A crop precision irrigation method and system based on algorithm fusion

Abstract

The application discloses a kind of crop precision irrigation method and system based on algorithm fusion, it is related to intelligent agriculture and precision irrigation control technical field, method includes: the time series data of meteorology, soil, crop physiology and irrigation process are collected, carry out exception elimination, missing completion and standardization processing and construct derivative feature;The historical characteristic sequence is input to the bidirectional time series network with attention to obtain time series representation, input neural fuzzy inference network to output future water demand prediction value and corresponding prediction interval, obtain confidence from prediction interval;Water demand prediction value, confidence and crop stress are input to the first layer fuzzy inference to obtain basic irrigation duration;Soil water potential and post-irrigation feedback are input to the second layer fuzzy inference to obtain valve fine tuning amount;Fusion generates irrigation instruction and drives execution equipment.By outputting water demand prediction interval and constructing confidence, the prediction reliability is improved, and by double-layer fuzzy inference, soil water potential and post-irrigation feedback are combined to realize executable and safety protection of irrigation instruction.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture and precision irrigation control technology, specifically to a crop precision irrigation method and system based on algorithm fusion. Background Technology

[0002] With the increasing demand for large-scale and refined management in agricultural production, irrigation has gradually shifted from traditional experience-based irrigation to data-driven precision irrigation. The core of precision irrigation lies in reducing water waste and the risk of over-irrigation while meeting crop water requirements, and ensuring the stability and safety of the irrigation process. Current irrigation management practices commonly include triggering irrigation based on soil moisture content thresholds or simple meteorological indicators, estimating water requirements using empirical formulas, or using fixed-rule control strategies to generate irrigation duration and valve opening / closing actions. However, in actual field environments, crop water requirements are influenced by a combination of factors, including weather changes, soil hydraulic properties, crop growth stages, and stress conditions. Furthermore, data collection suffers from noise, missing data, and abnormal fluctuations, making it difficult for the above methods to consistently provide accurate and executable irrigation strategies over the long term.

[0003] On the one hand, existing water demand forecasting or irrigation decision-making methods often fail to adequately address "data quality issues" and "multi-source heterogeneous time-series characteristics." Meteorological data, soil moisture and water potential data, crop physiological data, and irrigation process data collected by field sensors typically exhibit inconsistent sampling frequencies, numerous missing points, and frequent anomalous changes. Simply using filtering or direct input into the model can easily lead to the accumulation of prediction bias and control errors, thus affecting the reliability of irrigation decisions. On the other hand, many existing solutions typically output a single water demand forecast or a single irrigation duration suggestion, lacking an expression of prediction uncertainty. This makes it difficult to provide robust strategies in the face of sudden weather changes, abnormal crop conditions, or declining data quality, easily leading to over-irrigation or under-irrigation.

[0004] On the other hand, existing control strategies often rely on single-layer rules or single feedback variables, lacking a collaborative mechanism of "forward prediction and fine-tuning." Actual irrigation processes are affected by fluctuations in pipeline pressure, valve operation deviations, and differences in soil infiltration capacity. Even with a reasonable irrigation duration, variations in the response at the execution end can lead to insufficient root zone water replenishment or deep leakage. If the control strategy is based solely on a one-time calculated irrigation duration or simple threshold-based valve opening and closing, it is difficult to make timely adjustments and protective controls based on changes in soil water potential and the actual post-irrigation water replenishment effect, reducing adaptability and safety in complex field conditions.

[0005] Furthermore, crop growing environments exhibit significant time-varying and regional differences. Factors such as soil texture, irrigation facility status, and changes in crop growth stages necessitate continuous updates to model parameters and rule bases. Many existing methods lack effective post-irrigation evaluation and online incremental update mechanisms, making it difficult to utilize post-irrigation response data to form closed-loop optimization. This leads to long-term drift in model and rule performance degradation, hindering the achievement of continuous and stable precision irrigation management. Summary of the Invention

[0006] Based on the shortcomings of the prior art described above, the purpose of this invention is to provide a crop precision irrigation method and system based on algorithm fusion to solve the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a crop precision irrigation method based on algorithm fusion, comprising:

[0008] Collect time-series data on meteorology, soil, crop physiology, and irrigation processes;

[0009] The time-series data is subjected to anomaly removal, missing data completion, and standardization, and derived features are constructed.

[0010] The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation. The input is then input into a neural fuzzy inference network to output the predicted value of future water demand and the corresponding prediction interval. The confidence level is obtained from the prediction interval. The confidence level is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing it.

