteristics. At the physical level, soil modulus... Its moisture content There exists a nonlinear functional relationship between them. This relationship can be described using constitutive models in geotechnical engineering, such as the soil-water characteristic curve (SWCC) combined with the effective stress principle. Define the soil stiffness function. Indicates soil elastic modulus With moisture content The pattern of change: ; Generally, for unsaturated cohesive soils or sandy soils, as the water content increases... As the soil becomes wetter (from dry to moist), the matric suction between soil particles decreases, the effective stress decreases, and the macroscopic elastic modulus of the soil decreases. The modulus decreases. Conversely, when soil loses water and dries out, its modulus increases accordingly. Therefore, changes in local soil moisture content... This will directly lead to changes in local soil modulus. This, in turn, through the aforementioned fluid-structure interaction mechanism, causes the local system wave velocity to... offset This physical chain, from water content to wave velocity, constitutes the core sensing sensitivity of the system.

[0052] Besides changes in wave velocity, alterations in soil physical state can also cause other changes in the characteristics of pressure wave signals. Specific physical response characteristics include: Crest delay: When a pressure wave pulse passes through a region with high moisture content (low modulus region), the propagation time of the wave pulse in this section will be prolonged due to the local reduction in wave velocity. This manifests as a measurable delay in the arrival time of the characteristic peak of the reflected or transmitted wave signal received by the head sensor relative to the reference time under dry conditions.

[0053] Amplitude attenuation: The viscoelastic properties of soil lead to energy dissipation. When a pressure wave undergoes radial deformation in the pipe wall, some acoustic energy is transferred into the soil and converted into heat energy due to internal soil friction. The viscous damping coefficient of the soil is also related to its moisture content. Therefore, soils with different moisture contents will absorb pressure waves to varying degrees, resulting in changes in the amplitude attenuation rate of the received signal.

[0054] Waveform distortion: Pressure wave propagation exhibits frequency dependence, meaning that different frequency components have different propagation velocities and attenuation rates (dispersion effect). The soil medium also responds differently to sound waves of different frequencies. When a broadband pressure pulse passes through a pipe-soil coupling system, different frequency components experience different phase shifts and attenuations, leading to broadening, tilting, and other distortions in the received pulse waveform. The characteristic patterns of this distortion directly reflect the viscoelastic spectral properties of the soil medium.

[0055] In summary, the time-series pressure data acquired by the first high-frequency dynamic pressure sensor The data not only includes information about wave propagation time (related to wave speed), but also information about amplitude and waveform morphology changes. These complex physical response characteristics together constitute a comprehensive fingerprint for inverting the soil conditions along the pipeline network.

[0056] The transient signal feature extraction and decoupling module is responsible for processing the raw pressure data acquired by the head sensor. Its function is to suppress irrelevant noise and separate effective feature components carrying specific physical information from complex superimposed signals. This module mainly includes two parts: signal preprocessing and denoising, and the MISO signal decoupling and localization algorithm. The primary task of the signal preprocessing and denoising process is to process the raw time-series pressure data output by the high-frequency dynamic pressure sensor. The pressure sensor continuously samples at a sampling frequency of at least 10 kHz, generating a high-density data stream. The raw data is usually mixed with various interference signals, mainly including periodic mechanical vibration noise related to pump operation, high-frequency electronic white noise, and possible electromagnetic interference.

[0057] To extract the excited transient pressure wave response, this procedure employs a time-frequency analysis method based on wavelet transform. Considering the abrupt change and compact support characteristics of water hammer waves, this embodiment preferably uses the Daubechies wavelet system (e.g., db4 or db8) as the mother wavelet function. The fluid-structure interaction transient pressure response signal is then decomposed into multiple scales using the selected mother wavelet function. Wavelet transform can simultaneously provide localization information of the signal in both the time and frequency domains, making it particularly suitable for analyzing non-stationary transient signals.

[0058] This process employs a time-frequency analysis method based on wavelet transform. This method can analyze the original pressure signal... The signal is decomposed into the time-frequency plane, generating a set of wavelet coefficients. These coefficients visually demonstrate the distribution of signal energy at different times and frequencies.

