An agricultural big data water and fertilizer demand prediction method and system

Abstract

The application discloses an agricultural big data water and fertilizer demand prediction method and system, relates to the technical field of intelligent irrigation, and aims at the problems that the prior art is difficult to extract long tail parameters and the model precision is insufficient when medium distortion and pipe network disturbance occur.The application collects electric water pressure data, combines pre-stored transpose matrices to obtain an abnormal sensitivity vector, calculates a differential absolute value to extract a distortion gradient, and maps the distortion gradient into a fractional order order representing water characteristics; a double-level observer is used to decouple pipe network mechanical disturbance to extract soil back absorption equivalent pressure; and the order and the pressure are used to enhance the physical characteristics of time series data and construct a fractional order network output prediction quantity.The application can efficiently detect underground preferential leakage and greatly improve the prediction accuracy under poor working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent irrigation technology, and in particular to a method and system for predicting water and fertilizer demand in agriculture using big data. Background Technology

[0002] In agricultural production, field water and fertilizer management is a key link affecting crop growth quality and yield. With the advancement of agricultural modernization, the use of agricultural big data and intelligent irrigation technology for water and fertilizer demand prediction is of great significance for improving water and fertilizer utilization efficiency and reducing agricultural production costs. However, the field environment is complex and changeable. The movement of water and nutrients is not only affected by macro factors such as weather and crop canopy, but also closely related to the deep physical structure of the soil and the hydrodynamic characteristics of the pipe network. Patent CN120561557B discloses an intelligent field water and fertilizer control method based on the Internet of Things. It collects images of the crops to be controlled in the field, collects specified water and fertilizer data and weather forecast data of the current field soil, uses a cloud server to identify crop types and their growth stages, and then calls the basic water and fertilizer parameter library of the corresponding crops to generate an initial water and fertilizer control plan. At the same time, it creates a soil water and fertilizer prediction model to predict the trend of change and formulates a dynamic water and fertilizer control plan in combination with the growth cycle characteristics. Current technologies typically rely solely on point sensors to collect macroscopic soil moisture and nutrient data, assuming the soil is a homogeneous medium. This approach fails to effectively detect and quantify changes in the soil's internal topology. When soil develops cracks or large pores due to alternating wet and dry conditions, saline fertilizers can rapidly leak into deeper layers along these channels, resulting in severe preferential seepage. Existing methods cannot detect this microscopic distortion, leading to a lack of crucial physical boundary conditions in the predictive input data. Furthermore, existing predictive models often employ integer-order deep learning architectures such as Long Short-Term Memory (LSTM) networks. These purely data-driven models lack constraints on underlying physical processes and cannot accurately characterize the development of water and fertilizer. The strong capillary hysteresis and non-Fick long-tail memory effect during transport in cracked and deteriorated soils make it difficult for this fixed memory mechanism to adaptively respond to dynamic dynamic characteristics under different leakage severity levels. Existing technologies lack the ability to dynamically decouple the pipeline operation status and soil feedback resistance. They cannot extract the soil backflow equivalent pressure, which characterizes low-frequency dynamic features, from the high-frequency mechanical noise mixed with pump motor pulsation and pipeline resonance. Therefore, it cannot be introduced as a key physical constraint into the time series prediction of big data, ultimately resulting in insufficient accuracy of water and fertilizer demand prediction and often serious control lag when facing sudden water and fertilizer leakage. Summary of the Invention

[0003] The technical problem solved by this invention is that existing technologies are unable to detect medium topological anomalies and extract physical constraint parameters characterizing the long-tail memory effect of water transport when soil internal structure is distorted, such as when crack development leads to preferential seepage and irrigation networks are subject to complex mechanical disturbances. Furthermore, traditional pure data-driven models based on integer order cannot adaptively adjust the memory weight of historical water loss to current demand, resulting in insufficient accuracy in predicting water and fertilizer demand under complex field conditions and the inability to effectively and timely block deep leakage.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for predicting water and fertilizer demand in agriculture using big data, comprising the following steps: Step S1: Collect voltage and water pressure data; Step S2: Based on the voltage data and the pre-stored sensitivity transpose matrix, obtain the abnormal distribution sensitivity vector; Step S3: Arrange the abnormal distribution sensitivity vector according to the three-dimensional coordinates of the spatial nodes, calculate the sum of the absolute differences of the abnormal distribution sensitivity values ​​of adjacent nodes along the depth direction as the total sensitivity change of each vertical column, and take the maximum value of the total sensitivity change of all vertical columns as the distortion gradient. Step S4: Based on the distortion gradient, calculate the fractional order by constructing a monotonic mapping from the distortion gradient to the fractional order. Step S5: Input the water pressure data into the first observer to obtain the first water pressure, input the first water pressure into the second observer to obtain the soil reabsorption equivalent pressure, and the integration order of the second observer is determined by the fractional order; calculate the target negative pressure peak based on the soil reabsorption equivalent pressure; Step S6: Using the fractional order and the equivalent soil reabsorption pressure, construct and train a fractional-order long short-term memory network prediction model to output the predicted water and fertilizer demand for the next cycle.

[0005] Further, step S1 includes the following sub-steps: Step S101: Collect voltage data through a probe array and construct a voltage matrix; Step S102: Collect water pressure data using a pressure sensor.

[0006] Further, step S2 includes the following sub-steps: Step S201: Stretch the effective elements of the differential voltage matrix after removing the diagonal elements into a column vector, which serves as the potential anomaly vector. Step S202: Multiply the pre-stored sensitivity transpose matrix with the potential anomaly vector to obtain the anomaly feature distribution vector; Step S203: Perform a Hadamard product operation on the abnormal feature distribution vector and the pre-stored static reference potential vector, and perform median filtering on the operation result to obtain the abnormal distribution sensitivity vector.

[0007] Further, step S3 includes the following sub-steps: Step S301: Based on the pre-stored index coordinate mapping table, arrange each element in the abnormal distribution sensitivity vector in association with the three-dimensional coordinates of the spatial nodes. Step S302: Divide the detection area into multiple vertical columns according to the vertical coordinates. For each vertical column, calculate the sum of the absolute differences of the sensitivity values ​​of the abnormal distribution of adjacent nodes along the depth direction, which is used as the total sensitivity change of the column. Step S303: Take the maximum value among the total sensitivity changes of all vertical columns as the distortion gradient.

[0008] Further, step S4 includes the following sub-steps: Step S401: Construct a monotonic mapping relationship from the distorted gradient to the fractional order; Step S402, and calculate the fractional order corresponding to the distortion gradient based on the monotonic mapping relationship, wherein the value of the fractional order is in the interval (0,1).

[0009] Further, step S5 includes the following sub-steps: Step S501: Input the water pressure data into the first observer and output the first water pressure; The first observer adopts an integer-order extended state observer structure; In each water pressure control cycle, the first observer calculates the state variable value for the next cycle based on the water pressure data collected at the current moment, the state variable value of the previous cycle, and the valve control quantity at the current moment through the discrete-time state update equation.

[0010] Furthermore, step S5 also includes the following sub-steps: Step S502: Input the first water pressure into the second observer and output the soil backflow equivalent pressure. The second observer adopts a fractional-order expansion state observer structure.

[0011] Furthermore, step S5 also includes the following sub-steps: Step S503: Calculate the negative pressure change based on the soil backflow equivalent pressure, and obtain the target negative pressure peak value by combining it with the steady-state operation benchmark pressure of the pipeline network; The change in negative pressure is obtained by adding the water hammer pressure term, the gravitational potential energy term, and the soil reabsorption equivalent pressure. The target negative pressure peak value is the difference between the steady-state operating reference pressure of the pipeline network and the negative pressure change.

[0012] Step S504: Calculate the valve control quantity based on the target negative pressure peak value and the state estimate value of the second observer.