[0011] The water demand forecast, confidence level, and crop stress level are input into the first layer of fuzzy inference to obtain the basic irrigation duration;

[0012] The soil water potential and post-irrigation feedback are input into the second layer of fuzzy inference to obtain the valve fine-tuning amount. The post-irrigation feedback is the rate of increase of soil moisture content within a preset time window after the last irrigation.

[0013] The irrigation command is generated and the execution device is driven. The irrigation command includes at least the irrigation start time, the final irrigation duration, and the valve opening sequence.

[0014] The present invention is further configured such that the meteorological data includes temperature, humidity, wind speed, radiation and precipitation; the soil data includes soil moisture content at multiple depths, soil temperature, electrical conductivity and soil water potential; the crop physiological data includes canopy infrared temperature data and / or stem flow data; and the irrigation process data includes flow rate, pressure, valve opening and closing status and valve opening degree.

[0015] The present invention is further configured such that: anomaly removal includes joint detection based on constraints of change magnitude between adjacent time points and outlier discrimination; missing data completion includes completion based on temporal interpolation or nearest neighbor interpolation; and standardization processing includes dimensional unification based on sliding window statistics.

[0016] The present invention is further configured such that the derived features include reference evapotranspiration, soil available water content, crop growth process characteristics, and crop stress; wherein, reference evapotranspiration is calculated from temperature, humidity, wind speed, and radiation; soil available water content is calculated from soil moisture content combined with soil field capacity and wilting coefficient; and crop stress is obtained by standardizing the difference between canopy infrared temperature and meteorological temperature.

[0017] The present invention is further configured such that the historical feature sequence is composed of multi-source derived features within a preset historical time window aligned by time; when outputting the future water demand forecast value and the corresponding forecast interval, the meteorological forecast features of the preset number of days in the future are jointly input with the historical feature sequence to generate the water demand forecast value and the corresponding forecast interval for the preset number of days in the future.

[0018] The present invention is further configured such that the neural fuzzy inference network includes an input fuzzification layer, a rule activation layer, a normalized weighted layer, and a defuzzification layer; the input fuzzification layer fuzzifies the temporal representation and its spliced ​​features based on a trainable membership function; the defuzzification layer outputs the predicted value of future water demand, the lower bound of the prediction interval, and the upper bound of the prediction interval, respectively, forming the corresponding prediction interval.

[0019] The present invention is further configured such that the input of the first layer of fuzzy inference includes at least the difference between the predicted future water demand and the available soil water content, the future water demand change trend, the confidence level, and the crop stress level; the rule base of the first layer of fuzzy inference is configured to reduce the basic irrigation duration and / or split the basic irrigation duration into multiple segments for execution when the confidence level decreases, in order to suppress the risk of over-irrigation caused by prediction uncertainty.

[0020] The invention is further configured such that the input of the second layer of fuzzy inference includes the real-time value of the root zone soil water potential and post-irrigation feedback, and the valve fine-tuning amount represents the adjustment range of the valve opening sequence; the final irrigation duration is obtained by combining the basic irrigation duration with the adjustment coefficient corresponding to the valve fine-tuning amount, and the valve opening sequence is obtained by superimposing the valve fine-tuning amount on the basic opening sequence. When the soil water potential reaches the preset safety threshold, a protective action of reducing the opening or closing the valve is triggered.

[0021] The invention is further configured to include an effect score, which is jointly determined by the sufficiency of water replenishment, the uniformity of water distribution, and the trend of crop stress relief within a preset time window after irrigation. A rolling sample set is constructed based on the effect score, the prediction model is incrementally updated, and the weights and membership function parameters of the fuzzy rules are self-tuned.

[0022] This invention also provides a precision crop irrigation system based on algorithm fusion, used to implement the above-mentioned precision crop irrigation method based on algorithm fusion, comprising: a processor, a memory, and a communication interface, wherein the memory stores a program executable by the processor, and the processor executes the program to implement the following functional units:

[0023] Data acquisition module: Collects time-series data on meteorology, soil, crop physiology, and irrigation processes;

[0024] Data processing module: performs anomaly removal, missing data completion, and standardization on time-series data, and constructs derived features;

[0025] Fusion prediction module: The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation. The input is then input into a neural fuzzy inference network to output the predicted value of future water demand and the corresponding prediction interval. The confidence level is obtained from the prediction interval. The confidence level is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing it.