[0059] The wavelet coefficient matrix obtained through calculation Different signal components can be clearly distinguished on the time-frequency plane. The transient water hammer wave signal excited by the inverse perturbation coding subsystem is represented on the time-frequency diagram as occurring at a specific time. The noise appears as a localized high-energy region with energy concentrated in a specific frequency band (determined by the characteristics of water hammer waves). Periodic noise from mechanical equipment such as water pumps, on the other hand, manifests as a horizontal bright band spanning the entire time axis, with energy concentrated at a few fixed frequencies (fundamental frequency and harmonics). High-frequency electronic noise, however, diffuses throughout the entire high-frequency region, with a lower energy amplitude.

[0060] The denoising process is based on wavelet domain threshold filtering. The system sets an energy threshold and applies it to the wavelet coefficient matrix. Coefficients with amplitudes below this threshold are set to zero, while high-energy coefficients that match the characteristics of transient pressure waves are retained. A clean pressure response signal is reconstructed by performing an inverse wavelet transform on the filtered wavelet coefficients. This signal has filtered out most of the steady-state mechanical noise and random electronic noise, retaining the core dynamic pressure response information excited by the distributed valves and modulated by the pipe-soil coupling system.

[0061] The MISO signal decoupling and localization algorithm primarily addresses the signal aliasing problem in multi-input single-output systems. Since multiple solenoid valves act as signal sources in field pipelines, when these valves operate concurrently according to their respective pseudo-random binary sequences (PRBS), the denoised signal received by the head sensor... It is a linear superposition of responses from multiple independent signal sources in space and time. In order to obtain soil feedback information for a specific geographical location (a specific valve control area), the response component corresponding to each individual valve source must be separated from the total superimposed signal.

[0062] The core of the algorithm utilizes cross-correlation analysis techniques. It calculates the total response signal received by the sensor. With each specific valve Preset action sequence Different time delays The correlation under given conditions yields a cross-correlation function. Due to the quasi-orthogonality of the input sequences, this function only applies at a specific time delay. A significant peak is observed at this point, which corresponds to the pressure wave originating from the [missing information]. The propagation time from the valve to the sensor.

[0063] By examining each valve ( The algorithm performs the cross-correlation operation described above, decomposing the single aliased total signal into independent sub-channel response functions. Each sub-channel response function actually represents the system impulse response function along a specific pipeline path.

[0064] After completing channel separation, the algorithm further performs reference drift correction. This is because the gas content of the fluid is actively changed. This causes a shift in the global reference wave velocity. To extract the local wave velocity disturbance caused purely by changes in soil modulus, the theoretical reference wave velocity calculated using the aforementioned Wood equation needs to be used, and the global delay component introduced by the air bubble needs to be subtracted from the measured travel time. After this correction step, the final separated feature signal contains only wave velocity distortion information caused by changes in the constraint stiffness of the soil outside the pipe, achieving precise positioning and feature decoupling of the soil physical state in different irrigation areas.

[0065] Reference Appendix Figure 2 The Physical Information Spatiotemporal Inversion Module is responsible for transforming the denoised and decoupled signal features into a distribution of physical parameters across the entire pipeline network. This module utilizes a computational framework combining deep learning and physical laws to reconstruct continuous fluid dynamic states and soil mechanical parameters from sparse boundary observation data. This module mainly comprises three parts: the Physical Information Neural Network (PINN) architecture design, the embedding of physical constraint equations, and the construction and optimization strategies for the loss function.

[0066] The Physical Information Neural Network (PINN) architecture was designed to construct a deep neural network model for approximating spatiotemporal physical fields. Unlike traditional black-box neural networks that rely solely on large amounts of labeled data, the PINN architecture explicitly incorporates the governing equations of the physical fields.