[0013] Further, step S6 includes the following sub-steps: Step S601: Collect meteorological data, crop canopy data, and historical irrigation data to construct a multidimensional agricultural big data time series. Step S602: The fractional order is used as the index of the memory forgetting factor and embedded into the cell state update equation of the long short-term memory network to construct a fractional order long short-term memory network model. Step S603: The equivalent pressure of soil reabsorption is added as a physical constraint term to the loss function of the fractional-order long short-term memory network model; Step S604: Input the multidimensional agricultural big data time series into the trained fractional-order long short-term memory network model and output the predicted water and fertilizer demand for the next period.

[0014] An agricultural big data water and fertilizer demand forecasting system includes: The data acquisition module is used to collect voltage and water pressure data. An anomaly detection module is used to obtain an anomaly distribution sensitivity vector based on the voltage data and a pre-stored sensitivity transpose matrix. The gradient calculation module is used to arrange the abnormal distribution sensitivity vector according to the three-dimensional coordinates of the spatial nodes, calculate the sum of the absolute differences of the abnormal distribution sensitivity values ​​of adjacent nodes along the depth direction as the total sensitivity change of each vertical column, and take the maximum value of the total sensitivity change of all vertical columns as the distortion gradient. The order mapping module is used to calculate the fractional order based on the distortion gradient by constructing a monotonic mapping from the distortion gradient to the fractional order. The pressure estimation module is used to input the water pressure data into a first observer to obtain a first water pressure, input the first water pressure into a second observer to obtain the soil reabsorption equivalent pressure, wherein the integration order of the second observer is determined by the fractional order; and calculate the target negative pressure peak based on the soil reabsorption equivalent pressure. The demand forecasting module is used to construct and train a fractional-order long short-term memory network prediction model based on the fractional order and the equivalent soil reabsorption pressure, and output the predicted water and fertilizer demand for the next cycle.

[0015] The beneficial effects of this invention are as follows: This invention constructs a two-order extended state observer architecture, which utilizes an integer-order observer without historical integral memory to quickly absorb high-frequency mechanical noise such as water pump pulsation, and uses a fractional-order observer approximated by an Oustaloup filter to accurately extract the soil backflow equivalent pressure with low-frequency long-tail memory characteristics. This invention directly embeds the fractional-order obtained by physical detection into the cell state update equation of the long short-term memory network as a derivative operator, so that the model can adaptively adjust the memory weight of historical irrigation and water loss to the current prediction according to the current dynamic development degree of soil cracks. This invention incorporates the extracted soil backflow equivalent pressure as a physical constraint term into the loss function, and applies additional penalties to the prediction bias that ignores soil deterioration and water retention capacity decline during model training. Attached Figure Description

[0016] Figure 1 The flowchart illustrates the steps of an agricultural big data water and fertilizer demand prediction method according to an embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1, referring to Figure 1 This paper provides a method for predicting agricultural water and fertilizer demand using big data, which includes the following steps: Step S1: Collect voltage and water pressure data; Step S2: Based on the voltage data and the pre-stored sensitivity transpose matrix, obtain the abnormal distribution sensitivity vector; Step S3: Arrange the anomaly distribution sensitivity vector according to the three-dimensional coordinates of the spatial nodes, calculate the sum of the absolute differences of the anomaly distribution sensitivity values ​​of adjacent nodes along the depth direction as the total sensitivity change of each vertical column, and take the maximum value of the total sensitivity change of all vertical columns as the distortion gradient. Step S4: Based on the distortion gradient, calculate the fractional order by constructing a monotonic mapping from the distortion gradient to the fractional order. Step S5: Input the water pressure data into the first observer to obtain the first water pressure, input the first water pressure into the second observer to obtain the soil reabsorption equivalent pressure, and the integration order of the second observer is determined by the fractional order; calculate the target negative pressure peak based on the soil reabsorption equivalent pressure. Step S6: Construct and train a fractional-order long short-term memory network prediction model using fractional order and combined with soil reabsorption equivalent pressure, and output the predicted water and fertilizer demand for the next cycle.

[0019] This invention efficiently maps boundary electric field anomalies into the distortion gradient of three-dimensional spatial nodes, enabling the quantification of the overall leakage capacity of large pore channels within the brief window of priority leakage. It utilizes a dual-level observer to decouple complex disturbances in the pipeline network and extract the soil backflow equivalent pressure characterizing low-frequency dynamic properties, providing a crucial physical boundary input for long short-term memory networks. By introducing fractional-order dynamic operators into cell state updates and implementing physical loss constraints, it can adaptively adjust the memory weights of historical loss data, significantly improving the robustness and accuracy of water and fertilizer demand prediction for the next cycle under soil structure damage scenarios.

[0020] Step S1 includes the following sub-steps: Step S101: Collect voltage data through a probe array and construct a voltage matrix; Step S102: Collect water pressure data using a pressure sensor.

[0021] In a specific embodiment of the present invention, voltage data is acquired by a micro-impedance probe array, which consists of 16 304 stainless steel probes arranged at equal intervals in a 4×4 grid, and is horizontally buried at a depth of 30cm below the ground surface. The probes are connected by two 16-channel analog multiplexers, and the adjacent electrode excitation-measurement mode is adopted: two adjacent probes are selected in sequence as current injection electrode pairs, and an alternating current with a frequency of 10kHz and an amplitude of 1mA is applied; under each pair of excitation electrodes, the voltage amplitude between the remaining adjacent probe pairs is measured in sequence. The 16 probes form a total of 16 adjacent excitation electrode pairs, and 13 adjacent voltage measurements can be performed under each excitation pair, for a total of 208 voltage measurement values. Within each 50ms detection cycle, five samples are acquired at a frequency of 100Hz. The average value is then used to construct a 16×16 voltage matrix, with units in millivolts. The row numbers of the voltage matrix correspond to the excitation electrode pair numbers, and the column numbers correspond to the measurement electrode pair numbers. The element in the i-th row and j-th column of the matrix represents the voltage amplitude measured by the j-th measurement electrode pair under the i-th excitation electrode pair. For positions that are not actually measured, i.e., i=j and non-adjacent electrode pair positions, the matrix elements are set to 0. Perform min-max normalization on each element of the voltage matrix to linearly map the voltage magnitude to the dimensionless interval [0,1]. Water pressure data is acquired through a piezoresistive pressure transmitter, which is installed on the drip irrigation trunk line downstream of the variable frequency solenoid main control valve. The transmitter continuously acquires instantaneous pressure values ​​at a frequency of 1000Hz, obtaining one pressure data point within 1ms of each water pressure control cycle, with the unit being kilopascals.

[0022] Step S2 includes the following sub-steps: Step S201: Stretch the effective elements of the differential voltage matrix after removing the diagonal elements into a column vector, which serves as the potential anomaly vector; Step S202: Multiply the pre-stored sensitivity transpose matrix with the potential anomaly vector to obtain the anomaly feature distribution vector; Step S203: Perform a Hadamard product operation on the abnormal feature distribution vector and the pre-stored static reference potential vector, and perform median filtering on the operation result to obtain the abnormal distribution sensitivity vector.