[0026] Duration Decision Module: Input the water demand forecast, confidence level, and crop stress into the first-level fuzzy inference to obtain the basic irrigation duration;

[0027] Valve position fine-tuning module: The soil water potential and post-irrigation feedback are input into the second layer of fuzzy inference to obtain the valve fine-tuning amount. The post-irrigation feedback is the rate of increase of soil moisture content within a preset time window after the last irrigation.

[0028] Command-driven module: Integrates and generates irrigation commands and drives the execution equipment. The irrigation commands include at least the irrigation start time, the final irrigation duration, and the valve opening sequence.

[0029] This invention provides a precision crop irrigation method and system based on algorithm fusion. The method collects time-series data on meteorology, soil, crop physiology, and irrigation processes; performs anomaly removal, missing data completion, and standardization on the time-series data to construct derived features; inputs historical feature sequences into an attention-based bidirectional time-series network to obtain time-series representations, inputs these representations into a neural fuzzy inference network to output future water demand predictions and corresponding prediction intervals, and obtains confidence scores from the prediction intervals, where the confidence score is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing; inputs the water demand predictions, confidence scores, and crop stress into the first layer of fuzzy inference to obtain the basic irrigation duration; inputs soil water potential and post-irrigation feedback into the second layer of fuzzy inference to obtain valve fine-tuning amounts, where the post-irrigation feedback is the rate of increase in soil moisture content within a preset time window after the last irrigation; and fuses these to generate irrigation commands and drive the execution equipment, wherein the irrigation commands at least include the irrigation start time, the final irrigation duration, and the valve opening sequence. The beneficial effects include:

[0030] 1. Improve prediction reliability: By integrating a bidirectional time-series network with attention and a neural fuzzy inference network, the predicted water demand value and prediction interval are output, and a confidence level is constructed from the interval, so that the prediction results have an uncertainty representation; when there are weather fluctuations or data quality decline, more robust irrigation decisions can be made based on the confidence level, reducing the risk of over-irrigation or under-irrigation.

[0031] 2. Enhance decision-making feasibility and safety: A two-layer fuzzy reasoning is adopted. The first layer generates the basic irrigation duration, and the second layer combines soil water potential and post-irrigation feedback to generate valve fine-tuning amount, forming an irrigation instruction that includes the start time, final duration and valve opening sequence; and triggers the reduction of opening or valve closure protection when the water potential reaches the safety threshold, improving execution consistency and operational safety.

[0032] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0034] Figure 1 A flowchart illustrating an algorithm-based precision irrigation method for crops, as shown in an exemplary embodiment of the present invention;

[0035] Figure 2 This is a schematic diagram of the structure of a precision crop irrigation system based on algorithm fusion, which is an exemplary embodiment of the present invention. Detailed Implementation

[0036] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention and not for limiting the scope of protection of the present invention.

[0037] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0038] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0039] Example 1:

[0040] A precision crop irrigation method based on algorithm fusion, such as Figure 1 As shown, it includes:

[0041] Collect time-series data on meteorology, soil, crop physiology, and irrigation processes;

[0042] The time-series data is subjected to anomaly removal, missing data completion, and standardization, and derived features are constructed.

[0043] The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation. The input is then input into a neural fuzzy inference network to output the predicted value of future water demand and the corresponding prediction interval. The confidence level is obtained from the prediction interval. The confidence level is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing it.

[0044] The water demand forecast, confidence level, and crop stress level are input into the first layer of fuzzy inference to obtain the basic irrigation duration;

[0045] The soil water potential and post-irrigation feedback are input into the second layer of fuzzy inference to obtain the valve fine-tuning amount. The post-irrigation feedback is the rate of increase of soil moisture content within a preset time window after the last irrigation.

[0046] The irrigation command is generated and the execution device is driven. The irrigation command includes at least the irrigation start time, the final irrigation duration, and the valve opening sequence.

[0047] Specifically, meteorological data acquisition units, soil data acquisition units, crop physiological data acquisition units, and irrigation process data acquisition units are set up in drip-irrigated fields to generate four types of time-series data.

[0048] The meteorological data includes temperature, humidity, wind speed, radiation, and precipitation; soil data includes soil moisture content at multiple depths, soil temperature, electrical conductivity, and soil water potential; crop physiological data includes canopy infrared temperature data and / or stem flow data; and irrigation process data includes flow rate, pressure, valve opening / closing status, and valve opening degree. Each data acquisition unit is equipped with a unique device identifier, recording its installation location, corresponding field number, and sampling channel number to ensure that subsequent time-series data can be traced back to the specific data acquisition point.