[0067] The core of the network is a fully connected deep feedforward neural network, denoted as... The network's input layer receives spatiotemporal coordinate vectors. and the control parameters for experimental excitation (such as the current gas content) (or reference wave velocity). To handle complex nonlinear mappings, the network contains multiple hidden layers (e.g., 5 to 10 layers), each containing tens to hundreds of neurons (e.g., 50 to 100 neurons). Each neuron employs a nonlinear activation function, such as the hyperbolic tangent function (tanh) or the Swish function, to ensure sufficient continuity of higher-order derivatives, which is a prerequisite for subsequent calculation of physical partial derivatives.

[0068] The network's output layer is designed to directly predict the physical quantities of interest. For an irrigation network system, the output layer contains three main physical variables: Shuitouchang : Corresponds to the pressure distribution of fluid within the pipeline network.

[0069] Flow field : Corresponds to the flow distribution of fluid within the pipeline network.

[0070] Wave speed field : Corresponds to the system wave velocity distributed along the pipeline.

[0071] In particular, wave velocity field In the time dimension, it is usually considered a quasi-static variable (remaining unchanged during a single transient probe), but in the spatial dimension... The above is subject to change. Therefore, a subnetwork or independent branch is typically designed in the network architecture specifically for prediction. This branch is based solely on spatial coordinates. As input.

[0072] To compute the physical constraints, the PINN architecture utilizes automatic differentiation. Unlike numerical difference, automatic differentiation uses the chain rule to precisely calculate the partial derivatives of the network output with respect to the input coordinates within the range of computer precision. For example... , These partial derivative terms do not directly participate in the numerical prediction of forward propagation, but are fed as intermediate variables into the subsequent physical constraint module to construct the physical residual equation, thereby forcing the network output to not only conform to the training data, but also to more strictly follow the physical laws of fluid dynamics.

[0073] The embedding of physical constraint equations is the core mechanism that distinguishes Physical Information Neural Networks (PINNs) from traditional data-driven neural networks. This process encodes prior physical knowledge of fluid dynamics and fluid-structure interaction into the training process of the neural network, ensuring that the physical fields predicted by the network are consistent in space and time and conform to physical laws.

[0074] The embedding process is achieved by constructing physical residual terms. Using automatic differentiation techniques, the physical fields predicted by the neural network can be accurately calculated. and Relative to time and space The partial derivatives are then substituted into the known governing equations for transient fluid flow. For transient flow in a pipe, the core governing equations are the one-dimensional Saint-Venant equations, specifically including: Continuity Equation: This equation describes the law of conservation of fluid mass. In transient flow in a pipe, considering the elasticity of the pipe wall and the compressibility of the fluid, the equation expresses the relationship between the rate of change of head and the rate of change of flow space. Substituting the physical quantities predicted by the network, the resulting residual term of the continuity equation... Defined as: ; Momentum Equation: This equation describes the law of conservation of fluid momentum, taking into account the balance between pressure gradient, inertial force, and frictional resistance. Residual terms of the momentum equation. Defined as: ; in, It is the acceleration due to gravity. The cross-sectional area of ​​the pipe. This is the friction coefficient of Darcy Weisbach. Let be the pipe diameter. These are all known physical or geometric parameters. Specifically, the system wave velocity... As an output of the neural network, it is directly coupled to these two physical equations.

[0075] In addition to the fluid dynamics equations, this module also embeds the established fluid-structure interaction constitutive relations as physical constraints. The system wave velocity field predicted by the neural network... The previously defined pipe-soil coupled wave velocity model must be satisfied. The network predictions... Substituting into the wave velocity model equation and rearranging, we can obtain a value for the unknown soil modulus. The expression for this constraint allows the neural network to not only learn fluid dynamics but also simultaneously invert the soil elastic modulus distributed along the pipeline. This enables the inversion of physical parameters from fluid fields to solid media.

[0076] By using the residual terms of these physical equations and As a penalty term added to the network's training objective, the network is forced during optimization to find a set of weights and biases that not only fit the observed data in its predicted physics but also make these physical residuals approach zero across the entire computational domain. This is equivalent to applying strong physical regularization to the solution space, guiding the network to converge to a physically valid solution, enabling it to make reasonable physical inferences even in regions without observed data.