[0023] In one specific embodiment of the present invention, a differential voltage matrix is ​​calculated based on a voltage matrix and a reference voltage matrix. The elements in the differential voltage matrix are obtained by subtracting the elements in the voltage matrix from the elements in the reference voltage matrix element by element. The voltage matrix, the reference voltage matrix, and the differential voltage matrix have the same dimensions. The reference voltage matrix is ​​collected and stored during the initialization phase. Two hours after the end of full irrigation, when the soil is uniformly moist and there is no preferential leakage, the voltage matrix of 10 detection cycles is collected in the same way as in step S1. The average value is taken as the reference voltage matrix. The dimension of the reference voltage matrix is ​​16×16 and the unit is the same as that of the voltage matrix, which is mV. The process of obtaining the differential voltage matrix is ​​represented as follows: under uniform and crack-free soil conditions, the potential distribution between each electrode is relatively stable. When salt-containing fertilizer rapidly infiltrates along large pores or cracks, the conductivity of the local area increases sharply, causing the measured potential of the corresponding electrode pair to drop instantly. The non-zero elements in the differential voltage matrix essentially reflect the local potential collapse caused by preferential leakage, thus removing the influence of the soil medium. Preferred leakage refers to the phenomenon that water and fertilizer do not permeate evenly through the soil micropores, but instead leak directly and rapidly into the deep underground layers along cracks and large pores. In smart agriculture, low-cost microcontrollers, namely agricultural microcontrollers, are usually used as the control core to achieve large-scale deployment and cost control. Although these microcontrollers have basic digital signal processing capabilities, their memory capacity and floating-point operation speed are extremely limited, making them unable to support complex numerical calculations. Among existing soil foreign object detection technologies, electrical impedance imaging is a commonly used technique. It infers the internal conductivity distribution by measuring the voltage on the object's surface, thereby identifying cracks or cavities. However, traditional methods have significant drawbacks. They heavily rely on iterative inversion of a large number of nonlinear matrices to reconstruct images. This process takes several seconds on a regular computer. For agricultural microcontrollers with extremely low computing power, it is impossible to complete the calculation within the short window of priority leakage, resulting in a technical problem where the calculation lags behind the physical process. To solve the above technical problems, this invention utilizes the adjoint variable method in electrical impedance imaging. The core of the adjoint variable method is that the boundary measurement value, i.e. the sensitivity of the electrode voltage to a certain local parameter inside the medium, i.e. the change in conductivity, can be directly obtained through the calculation of the adjoint field. This method is used for complex nonlinear iterative solution processes, and the dimension reduction and reconstruction is transformed into a linear algebraic dot product mapping. This characteristic allows the computational complexity to be reduced to the order of magnitude that a microcontroller can handle. The goal of this step is to convert the voltage anomaly values ​​measured between each electrode pair into the contribution of each underground spatial node to the anomaly measured in this study. If a spatial node happens to be located at the location of a crack, the change in conductivity of that node will have the greatest impact on the boundary voltage measurement value, and therefore the node will be assigned the greatest contribution. Specifically, the effective elements in the differential voltage matrix after removing the diagonal elements are stretched into a column vector in column priority order, which serves as the potential anomaly vector. Each element in the potential anomaly vector corresponds to the voltage difference value under a certain pair of excitation measurement combinations. In a uniform, crack-free soil area, the difference value of the corresponding channel is close to zero. When there are soil cracks on the measurement path corresponding to a certain channel, the infiltration of salt water fertilizer into the cracks causes the local conductivity to increase. The voltage measured by this channel drops relative to the steady-state reference, and the difference value shows a significant non-zero amplitude. Therefore, the distribution and amplitude of the non-zero elements in the potential anomaly vector directly reflect the spatial location and severity of the soil cracks. The voltage difference value under a certain pair of excitation measurement combinations refers to the deviation of the currently measured potential from the reference potential of uniform, crack-free soil under a specific pair of current injection electrodes and a pair of voltage measurement electrodes. The sign and magnitude of this deviation directly characterize the degree of change in the conductivity of the soil area traversed by the current path. Multiply the pre-stored sensitivity transpose matrix with the potential anomaly vector to obtain the anomaly feature distribution vector. The row index of the obtained anomaly feature distribution vector has a one-to-one correspondence with the coordinates of the three-dimensional spatial nodes. This correspondence is defined by the 8×8×16 grid used when the sensitivity transpose matrix is ​​constructed offline and is solidified in the form of an index coordinate mapping table. The anomalous feature distribution vector maps the instantaneous potential anomaly to the associated potential value of each node in the entire space through the associated reciprocity of the electric field. That is, the potential change caused by the internal medium disturbance at the boundary can be equivalent to the potential distribution generated by an associated source in the background field. The abnormal feature distribution vector is subjected to Hadamard product operation with the pre-stored static reference potential vector, and then subjected to 3×3×3 median filtering to obtain the abnormal distribution sensitivity vector. The abnormal distribution sensitivity vector is a vector that reflects the degree of influence of medium topological changes on boundary measurement. Its maximum value region corresponds to the penetration position and outline of soil cracks in physical space. The process of obtaining the sensitivity transpose matrix specifically includes: Discretize the three-dimensional space of the main root layer, that is, the layer 20cm to 40cm below the soil surface, into... The nodes are discretized using a uniform 8×8×16 cubic mesh, with each cubic element having a side length of 1.25 cm. The total number of nodes is [number missing]. The three-dimensional coordinates of each node Defined as: cm, cm, cm, of which , ; The soil background electrical conductivity was determined by taking the measured average value of 0.35 S / m in the target agricultural area under normal irrigation conditions. The sampling depth was consistent with the probe burial depth, ranging from 20 cm to 40 cm. Five sampling points were randomly selected in the target agricultural area, and the four-electrode method was used to measure the conductivity under field water holding capacity conditions. The average value of the five measurements was taken as 0.35 S / m. Based on the geometric position of the electrode array and the aforementioned soil background conductivity, the Laplace equation is solved using the finite element method. For each excitation channel, a unit positive current and a unit negative current are applied to the corresponding two electrode nodes, respectively. With infinity as the zero potential boundary condition, the potential distribution at each node is calculated. For each measurement channel, according to the reciprocity theorem, the sensitivity of that channel to changes in the conductivity of each node is calculated using the inner product of the excitation field and the measurement field. The mathematical expression for the sensitivity is: ; in, For the first The excitation measurement channel for the first... The sensitivity of each node, For the first The potential distribution generated by the excitation current For the first Potential distribution generated by a unit current source at each node. For the first Traverse all the cubic unit regions corresponding to each spatial node. After obtaining one excitation measurement channel, Line × The original sensitivity matrix of the column; After transposing the original sensitivity matrix, L2 norm normalization is performed to obtain the sensitivity transpose matrix; The number of rows in the sensitivity transpose matrix is , In this embodiment, the total number of three-dimensional spatial nodes is used to discretize the underground detection area. The number of columns is , Each row index in the sensitivity transpose matrix corresponds to the three-dimensional coordinates of a three-dimensional spatial node. The correspondence between the row indexes and the three-dimensional coordinates of the spatial nodes is used to construct an index coordinate mapping table. The process of obtaining the static reference potential vector specifically includes: The static reference potential vector is also one The vector, consisting of 1 row and 1 column, is obtained by simulating the exact same current excitation mode as the actual measurement, under the same conditions as calculating the sensitivity transpose matrix, and then calculating this... The theoretical potential response at each spatial node.

[0024] This invention utilizes the adjoint variable method in electrical impedance imaging to reduce the dimensionality of the nonlinear high-dimensional matrix required for traditional boundary voltage inversion to a linear algebraic dot product mapping of the potential anomaly vector and the pre-stored sensitivity transpose matrix. By introducing a reference voltage matrix solidified after sufficient irrigation to construct differential features, the inherent influence of the static soil background medium on the potential distribution is successfully removed. This allows non-zero elements to be directly focused on the local potential collapse caused by preferential leakage, reducing the computational complexity of image reconstruction to the level that can be handled by a regular microcontroller. Combined with Hadamard product and three-dimensional median filtering, the sensitivity vector reflecting the topological changes of the medium can be extracted efficiently and smoothly, solving the technical contradiction between limited computing resources and lagging physical processes when deploying smart agriculture on a large scale at low cost.

[0025] Step S3 includes the following sub-steps: Step S301: Based on the pre-stored index coordinate mapping table, associate and arrange each element in the anomaly distribution sensitivity vector with the three-dimensional coordinates of the spatial nodes. Step S302: Divide the detection area into multiple vertical columns according to the vertical coordinates. For each vertical column, calculate the sum of the absolute differences of the sensitivity values ​​of the abnormal distribution of adjacent nodes along the depth direction, which is used as the total sensitivity change of the column. Step S303: Take the maximum value among the total sensitivity changes of all vertical columns as the distortion gradient.