[0049] A small weather station is installed in an unobstructed upwind location within the field. The station integrates a temperature sensor, humidity sensor, wind speed sensor, radiation sensor, and rain gauge. The weather station generates meteorological time-series data frames at a fixed sampling period. Each frame contains at least the sampling time, temperature value, humidity value, wind speed value, radiation value, and cumulative precipitation value. It can also include power supply voltage, signal strength, and sensor self-test status as quality marker fields to identify power supply fluctuations and sensor inaccuracy risks.

[0050] A multi-depth soil probe group is vertically deployed in the crop root activity zone. The probe group includes at least: a multi-depth soil moisture content probe, a multi-depth soil temperature probe, a soil electrical conductivity probe, and a soil water potential probe.

[0051] Specifically, moisture content, temperature, and conductivity probes are deployed at multiple depth layers, and each layer is assigned a depth label. Soil water potential probes are deployed at at least one representative depth layer to reflect changes in root zone water suction. The acquisition terminal polls each depth channel at a preset period to form soil time-series data frames. Each frame includes at least the acquisition time, depth label, moisture content value, soil temperature value, conductivity value, and soil water potential value, along with probe communication status and over-range flags for subsequent anomaly identification and data reliability management.

[0052] A canopy infrared temperature measurement device is fixedly installed above the crop canopy, with its measurement field of view covering the canopy area of ​​representative plants, for outputting canopy infrared temperature time series data; a stem flow sensor is installed on the representative plants for outputting stem flow intensity time series data.

[0053] The canopy infrared temperature data frame should at least include the acquisition time, temperature value, and measurement field of view identifier; the stem flow data frame should at least include the acquisition time, plant number, and stem flow measurement value. To avoid bias caused by individual differences, multiple plants can be selected for parallel acquisition by the stem flow sensor, and the corresponding plant number can be recorded in the data frame to support subsequent aggregation by plant or by group.

[0054] Install flow meters and pressure sensors on the main irrigation line or branch lines to collect flow and pressure time-series data, respectively; read the valve opening and closing status and valve opening degree at the valve actuator.

[0055] The valve's open / closed status can be obtained from feedback by the valve limit switch or controller, while the valve opening degree can be obtained from the valve position encoder, angle sensor, or opening degree feedback potentiometer. During irrigation, the data acquisition unit increases the sampling frequency to capture valve actions, pressure fluctuations, and flow changes; during non-irrigation periods, a lower sampling frequency is used to maintain basic monitoring and reduce data redundancy. Each frame of irrigation process data includes at least the acquisition time, flow rate, pressure value, valve open / closed status, and valve opening degree, along with a valve control command acknowledgment marker to distinguish execution anomalies where "the command has been issued but not yet received."

[0056] Each acquisition unit synchronizes the acquisition time using a unified time base. The acquisition terminal writes a timestamp for each frame of data and packages it according to the categories of "meteorological frame, soil frame, physiological frame, and process frame" to form a time-series data stream that is reported to the edge gateway or central controller.

[0057] The edge gateway aligns and caches data frames from different sources by timestamp, ensuring that meteorological, soil, crop physiological and irrigation process data can be obtained within the same time window, thereby meeting the input requirements for subsequent modeling based on historical feature sequences.

[0058] The collected meteorological, soil, crop physiological, and irrigation data are aligned according to timestamps to form a unified time-series data frame sequence. Each frame contains at least the following fields: collection time, temperature, humidity, wind speed, radiation, soil moisture content (at least one depth layer), soil water potential, canopy infrared temperature (or stem flow), flow rate, pressure, valve status, etc., and records equipment status markers and communication status markers. After alignment, several consecutive frames are combined to form the original sequence within a preset historical time window, which serves as the input for subsequent anomaly detection, interpolation, and standardization.

[0059] Joint anomaly detection is performed on each type of time series field within the preset historical time window. Anomaly removal includes joint detection of adjacent time-to-time variation constraints and outlier detection.

[0060] Variation amplitude constraint detection: The difference value of any field at adjacent time points is constrained and judged. If the variation amplitude of a certain sampling point relative to the previous time point exceeds the preset variation upper limit corresponding to the field, the sampling point is marked as an abnormal change point. The preset variation upper limit can be determined based on the sensor range and field change rate of the field, and can be configured according to the season or growth stage.

[0061] Outlier detection: Within the sliding detection window, outlier detection is performed on the sampled points of this field. If a sampled point deviates significantly from the majority of samples in the window, it is marked as an outlier. Outlier detection can be implemented using discrimination logic based on quantile bands, density, or neighborhood consistency.