[0077] The loss function construction and optimization strategy establishes the training objective and parameter update mechanism of Physical Information Neural Network (PINN), mainly including the mathematical definition of the composite loss function and the gradient-based hybrid optimization algorithm.

[0078] Composite loss function It is a quantitative metric that measures the difference between network predictions and actual physical states. To balance the accuracy of the fit to the observed data with the degree of adherence to physical laws, the loss function is constructed as a weighted combination of data error terms and physical residual terms: ; In the formula, the observation loss term Defined at the physical boundary location of the head sensor (e.g.) =0). The network's output prediction at the boundary (stress). With traffic ) and the actual time series data collected by high-frequency sensors ( and The mean squared error (MSE) between the observations constitutes a data-driven constraint. This term ensures that the inversion results are consistent with physical facts at the observation points.

[0079] The physical loss term is defined within the spatiotemporal computational domain of the entire pipeline. A large number of collocation points are randomly selected within the computational domain, and the residuals of the continuity equation and momentum equation at these points are calculated using the aforementioned automatic differentiation technique. The physical loss term is the mean of the squares of these residuals. Minimizing this term forces the network-predicted fluid state to strictly adhere to the laws of fluid dynamics conservation even in the unobserved region (i.e., deep within the pipeline field), thus achieving a mathematical extension of the head boundary information to the global physical field.

[0080] The optimization strategy employs a multi-stage hybrid gradient descent scheme to solve the aforementioned non-convex optimization problem. In the initial training phase, the Adam optimizer (Adaptive Moment Estimator) is activated, utilizing its momentum mechanism to quickly traverse the flat regions of the loss function and avoid getting trapped in shallow local minima. When the loss function decreases to a preset threshold or the gradient change becomes gradual, the system automatically switches to the LBFGS optimizer (Limited-Memory Quasi-Newton Method). LBFGS uses approximate information from the second-order Hessian matrix for parameter updates, achieving higher convergence accuracy and enabling precise fine-tuning of network weights to meet stringent physical constraints.

[0081] To balance the gradient competition between data fitting and physical constraints, a dynamic weight adaptive algorithm is used to adjust the coefficients. and The algorithm monitors the norm distribution of the gradients of each loss function in real time, dynamically amplifying the weights of terms with smaller gradients to prevent ill-conditioned convergence during training, where the network either only fits the data and ignores the physical equations, or only satisfies the equations and deviates from the observed data. After training, the network output... This branch presents the spatial function of the wave velocity distribution across the entire pipe network. Using the aforementioned fluid-structure interaction wave velocity model, the wave velocity distribution... Algebraic mapping back to soil modulus distribution Ultimately, this will enable non-contact, full-domain inversion of the physical state of the soil along the field pipeline network.

[0082] The closed-loop decision-making and precision execution module, serving as the system's application terminal and control center, is responsible for transforming the physical parameters obtained from the preceding modules into a visualized agronomic state map, and generating automated irrigation execution commands accordingly. This module mainly comprises two parts: soil state parameter mapping and reconstruction, and spatiotemporal variable precision irrigation strategies.

[0083] The soil state parameter mapping and reconstruction mechanism aims to solve the semantic conversion from mechanical parameters to agronomic parameters, as well as the spatial reconstruction problem from one-dimensional pipeline data to two-dimensional field plane.

[0084] In the preceding module, the Physical Information Neural Network (PINN) outputs the soil elastic modulus distributed along the pipeline. Although this parameter is physically precise, it is difficult to directly guide agricultural production. Therefore, the system first performs a parameter mapping process, using the inverse function of the constitutive relation in soil physics to map the elastic modulus. Converted to volumetric water content and soil compaction Key agricultural indicators, etc.