[0026] In one specific embodiment of the present invention, based on an offline fixed index coordinate mapping table, each element in the anomaly distribution sensitivity vector is mapped according to the three-dimensional coordinates of its associated spatial node. Arrange; The entire detection area comprises 64 vertical columns, each corresponding to a fixed... The coordinates are 16 depth nodes per column. For each column, the difference in sensitivity values ​​of the abnormal distribution of adjacent nodes along the depth direction is calculated. The absolute values ​​are summed to obtain the total sensitivity change of the column. The maximum value among all total sensitivity changes is taken as the distortion gradient. In the above calculation process, the absolute values ​​of the sensitivity differences along the depth direction of each vertical column are summed. The summation result is simultaneously affected by the cross-sectional area of ​​the crack, which determines the magnitude of the difference, and the penetration depth, which determines the number of differences accumulated. Therefore, the distortion gradient is positively correlated with the product of the cross-sectional area of ​​the crack and the penetration depth. The distortion gradient can accurately characterize the overall leakage capacity of the preferred leakage channel. The larger the distortion gradient, the more unobstructed the large pore channel and the more intense the liquid flow.

[0027] This invention deconstructs the detection area into multiple vertical columns, and its differential accumulation result can be driven by both the cross-sectional area of ​​the crack (which determines the differential amplitude) and the penetration depth (which determines the number of accumulation terms). Therefore, its maximum value can serve as a strongly correlated physical characterization of the overall leakage capacity of the priority leakage channel.

[0028] Step S4 includes the following sub-steps: Step S401: Construct a monotonic mapping relationship from the distorted gradient to the fractional order; Step S402, and calculate the fractional order corresponding to the distortion gradient based on the monotonic mapping relationship. The value of the fractional order is in the interval (0,1).

[0029] In one specific embodiment of the present invention, the geometry of soil cracks not only determines the size of the physical channels for leakage, but also affects the dynamic characteristics of water transport within them; Specifically, when cracks are severely developed and penetrate deep, the flow of water and fertilizer in the cracks is close to free gravity infiltration, not dragged by soil capillary adsorption, and the flow rate is fast. The historical state has a weak influence on the current flow. When cracks are slight and only micropores are developed, the transport of water and fertilizer in the soil micropores is subject to strong capillary retardation, the flow rate is slow, and the historical state has a continuous long-tail effect on the current flow. In fractional-order dynamics theory, the duration of the influence of the historical state on the current flow is characterized by the fractional order. The closer the fractional order is to 1, the weaker the memory effect, and the closer it is to an integer-order process without memory; the closer the fractional order is to 0, the stronger the memory effect. This scheme characterizes the degree of crack development through distortion gradient. Therefore, this invention constructs a monotonic mapping from distortion gradient to fractional order. The larger the distortion gradient, the more developed the crack and the more porous the medium, and the closer the corresponding fractional order is to 1. Based on the above physical relationship, this embodiment calculates the first order by constructing a monotonic mapping. The mathematical expression for the first order is: ; in, For the first order, The first mapping constant, The second mapping constant, For distortion gradient; The first-order mathematical expression uses an exponential decay function to ensure the saturation characteristic of the mapping, when the cracks are extremely developed... At its maximum, Approaching 0, When the value approaches 1, water transport is close to memoryless integer-order diffusion, and the capillary adsorption retardation effect of the soil can be ignored; when leakage is slight, When smaller, Approaching Moisture transport maintains strong fractional-order characteristics, and the memory effect of capillary hysteresis adsorption is significant. The range of values ​​is limited to Within the interval, it conforms to the domain of the fractional order; First mapping constant The value of was determined through laboratory solute penetration experiments. One-dimensional solute penetration tests were conducted on 12 groups of undisturbed soil columns in the target agricultural area, using potassium bromide as a tracer. Penetration curves for each soil column were obtained. The fractional order of time in the fractional convection dispersion equation characterizes the non-Fick memory effect of solute transport, and its physical meaning is consistent with the time memory order in this scheme. Consistently, by backfitting the penetration curve, the baseline order of the uncracked soil was determined based on the fractional-order convection dispersion equation. The mathematical expression of the fractional-order convection dispersion equation is: ; in, This refers to the solute concentration. For time, For spatial coordinates, and They are the time and space fractional orders, respectively. The dispersion coefficient is... The average pore flow velocity is shown in Table 1-1. Table 1-1: Fitting results of solute penetration experiment on undisturbed soil column

[0030] The time fraction order of obtaining 12 groups of soil columns The values ​​range from 0.52 to 0.58, and the median of 0.55 is taken as the baseline order for uncracked soil in this embodiment. The benchmark value, therefore ; Second mapping constant The value of is related to the geometry of the probe array and the average electrical conductivity of the soil. Fifty three-dimensional models with different crack sizes were constructed using finite element simulation software, with crack widths ranging from 0.5 mm to 5 mm and depths from 5 cm to 40 cm. The spatial distortion gradient corresponding to each crack configuration was calculated. and the theoretical order based on fractional convection-diffusion equations ,by To constrain the mapping function, the Levenberg-Marquardt method nonlinear least squares algorithm is used. The fitting was performed, and the fitting results are shown in Table 1-2. Table 1-2: Fitting Table for the Second Mapping Constant

[0031] As can be seen from the table above, when When, goodness of fit The sum of squared residuals was 0.012, significantly better than other values. The recommended value range is This embodiment will Fixed at 2.0.

[0032] This invention establishes a distortion gradient to an exponentially decaying monotonic mapping relationship with saturation characteristics to quantify the non-Fick dispersion memory effect in porous media. Based on one-dimensional solute penetration experiments and finite element simulations to calibrate the mapping constant, the fractional order is strictly limited to the (0,1) interval. When soil cracks are severely developed and the distortion gradient is extremely large, the order adaptively approaches 1, characterizing the physical nature of water infiltration under near-free gravity and the weak influence of historical state on integer order. When leakage is slight and the medium compaction resistance is large, the order approaches the baseline value of 0.55, maintaining strong fractional order characteristics, thus conforming to the long-tailed memory effect brought about by soil capillary hysteresis adsorption.

[0033] Step S5 includes the following sub-steps: Step S501: Input the water pressure data into the first observer and output the first water pressure; The first observer adopts an integer-order extended state observer structure; In each water pressure control cycle, the first observer calculates the state variable value for the next cycle using the discrete-time state update equation based on the water pressure data collected at the current moment, the state variable value of the previous cycle, and the valve control quantity at the current moment.

[0034] In a specific embodiment of the present invention, the objective of this step is to separate the high-frequency, memoryless mechanical noise caused by pump motor pulsation, pipeline resonance, etc., and the disturbance of low-frequency capillary backflow equivalent pressure, which has a long-tail memory effect on water transport in deteriorated soil, from the water pressure signal mixed in the pipeline network, and to calculate the target negative pressure change required to block leakage based on the soil resistance after the interference is removed. The fractional order α calculated in each impedance detection cycle (50ms) is updated at the beginning of the next impedance detection cycle and remains unchanged in the subsequent 50ms water pressure control cycle as the order of the fractional integral operator in the second observer until the next update. This section adopts This serves as a discrete-time index (1ms period) for the water pressure control cycle, distinguishing it from the index (50ms period) used for the impedance detection cycle in step 1. The water pressure is input to the first observer, which adopts an integer-order extended state observer structure. Its discrete-time state update equations are as follows: ; ; ; in, and The water pressure at the current moment and the next moment, in kilopascals. It is the first observer's real-time estimate of the actual water pressure. and The rate of change of water pressure between the current moment and the next moment. and This represents the lumped mechanical disturbance value for the current and next time steps. The water pressure control step size is set to 0.001s; For discrete-time indexing; This represents the control quantity of the valve at the previous moment, dimensionless, with a range of values. Where -1 corresponds to all on and 1 corresponds to all off. , , The gain coefficient of the first observer. The first input gain coefficient; The water pressure measured at the current moment; The process of obtaining the first input gain coefficient through experimental calibration specifically includes: Send a step command signal to the variable frequency solenoid main control valve to switch the valve from the fully open state to the fully closed state and maintain it, so that the valve switches from the fully open state to the fully closed state in a very short time. The pressure transmitter installed downstream of the main control valve continuously records the water pressure after the valve is activated at a frequency of 1000Hz. The water pressure is recorded from 0.5 seconds before the step occurs to 2.0 seconds after the step occurs. The original water pressure sequence is formed by the measured water pressure. Using the valve opening command as input and the water pressure measured by the pressure sensor as output, a standard second-order transfer function is used as the dynamic model of the water pressure in the pipeline network. Through the least squares method, the values ​​of the pipeline network fluid volume gain constant, damping ratio, and frequency are continuously adjusted to minimize the sum of squared errors at each sampling point between the theoretical pressure curve calculated by G(s) under the same step input and the original water pressure sequence formed by the measured water pressure. After the fitting solution is completed, the ratio of the pipeline network fluid volume gain constant in this embodiment is obtained as 0.85, thereby determining the first input gain coefficient. ; In this embodiment, the mathematical expression for the water pressure dynamic model is: ; in, For water pressure dynamic model, The Laplace transform of the original water pressure sequence, For the Laplace transform of the step command signal, For Laplace variables, Let be the frequency to be solved, describing the speed of the water pressure response. Let be the damping ratio to be solved, which describes the degree of oscillation in the water pressure response; Let be the volumetric gain constant of the pipeline fluid to be solved; The process of determining the observer gain coefficient is based on the bandwidth parameterization design method for extended state observers, selecting the observer bandwidth. rad / s, then , , ; This observer relies on the instantaneous response characteristic of integer-order calculus operators, which have no historical integration memory, to quickly capture and absorb high-frequency oscillations in the signal, outputting a signal in each water pressure control cycle. The current value constitutes the first water pressure. Instantaneous value at the current moment , that is to say ; The initial state of the observer is set as follows: , , ,in, This serves as the reference pressure constant for steady-state operation of the pipeline network. The first observer, based on the instantaneous response characteristics of integer-order calculus operators without historical integral memory, can quickly capture and absorb high-frequency oscillations in the signal, and output the real-time estimate of the actual water pressure at the current moment as the first water pressure.

[0035] This invention employs an integer-order extended state observer with high bandwidth and specific gain design to process the original pipeline water pressure. Relying on the second-order transfer function model and the first input gain coefficient calibrated by step identification, the first observer can fully utilize the characteristics of integer-order calculus operators, which have no historical memory and are sensitive to instantaneous response. Within each discrete control step, the observer can quickly estimate the high-frequency memoryless mechanical noise caused by pump motor pulsation, pipeline resonance, etc., thereby outputting a high-fidelity real water pressure stripped of high-frequency oscillation interference, avoiding misjudgment caused by mechanical random disturbances in the subsequent low-frequency soil water potential feature extraction.

[0036] Step S5 also includes the following sub-steps: Step S502: Input the first water pressure into the second observer and output the soil backflow equivalent pressure. The second observer adopts a fractional-order expansion state observer structure. The fractional integral operator is approximated by the Oustaloup filter within a specified angular frequency range.

[0037] In a specific embodiment of the present invention, a first water pressure is input to a second observer. The second observer adopts an extended state observer structure based on an approximate fractional integrator using an Oustaloup filter. The dynamic equation of the second observer is in fractional integral form, and the mathematical expression of the dynamic equation of the second observer is: ; ; ; in, This represents the current effective soil water potential. The fractional rate of change of water potential at the current moment. This represents the equivalent soil resorption pressure at the current moment. for Fractional integral operator of order order, , and The gain coefficient of the second observer. The first water pressure input at the current moment. This is the second input gain coefficient. This represents the valve control value from the previous moment; express Fractional integral operator (transfer function) This operator is generated by the Oustaloup filter in a specified angular frequency range. In this embodiment, The approximation is performed within rad / s. This frequency band covers the low-frequency memory characteristics of soil moisture transport and the high-frequency dynamic characteristics of water hammer pressure, and can accurately approximate the amplitude and phase frequency characteristics of the fractional integral operator throughout the entire control process. The specific coefficients of the second observer were calculated based on the Oustaloup filter design method and verified through frequency domain simulation. within the rad / s band The amplitude frequency response approximation error is less than 1dB, and the phase frequency response error is less than... This satisfies the requirements of the present invention; The discrete transfer function and difference equation of the second observer are in the following forms: ; ; in The input signals of the three integrators are as follows: ; ; ; in, For the valve control input from the previous control cycle, each of the three channels runs a filter instance, with the outputs being respectively... , , The state update directly takes the filter output as: ; ; ; Since fractional integrators already incorporate a complete memory effect, this update method naturally aligns with fractional dynamics. , , Dimensions and The same, both are in kPa; The online update of Oustaloup filter coefficients uses a combination of table lookup and interpolation, while the offline stage pre-calculates the coefficients. exist The 41 sets of filter coefficients within the interval, with a step size of 0.01, are stored in a lookup table. When updated... Afterwards, according to the latest The value is read from the lookup table along with the corresponding coefficient. If the value is not in the lookup table node, linear interpolation is used to obtain the coefficient; because The parameters are updated only once every 50ms, and adjacent cycles... The changes are gradual, and the filter coefficients employ a linear weighted fade-in / fade-out algorithm with a transition period. When an update to the α parameter causes a change in the filter coefficients, a transition period of 10 control cycles is set. During this transition period, the final output of the filter... The output of the old coefficient filter and the output of the new coefficient filter Weighted composition: λ increases linearly from 0 to 1 over 10 periods to achieve a smooth transition and avoid abrupt numerical changes. After the transition period, the system is fully switched to the new coefficients. To avoid coefficient update conflicts, an update flag is set. If the process is in progress, the update is ignored and the next update is performed after the process is complete. The process of obtaining the second input gain coefficient through experimental calibration specifically includes: The second input gain coefficient is determined through offline identification experiments. A series of step commands with different opening degrees are sent to the electromagnetic main control valve, and the corresponding steady-state pressure values ​​are recorded. A static mapping relationship between valve opening degree and steady-state pressure is established. This mapping relationship is then linearly fitted, and the slope of the fitted line is the second input gain coefficient. In this embodiment... The second input gain coefficient indicates that for every unit change in valve opening, from 0 to 1, the steady-state water pressure in the pipeline network changes by 15.0 kPa. Using the same bandwidth design method as the first observer, the bandwidth of the second observer is taken. rad / s, to obtain , , ; The initial state of the second-level observer is set as follows: , , ,in, This serves as the reference pressure constant for steady-state operation of the pipeline network. Each water pressure control cycle is obtained The equivalent pressure of water reabsorption due to the current soil micropore structure, expressed in kilopascals, is used as the equivalent pressure for soil reabsorption. ,make The equivalent pressure of soil reabsorption is the output of the second observer; In addition to outputting the equivalent soil reabsorption pressure, the second observer also provides crucial model simplification functions, because The real-time estimation of the total disturbance, including the equivalent pressure of soil reabsorption, is used to compensate for the dynamic relationship between valve control quantity and pipeline pressure into an approximate integral series system as follows: ; in, Given the current pipeline pressure, For the next moment of pipeline pressure, This represents the valve control quantity at the current moment. This is the second input gain coefficient. The total disturbance estimated by the second observer; This approximate relationship indicates that within each control cycle, the pressure change is determined by the valve control quantity and the known total disturbance. A joint decision.