[0085] The mapping process is based on a pre-calibrated regional soil transformation model. According to the effective stress principle, there is a monotonically negative correlation between soil modulus and water content; the system calls the inverse function of this relationship. Perform point-by-point calculations: ; In this relation A vector of soil texture parameters (including clay content, sand content, and organic matter content) represents a specific plot of land, and these parameters determine the specific shape of the transformation curve. To eliminate the deviation between the theoretical model and the actual environment, the system introduces an anchor point correction mechanism. A small number of in-situ pin-type soil sensors are deployed at key locations in the field (e.g., one per 10 acres) as real reference points. The system compares the PINN inversion values ​​with the sensor measured values ​​in real time to calculate the global deviation correction coefficient. This coefficient is used to dynamically calibrate the global inversion results, thus combining the wide coverage advantage of distributed inversion with the high precision advantage of point sensing.

[0086] After completing the parameter mapping along the pipeline, the system obtained a set of one-dimensional high-resolution moisture content data streams distributed along the irrigation branch pipes. To construct a two-dimensional distribution map of the entire field, the system combined the spatial topology of the pipeline network recorded by a Geographic Information System (GIS) and employed the Kriging interpolation algorithm from geostatistics. The Kriging algorithm utilizes the spatial autocorrelation (semi-variogram) of soil properties, using data from the locations of the irrigation pipelines as a framework, to perform optimal linear unbiased estimation of soil moisture content in the gaps between the pipelines.

[0087] Through the above processing, the discrete pipeline network sensing data is reconstructed into a continuous, high-resolution field digital twin heat map. This heat map visually presents the drought distribution of the entire field with pixel-level precision (e.g., a 1m x 1m grid), and can clearly identify local drought patches caused by topographic relief, uneven soil texture, or differences in crop water consumption, providing a digital base map for subsequent variable irrigation decisions.

[0088] The intelligent decision control logic is an algorithmic entity that enables the transformation from passive irrigation to precise on-demand irrigation. It mainly includes two parts: a water demand decision mechanism based on spatial aggregation and an asynchronous execution strategy under hydraulic constraints.

[0089] The spatial aggregation-based water demand decision mechanism is responsible for converting high-resolution soil moisture content heatmaps into discretized irrigation commands for specific valve control units. Although the soil state map reconstructed by the preceding modules has extremely high spatial resolution (e.g., pixel-level), the execution granularity of the physical irrigation system is limited by the topological distribution of the solenoid valve array. Therefore, the logic unit first performs spatial aggregation operations.

[0090] The system defines the hydraulic control zone for each solenoid valve unit based on the physical layout of the field irrigation network. The boundary of the control zone is determined by the effective coverage of adjacent drip tapes or sprinklers. For each hydraulic control zone... The system extracts all soil moisture content pixel data covering the area and combines it with a target moisture content threshold corrected by the crop coefficient. The algorithm calculates the weighted average water deficit for the region. During this process, a non-uniform weighting factor is introduced, assigning higher decision weights to pixels located at the edge of the control domain or in areas with dense crop root systems, to prevent severe local drought from being masked by the average value.

[0091] Based on the calculated regional water deficit and the planned wetting depth of the crop root zone, the system generates a variable irrigation prescription map. The prescription map does not directly output water volume; instead, it converts the water demand into the corresponding solenoid valve opening duration. The conversion formula is based on the orifice outflow formula in fluid mechanics, combined with the measured water pressure of the current main pipeline. and valve flow coefficient Perform dynamic calculations: ; In the formula, For the first The control area of ​​each valve The depth of the root layer. This represents the current average moisture content of the region. Through this mechanism, the system generates independent, second-precise opening duration commands for each valve across the entire field, achieving differentiated water distribution for each valve.

[0092] The asynchronous execution strategy under hydraulic constraints is responsible for transforming the generated irrigation prescription map into a physically executable time-series control queue. Due to the limited capacity of the water supply pumping stations at the irrigation head and the maximum allowable flow limit in the pipeline design, opening all valves that require irrigation at the same time would cause a sharp drop in pipeline pressure, disrupting irrigation uniformity and even triggering water hammer accidents.

[0093] To address this multi-constraint scheduling problem, the control logic employs a dynamic time-slicing and pressure-balancing algorithm. The algorithm first reads the rated water supply flow rate of the head pump station. The system then places all valves to be opened into an execution pool and sorts them according to priority (determined by the degree of water shortage).