[0038] This invention introduces a second-stage extended state observer based on an approximate fractional integrator using an Oustaloup filter. Its integral operator operates within a specific angular frequency range, covering the low-frequency memory band of soil moisture transport. It uses table lookup and linear interpolation to update the fractional order generated by impedance sensing online, and combines it with a linear weighted fade-in / fade-out algorithm with a transition period to achieve adaptive evolution of the model structure while preventing numerical mutations. This observer can accurately extract the low-frequency equivalent pressure of soil micropore structure on water reabsorption from the pipeline water pressure signal, and online compensate and simplify the complex valve, pipeline, and soil nonlinear dynamic system into an approximate integral series system.

[0039] Step S5 also includes the following sub-steps: Step S503: Calculate the negative pressure change based on the equivalent pressure of soil reabsorption, and obtain the target negative pressure peak value by combining it with the steady-state operating benchmark pressure of the pipeline network; The change in negative pressure is obtained by adding the water hammer pressure term, the gravitational potential energy term, and the soil reabsorption equivalent pressure. The target negative pressure peak is the difference between the steady-state operating reference pressure of the pipeline network and the change in negative pressure.

[0040] In one specific embodiment of the present invention, if preferential seepage caused by soil fissures is occurring, water and fertilizer will seep rapidly downwards along the fissures under the action of gravity, forming a downward water flow. In order to retain this part of the water and fertilizer that is about to leak out in the main root layer, the present invention utilizes the negative pressure water hammer effect generated by rapidly closing the electromagnetic main control valve to create a reverse pressure wave in the drip irrigation pipeline. This pressure wave is transmitted to the underground fissures, forming a transient negative pressure in the fissures, which counteracts the downward momentum of the water flow and overcomes the gravity of the liquid column, drawing the water and fertilizer back from the fissure channel to the surrounding soil micropores 20cm to 40cm deep, and the water and fertilizer are redistributed inside the soil. Specifically, based on the equivalent pressure of soil reabsorption, the change in negative pressure is calculated, and the mathematical expression for the change in negative pressure is: ; in, The change in negative pressure is taken as 1050 kg / m³. The density of the water-fertilizer mixture, The water hammer wave velocity in the water-fertilizer mixture. This refers to the steady-state flow rate of the water-fertilizer mixture within the pipe network. Let gravitational acceleration be 9.8 m / s². The set leakage interception depth is 0.4m; The conversion factor for pipeline cross-sectional resistance is taken as 0.07; This is the equivalent pressure of soil reabsorption; The water hammer wave velocity is calculated using the water hammer wave velocity formula, and the mathematical expression of the water hammer wave velocity formula is: ; in, The bulk modulus of water is 2.2 GPa. The elastic modulus of PVC-U pipe is 3.0 GPa. The inner diameter of the pipe is 28mm. The wall thickness is 2.0 mm. With a support coefficient of 1.0, the calculation is as follows: m / s; The steady-state flow rate of the water-fertilizer mixture in the pipeline network is calculated based on the designed drip flow rate and the cross-sectional area of ​​the main pipe. The steady-state flow rate is the ratio of the drip flow rate to the cross-sectional area of ​​the drip irrigation main pipe. In this embodiment, the steady-state flow rate of the water-fertilizer mixture in the pipeline network is taken as 0.5 m / s. The process of obtaining the steady-state operating reference pressure of the pipeline network by using the difference between the steady-state operating reference pressure and the negative pressure change as the target negative pressure peak value specifically includes: Open the valve fully and run the water and fertilizer integrated machine to stabilize its output pressure at the rated working pressure. After the pipeline is completely vented, directly read the pressure value of the pressure transmitter to obtain the steady-state operating reference pressure of the pipeline network. In this embodiment, 120 kPa is used. Pipeline cross-sectional resistance conversion factor The method for determining the value is as follows: A three-dimensional computational fluid dynamics model is constructed, including the drip irrigation trunk line, branch lines, and capillary tubes. Different intensities of reverse suction pressure are simulated at the capillary tube outlet. The friction loss and local head loss along the pipeline are calculated. The average value of the ratio of total head loss to the equivalent soil backflow pressure under typical operating conditions is taken as the... The calibration result was 0.07; Obtain the conversion factor of pipeline cross-sectional resistance The process specifically includes: applying reverse suction pressure (negative pressure) of different intensities at the capillary outlet end to simulate the pipeline head loss when water hammer shock waves are transmitted in reverse. In the simulation model, by setting different soil physical parameters, including soil texture, compaction degree, and initial moisture content, various typical soil conditions from sandy soil to clay soil are simulated. Under each soil condition, the corresponding equivalent soil back suction pressure is calculated based on the virtual water pressure data generated by the simulation. At the same time, the total head loss of the pipeline network under this condition is recorded, including friction loss and local loss. The ratio of the total head loss to the equivalent soil back suction pressure is used as the road section resistance conversion coefficient. The results are shown in Table 2 below. Table 2: Calibration Results of Pipeline Cross-sectional Resistance Conversion Coefficient under Different Soil Types

[0041] As can be seen from the table above, the conversion factor for pipeline cross-sectional resistance is set to 0.07 in this embodiment.

[0042] Step S504: Calculate the valve control quantity based on the target negative pressure peak value and the state estimate from the second observer. The mathematical expression for the valve control quantity is: ; in, This is the valve control quantity at the current moment, dimensionless, and its value range is limited to the interval [-1, 1]. The gain is a proportional feedback gain, dimensionless, and in this embodiment, it is taken as... ; The target negative pressure peak value at the current moment, in units of ; The effective soil water potential at the current moment is estimated by the second observer, in units of ; The equivalent soil reabsorption pressure at the current moment, estimated by the second observer, in units of ; This is the second input gain coefficient; The term represents the deviation between the target negative pressure peak and the current estimated effective soil water potential, and serves as a proportional control term to drive valve action; Target negative pressure peak It is updated once at the end of each detection cycle and remains constant for the subsequent 50 water pressure control cycles, when no preferential leakage is detected. The value is taken as the reference pressure for steady-state operation of the pipeline network. At this point, the control law degenerates into a conventional constant pressure regulation. When preferential leakage is detected, The target negative pressure peak value is taken as S503. The control law drives the pipeline pressure to drop rapidly to this peak value in order to generate a negative pressure water hammer effect. Proportional feedback gain The value of is determined through frequency domain response experiments of the closed-loop system. The value that minimizes the pressure tracking rise time is chosen while ensuring system stability. Value, in this embodiment , The experimental control results for the selected values ​​are shown in Table 3. Table 3: Proportional Feedback Gain Experimental results of value selection

[0043] As can be seen from the above table, This embodiment selects [a specific method / mechanism] to achieve the fastest pressure tracking without causing significant overshoot and pipeline oscillation. .

[0044] When preferential leakage is detected, this invention can calculate the target negative pressure peak that can counteract the downward momentum of the liquid flow and overcome the gravity of the liquid column based on the precise water hammer wave velocity and the steady-state flow velocity of the pipeline. By quickly adjusting the valve control quantity, a negative pressure water hammer effect is triggered. This transient negative pressure is transmitted in reverse to the underground cracks, which can generate a reverse suction force. Without interfering with the normal nutrient supply, the water and fertilizer that are about to be lost are actively intercepted, drawn back and redistributed to the porous media of the root system, thus realizing the physical-level active interception of preferential leakage.