[0094] The scheduler employs a greedy or genetic algorithm to divide the tasks in the execution pool into several temporally consecutive and spatially independent rotational irrigation groups. When constructing these rotational irrigation groups, the algorithm strictly adheres to flow balance constraints, ensuring that the total flow of all open valves within each group is maintained. The flow rate shall not exceed the rated water supply capacity of the pumping station, nor be lower than the flow rate threshold corresponding to the minimum system pressure required to maintain the normal operation of the drip irrigation system.

[0095] Once the irrigation rotation sequence is determined, the main control computer sends opening and closing commands to the distributed solenoid valves in the field via fieldbus or wireless communication network. Unlike the high-frequency micro-vibration in detection mode, in irrigation execution mode, the solenoid valves perform a fully open action and maintain this position for the calculated duration. Furthermore, the system employs staggered start-stop logic. During the switching intervals between the end of one irrigation cycle and the start of the next, millisecond-level time differences are used to gradually close old valves and open new ones, smoothing pressure fluctuations within the pipeline and preventing damage to pipe fittings from transient water hammer effects. After a round of precise variable irrigation is completed, the system automatically enters a static infiltration period, waiting for moisture to redistribute and reach equilibrium in the soil. Subsequently, the system triggers the fluid excitation sequence again, initiating a new round of excitation-sensing inversion loop to verify irrigation effectiveness and adjust decision parameters for the next cycle, thus forming a complete closed-loop intelligent control circuit.

Claims

1. A precision irrigation system based on machine learning, characterized in that, include: Field pipeline systems, which consist of flexible pipes with elastic deformation capabilities, serve as physical carriers for transporting fluids and sensing the soil environment. A multi-source heterogeneous fluid excitation module is used to physically connect with the head of the field pipeline system, generate variable modulus fluid and generate coded wave source, and actively inject transient pressure waves into the field pipeline system. A fluid-structure interaction response acquisition module is installed at the head of the field pipeline system to acquire the fluid-structure interaction transient pressure response signal after the transient pressure wave is modulated by the soil constraint state along the field pipeline system. The transient signal feature extraction and decoupling module is used to receive the fluid-structure interaction transient pressure response signal and perform noise reduction processing to decouple the independent channel response features of a specific valve position from the fluid-structure interaction transient pressure response signal. The physical information spatiotemporal inversion module is used to receive the decoupled response features, construct a physical information neural network to invert the system wave velocity distribution of the entire field pipeline system, and derive the soil elastic modulus distributed along the pipeline. The closed-loop decision-making and precise execution module is used to receive the soil elastic modulus obtained by inversion and reconstruct it into agronomic state parameters. Based on the agronomic state parameters, it generates control commands and feeds them back to the multi-source heterogeneous fluid excitation module to execute irrigation or anomaly handling tasks.

2. The precision irrigation system based on machine learning according to claim 1, characterized in that, The multi-source heterogeneous fluid excitation module includes a gas-liquid two-phase rheological modulus control subsystem and a distributed valve-controlled inverse disturbance coding subsystem. The gas-liquid two-phase rheological modulus control subsystem is used to adjust the gas content in the fluid to set different detection reference wave velocities; the distributed valve-controlled reverse disturbance coding subsystem is used to control the valve action on the branch pipe of the field pipeline system according to the pseudo-random binary sequence to generate a water hammer pulse signal with orthogonality.

3. The precision irrigation system based on machine learning according to claim 1, characterized in that, The fluid-structure interaction response acquisition module operates based on the fluid-structure interaction sensing mechanism of flexible pipelines. In the field pipeline system, the flexible pipes undergo radial deformation under transient pressure, and the radial deformation is physically constrained by the elastic modulus of the soil outside the pipe, which determines the propagation speed of the transient pressure wave in the pipe. Based on the correlation between the propagation speed and the physical constraints, a quantitative mapping relationship between soil confining pressure, pipe wall thickness, pipe material modulus and system wave velocity is established using the system equivalent wave velocity formula that includes soil elastic modulus variables, thereby transforming changes in soil physical state into changes in wave velocity.