[0045] Step S6 includes the following sub-steps: Step S601: Collect meteorological data, crop canopy data, and historical irrigation data to construct a multidimensional agricultural big data time series. Meteorological data, crop canopy data, and historical irrigation data were collected. Meteorological data included: daily average temperature, daily average relative humidity, and cumulative solar synergistic radiation (CRN). (Unit: ...) Daily average wind speed; Crop canopy data, including the normalized vegetation index (NDVI), was acquired using a field multispectral sensor. The multispectral sensor was installed on a horizontal bar 2.5m above the ground in the field, with its field of view covering four rows of crop canopy. Data was collected every 30 minutes, and the value collected at 12:00 noon each day was taken as the representative NDVI value for that day. Historical irrigation volume includes the actual daily irrigation volume for the previous 7 days, in m³ / mu. The actual daily irrigation volume is collected by an electromagnetic flow meter installed on the main outlet pipe of the water and fertilizer machine. The flow meter records the instantaneous flow rate once per minute. After the irrigation is completed each day, the total flow rate of the day is accumulated and converted into the irrigation volume per unit area. The daily soil reabsorption equivalent pressure measured in step S5 of this method is used as the basis for this method. The daily average is used as an additional input; All collected data were normalized, and each input feature was obtained. These input features were then aligned by day to form a multidimensional time-series input sequence with a day as the time step. Where 7 represents the time window length of 7 days, and d represents the feature dimension, with a value of 7, namely 4 meteorological features, 1 canopy feature, 1 historical irrigation volume feature, and 1 soil reabsorption equivalent pressure feature, the output is... Forecast of water and fertilizer demand for the next day, unit: m³ / mu; Step S602: The fractional order is used as the index of the memory forgetting factor and embedded into the cell state update equation of the long short-term memory network to construct a fractional order long short-term memory network model. The fractional-order long short-term memory network model adopts a two-layer LSTM stacked structure, including an input layer, a first LSTM layer, a second LSTM layer, a dropout layer, and a fully connected output layer; The input layer dimension is The time step is 7, the feature dimension is 7, which corresponds to the multi-dimensional time-series input sequence constructed in step S601; The first LSTM layer contains 64 hidden units, and the activation function is... The recurrent activation function is sigmoid. The second LSTM layer contains 32 hidden units, and the activation function is... The cyclic activation function is sigmoid; The dropout layer is placed between two LSTM layers with a dropout rate of 0.2 to prevent overfitting.

[0046] The fully connected output layer contains one neuron with a linear activation function, and outputs the predicted water and fertilizer demand for the next day.

[0047] The difference between fractional-order Long Short-Term Memory (LSTM) networks and traditional LSTM networks lies in replacing the integer-order time derivative in the traditional LSTM cell state update equation with a fractional-order derivative. Specifically, the traditional LSTM cell state update is as follows: ; This invention modifies it into a fractional form, the mathematical expression of which is: ; in, For the definition based on Grünwald-Laitnikov Fractional derivative operators, The value is the first order obtained from step S4. ), Let be the cell state vector at time step t, i.e., the object of the fractional derivative. Output for the forget gate. For input gate output, Candidate cell state; The implementation of the fractional derivative in the discrete-time domain is consistent with the GL implementation method of the fractional integrator in step S5, by introducing a dynamic fractional order. The model can adaptively adjust the weighting of historical irrigation data on current demand forecasts based on the current development level of soil cracks, i.e., the severity of seepage and leakage. The more developed the cracks ( (Approaching 1), the faster the historical water loss, the more the model automatically shortens the memory window for historical water. Specifically, the discrete form of the fractional derivative based on the Grünwald-Laitnikov definition is as follows: ; in, It is an α-order fractional derivative operator. Let be the cell state vector at time step t. The discrete time step, in this invention, is equal to the daily time step for LSTM processing time-series data, i.e., 1 day. The coefficients are binomial coefficients. ,in, For gamma function, For the first Cell state vectors at each time step; Let the expansion, etc. Separate item ,available The explicit solution expression.

[0048] Step S603: The equivalent pressure of soil reabsorption is added as a physical constraint term to the loss function of the fractional-order long short-term memory network model; The fractional-order long short-term memory network model is trained using supervised learning, and the loss function is designed as follows: ; in, For batch size, this example uses 32. For the first The actual water and fertilizer requirements of each sample (unit: m³ / mu). For the model to the first The predicted value for each sample, For physical constraint weighting coefficients, this embodiment takes... , The threshold value for the equivalent pressure of healthy soil reabsorption was set. For the first The daily average equivalent soil reabsorption pressure for each sample during the corresponding time period; Healthy soil reabsorption equivalent pressure threshold Field experiments determined that, in the target agricultural area, control plots with intact soil structure and no cracks were selected, and the soil reabsorption equivalent pressure was continuously monitored for 30 days. The 25th percentile of the monitored values ​​was taken as the baseline. In this embodiment ; Penalty term in the loss function The physical meaning is that when the soil reabsorption equivalent pressure Below the health threshold When the soil has poor water retention capacity, it may be caused by preferential leakage due to cracks. The model will be penalized during training to encourage it to learn to appropriately increase the water and fertilizer prediction values ​​in this case to compensate for possible leakage. Step S604: Input the multidimensional agricultural big data time series into the trained fractional-order long short-term memory network model and output the predicted water and fertilizer demand for the next period.

[0049] The training process of the fractional-order long short-term memory network model is as follows: Daily field data from April to October of the past three growing seasons were collected from the target agricultural area. This included data on temperature, humidity, photosynthetically active radiation, and wind speed collected at field weather stations; normalized difference vegetation index collected using field multispectral sensors; actual irrigation volume collected using smart water meters; and the detection system of this invention, specifically the daily soil backflow equivalent pressure. Average value, collected from multiple valid daily samples; Z-score standardization was performed on all samples. All samples were divided into training set, validation set and test set in chronological order with a ratio of 6:2:2. The first two growth seasons were used as the training set, the first four months of the third growth season were used as the validation set, and the last three months of the third growth season were used as the test set. The optimizer uses the Adam algorithm with an initial learning rate of 0.001. The learning rate is reduced to 0.9 times the current rate after every 50 training epochs. The batch size is 32, and the number of training epochs is 300. Early stopping is used, and training stops when the validation set loss no longer decreases after 50 consecutive epochs. The model parameters with the minimum validation set loss are saved. All weight matrices in the network are initialized with a uniform Xavier distribution, and the bias vector is initialized to zero.

[0050] The network achieves physical enhancement of temporal features by introducing the equivalent pressure of soil reabsorption into the temporal input. By embedding dynamic fractional order into the cell state update equation of traditional LSTM, the traditional unbiased forget gate is upgraded to a fractional derivative operator with underlying physical mechanism, enabling the model to automatically adjust the memory window of historical moisture data according to the development of cracks and the severity of leakage. A physical penalty term for the deterioration of soil water retention capacity is introduced into the loss function, prompting the network to appropriately increase the prediction compensation for the working conditions of deteriorating water retention during training.

[0051] An agricultural big data water and fertilizer demand forecasting system includes: The data acquisition module is used to collect voltage and water pressure data. An anomaly detection module is used to obtain an anomaly distribution sensitivity vector based on voltage data and a pre-stored sensitivity transpose matrix; The gradient calculation module is used to arrange the anomaly distribution sensitivity vector according to the three-dimensional coordinates of the spatial nodes, calculate the sum of the absolute differences of the anomaly distribution sensitivity values ​​of adjacent nodes along the depth direction as the total sensitivity change of each vertical column, and take the maximum value of the total sensitivity change of all vertical columns as the distortion gradient. The order mapping module is used to calculate the fractional order based on the distortion gradient by constructing a monotonic mapping from the distortion gradient to the fractional order. The pressure estimation module is used to input water pressure data into the first observer to obtain the first water pressure, input the first water pressure into the second observer to obtain the soil reabsorption equivalent pressure, and the integration order of the second observer is determined by the fractional order; the target negative pressure peak is calculated based on the soil reabsorption equivalent pressure. The demand forecasting module is used to construct and train a fractional-order long short-term memory network prediction model based on the equivalent pressure of soil reabsorption, and output the predicted water and fertilizer demand for the next cycle.

[0052] This invention utilizes the adjoint variable method to reduce the dimension of traditional electrical impedance imaging based on complex nonlinear iterative inversion to linear algebraic dot product mapping, effectively avoiding the huge burden of matrix operations. This design significantly reduces the computational complexity to the level that can be handled by low-cost agricultural microcontrollers, and can calculate the distortion gradient characterizing the degree of soil fissure development in real time during the short window period when preferential leakage occurs, solving the problem that the calculation of existing detection technology is seriously lagging behind the physical leakage process. Based on this, the present invention constructs a two-order extended state observer architecture. It utilizes an integer-order observer without historical integral memory to quickly absorb high-frequency mechanical noise such as water pump pulsation, and uses a fractional-order observer approximated by an Oustaloup filter to accurately extract the soil backflow equivalent pressure with low-frequency long-tail memory characteristics. This not only calculates the target negative pressure and triggers the water hammer effect, but also actively draws back the water and fertilizer that are rapidly leaking along the large pores to the root micropores for redistribution. Furthermore, it provides a reliable physical boundary input for subsequent big data prediction. In terms of water and fertilizer demand prediction, this invention directly embeds the fractional order obtained from physical detection into the cell state update equation of the Long Short-Term Memory network as a derivative operator, enabling the model to adaptively adjust the memory weights of historical irrigation and water loss to the current prediction based on the current dynamic development of soil cracks. This invention also incorporates the extracted equivalent soil backflow pressure as a physical constraint term into the loss function, and applies additional penalties to the prediction bias that ignores soil deterioration and reduced water retention capacity during model training. This invention achieves deep coupling between agricultural time-series big data and the dynamics of porous media in the underlying soil, significantly improving the robustness and accuracy of water and fertilizer demand prediction under complex field conditions and scenarios with frequent seepage.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions can also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A method for predicting agricultural water and fertilizer demand using big data, characterized in that, Includes the following steps: Step S1: Collect voltage and water pressure data; Step S2: Based on the voltage data and the pre-stored sensitivity transpose matrix, obtain the abnormal distribution sensitivity vector; Step S3: Arrange the abnormal distribution sensitivity vector according to the three-dimensional coordinates of the spatial nodes, calculate the sum of the absolute differences of the abnormal distribution sensitivity values ​​of adjacent nodes along the depth direction as the total sensitivity change of each vertical column, and take the maximum value of the total sensitivity change of all vertical columns as the distortion gradient. Step S4: Based on the distortion gradient, calculate the fractional order by constructing a monotonic mapping from the distortion gradient to the fractional order. Step S5: Input the water pressure data into the first observer to obtain the first water pressure, input the first water pressure into the second observer to obtain the soil reabsorption equivalent pressure, and the integration order of the second observer is determined by the fractional order; calculate the target negative pressure peak based on the soil reabsorption equivalent pressure; Step S6: Using the fractional order and the equivalent soil reabsorption pressure, construct and train a fractional-order long short-term memory network prediction model to output the predicted water and fertilizer demand for the next cycle.

2. The method for predicting agricultural water and fertilizer demand using big data as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Collect voltage data through a probe array and construct a voltage matrix; Step S102: Collect water pressure data using a pressure sensor.

3. The method for predicting agricultural water and fertilizer demand using big data as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S201: Stretch the effective elements of the differential voltage matrix after removing the diagonal elements into a column vector, which serves as the potential anomaly vector; Step S202: Multiply the pre-stored sensitivity transpose matrix with the potential anomaly vector to obtain the anomaly feature distribution vector; Step S203: Perform a Hadamard product operation on the abnormal feature distribution vector and the pre-stored static reference potential vector, and perform median filtering on the operation result to obtain the abnormal distribution sensitivity vector.

4. The method for predicting agricultural water and fertilizer demand using big data as described in claim 3, characterized in that, Step S3 includes the following sub-steps: Step S301: Based on the pre-stored index coordinate mapping table, arrange each element in the abnormal distribution sensitivity vector in association with the three-dimensional coordinates of the spatial nodes. Step S302: Divide the detection area into multiple vertical columns according to the vertical coordinates. For each vertical column, calculate the sum of the absolute differences of the sensitivity values ​​of the abnormal distribution of adjacent nodes along the depth direction, which is used as the total sensitivity change of the column. Step S303: Take the maximum value among the total sensitivity changes of all vertical columns as the distortion gradient.

5. The method for predicting agricultural water and fertilizer demand using big data as described in claim 4, characterized in that, Step S4 includes the following sub-steps: Step S401: Construct a monotonic mapping relationship from the distorted gradient to the fractional order; Step S402, and calculate the fractional order corresponding to the distortion gradient based on the monotonic mapping relationship, wherein the value of the fractional order is in the interval (0,1).

6. The method for predicting agricultural water and fertilizer demand using big data as described in claim 5, characterized in that, Step S5 includes the following sub-steps: Step S501: Input the water pressure data into the first observer and output the first water pressure; The first observer adopts an integer-order extended state observer structure; In each water pressure control cycle, the first observer calculates the state variable value for the next cycle based on the water pressure data collected at the current moment, the state variable value of the previous cycle, and the valve control quantity at the current moment through the discrete-time state update equation.

7. The method for predicting agricultural water and fertilizer demand using big data as described in claim 6, characterized in that, Step S5 further includes the following sub-steps: Step S502: Input the first water pressure into the second observer and output the soil backflow equivalent pressure. The second observer adopts a fractional-order expansion state observer structure.

8. The method for predicting agricultural water and fertilizer demand using big data as described in claim 7, characterized in that, Step S5 further includes the following sub-steps: Step S503: Calculate the negative pressure change based on the soil backflow equivalent pressure, and obtain the target negative pressure peak value by combining it with the steady-state operation benchmark pressure of the pipeline network; The change in negative pressure is obtained by adding the water hammer pressure term, the gravitational potential energy term, and the soil reabsorption equivalent pressure. The target negative pressure peak value is the difference between the steady-state operating reference pressure of the pipeline network and the negative pressure change. Step S504: Calculate the valve control quantity based on the target negative pressure peak value and the state estimate value of the second observer.

9. The method for predicting agricultural water and fertilizer demand using big data as described in claim 8, characterized in that, Step S6 includes the following sub-steps: Step S601: Collect meteorological data, crop canopy data, and historical irrigation data to construct a multidimensional agricultural big data time series. Step S602: The fractional order is used as the index of the memory forgetting factor and embedded into the cell state update equation of the long short-term memory network to construct a fractional order long short-term memory network model. Step S603: The equivalent pressure of soil reabsorption is added as a physical constraint term to the loss function of the fractional-order long short-term memory network model; Step S604: Input the multidimensional agricultural big data time series into the trained fractional-order long short-term memory network model and output the predicted water and fertilizer demand for the next period.

10. An agricultural big data water and fertilizer demand forecasting system, which is applied in the agricultural big data water and fertilizer demand forecasting method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect voltage and water pressure data. An anomaly detection module is used to obtain an anomaly distribution sensitivity vector based on the voltage data and a pre-stored sensitivity transpose matrix. The gradient calculation module is used to arrange the abnormal distribution sensitivity vector according to the three-dimensional coordinates of the spatial nodes, calculate the sum of the absolute differences of the abnormal distribution sensitivity values ​​of adjacent nodes along the depth direction as the total sensitivity change of each vertical column, and take the maximum value of the total sensitivity change of all vertical columns as the distortion gradient. The order mapping module is used to calculate the fractional order based on the distortion gradient by constructing a monotonic mapping from the distortion gradient to the fractional order. The pressure estimation module is used to input the water pressure data into a first observer to obtain a first water pressure, input the first water pressure into a second observer to obtain the soil reabsorption equivalent pressure, wherein the integration order of the second observer is determined by the fractional order; and calculate the target negative pressure peak based on the soil reabsorption equivalent pressure. The demand forecasting module is used to construct and train a fractional-order long short-term memory network prediction model based on the fractional order and the equivalent soil reabsorption pressure, and output the predicted water and fertilizer demand for the next cycle.

Images