4. The precision irrigation system based on machine learning according to claim 1, characterized in that, The transient signal feature extraction and decoupling module employs a time-frequency analysis method based on continuous wavelet transform when performing denoising processing, specifically including: The fluid-structure interaction transient pressure response signal is decomposed into multiple scales by selecting a mother wavelet function that matches the characteristics of the signal. Steady-state mechanical noise and high-frequency electronic noise are filtered out using wavelet domain threshold filtering technology. Finally, a pure pressure response signal that retains the transient waveform characteristics is reconstructed by inverse wavelet transform.

5. The precision irrigation system based on machine learning according to claim 1, characterized in that, The transient signal feature extraction and decoupling module employs the MISO signal decoupling and localization algorithm when decoupling independent channel response features, specifically including: Using cross-correlation analysis, the cross-correlation function between the fluid-structure interaction transient pressure response signal and the preset pseudo-random binary encoding sequence of each valve is calculated; By identifying the peak delay of the cross-correlation function, the sub-channel response corresponding to a specific valve position is separated, and the global transmission delay introduced by the gas content is corrected based on the theoretical reference wave velocity.

6. The machine learning-based precision irrigation system according to claim 1, characterized in that, The physical information neural network in the physical information spatiotemporal inversion module adopts a fully connected deep feedforward architecture, specifically including: Establish a nonlinear mapping relationship from the spatiotemporal coordinate vector of the input layer to the fluid head field, flow field, and system wave velocity field of the output layer; Based on the aforementioned nonlinear mapping relationship, the partial derivatives of the output layer variables with respect to the input layer variables are calculated using an automatic differentiation mechanism, and these partial derivatives are used as the basis for constructing physical constraint terms to constrain the training of the physical information neural network.

7. The precision irrigation system based on machine learning according to claim 6, characterized in that, The physical information spatiotemporal inversion module embeds physical constraint equations during the training process of the physical information neural network. The physical constraint equations include a continuity equation describing the conservation of mass in transient flow, a momentum equation describing the conservation of momentum, and a constitutive equation describing the nonlinear mapping relationship between system wave velocity and soil elastic modulus. The goal of network training is to minimize the weighted sum of the fitting error of boundary observation data and the residual of the global physical equation.

8. The precision irrigation system based on machine learning according to claim 7, characterized in that, The physical information spatiotemporal inversion module adopts a composite loss function and a hybrid optimization strategy; The composite loss function is composed of a weighted average of the fitting error and the physical residual; The hybrid optimization strategy is used to achieve fast convergence using an adaptive moment estimation optimizer in the early stages of training, and to switch to a finite-memory quasi-Newton optimizer for fine-grained search in the later stages of training. The weight coefficients of each loss term are dynamically adjusted based on the statistical characteristics of the gradient during training.

9. The machine learning-based precision irrigation system according to claim 7, characterized in that, The closed-loop decision-making and precise execution module reconstructs the soil elastic modulus into agronomic state parameters, including the following steps: Using the inverse function of the soil constitutive relation, the soil elastic modulus obtained by inversion is mapped to the volumetric water content, and then corrected by combining the measured data of the in-situ anchor point sensor. Furthermore, by utilizing the Kriging interpolation algorithm in conjunction with the topology of the field pipeline system, the one-dimensional parameters distributed along the pipeline were reconstructed into a two-dimensional humidity distribution map covering the entire area.

10. The machine learning-based precision irrigation system according to claim 7, characterized in that, The closed-loop decision-making and precise execution module also executes intelligent decision-making control logic, specifically including: The zoning decision algorithm calculates the water demand of each irrigation zone based on the two-dimensional humidity distribution map, and converts the water demand into valve opening time based on fluid dynamics principles. A hydraulically constrained scheduling strategy is used to divide all initiated tasks into several irrigation groups using a dynamic time-slicing algorithm. An abnormal response mechanism is used to generate control commands to drive the multi-source heterogeneous fluid excitation module to start the pure bubble oxygenation mode when an abnormal modulus corresponding to soil compaction or cavities is detected.

Images