[0062] Joint determination: When the same sampling point meets both the conditions of change anomaly and outlier anomaly, it is determined as an anomaly to be removed; when it meets only one of the conditions, it is determined as a suspicious point and the anomaly mark is retained for priority processing in the subsequent interpolation stage.

[0063] The output of anomaly removal is a cleaned sequence with anomaly markers, where the field values ​​of the anomalies to be removed are set to empty, and the missing data completion step is initiated.

[0064] Missing defects in the cleaning sequence caused by communication interruption, sensor failure, or abnormal removal are filled with missing information. Missing information filling includes filling based on temporal interpolation or nearest neighbor interpolation.

[0065] Temporal interpolation: When there are valid sampling points on both sides of the missing value in the same field and the missing span does not exceed the preset length, the missing interval is interpolated in time order to obtain the interpolated value; the interpolation method can be linear interpolation or interpolation based on local trends to maintain the continuity of short-term changes.

[0066] Nearest neighbor imputation: When the gap span is long or there are insufficient effective points on both sides, the nearest neighbor imputation method is adopted. Historical moments that are close to the current moment and have similar meteorological background or irrigation status are selected as nearest neighbor samples. The field values ​​corresponding to the nearest neighbor samples are mapped to the fill value of the current gap. Among them, the selection of nearest neighbor samples is determined at least based on the similarity of time and the similarity of key driving fields.

[0067] The missing completion output is a complete sequence of time-series data frames without any gaps, while retaining the interpolation source marker for reliability control in subsequent modeling stages.

[0068] The completed time-series data frame sequence undergoes standardization, which includes unifying the dimensions based on sliding window statistics. A sliding window is set for each field, and window statistics are calculated for samples within the window. These statistics are then mapped to a unified dimension space, allowing data from different dimensions to be used as model input. The window statistics include at least the window central tendency and the window fluctuation scale, used to eliminate differences in dimensions between different sensors and seasonal drift. The standardized output is a standardized time-series sequence, retaining the window statistics corresponding to each field to support backtracking and online updates.

[0069] After obtaining the standardized time series, derived features are further constructed, including reference evapotranspiration, soil available water content, crop growth process characteristics, and crop stress levels.

[0070] Reference evapotranspiration construction: Reference evapotranspiration is calculated at each time step based on temperature, humidity, wind speed and radiation, and a time series characteristic of reference evapotranspiration is formed; Reference evapotranspiration is used to characterize the potential water consumption intensity driven by meteorology and is correlated with subsequent water demand forecasts;

[0071] Soil available water content construction: The soil moisture content at each depth is converted and combined with the soil field holding capacity and wilting coefficient to obtain the available water content that can be used by crops in the root zone; when there are multiple depth layers, the available water content in the root zone is further aggregated to generate the temporal characteristics of the available water content in the root zone.

[0072] Crop growth process feature construction: Growth process features are generated based on time series and crop planting and management information to characterize the impact of crop growth stage changes on water demand; these features can be expressed in the form of accumulated temperature, number of days after transplanting, or stage coding, and are synchronized with historical feature sequences.

[0073] Crop stress quantity construction: The difference between canopy infrared temperature and meteorological temperature is calculated at the same time, and the difference is standardized to obtain the crop stress quantity; the crop stress quantity is used to characterize the degree of crop transpiration restriction or water shortage stress, and serves as one of the input variables for subsequent irrigation decisions;

[0074] The output of the derived feature construction is: a derived feature vector is formed at each sampling time, and the derived feature vectors are arranged in chronological order to form a historical feature sequence for subsequent prediction and decision-making modules to call.

[0075] Within a preset historical time window, the multi-source derived features are time-aligned according to the sampling time to form a historical feature sequence. The multi-source derived features include at least reference evapotranspiration, soil available water content, crop growth process characteristics, and crop stress, and may include pre-processed basic fields (such as soil water potential and irrigation process status) as supplementary input features.

[0076] Specifically, the derived feature vectors of each sampling moment within the historical time window are arranged chronologically to obtain the historical feature sequence. When the sampling frequencies of different sensors are inconsistent, they are first mapped to the same set of sampling moments according to the unified timestamp alignment rule. Missing parts are filled with values ​​that are padded with preceding or interpolated values ​​to ensure that the historical feature sequence has complete feature dimensions at each moment.

[0077] To generate water demand forecasts and corresponding forecast intervals for a predetermined number of days in the future, meteorological forecast features for the predetermined number of days in the future are obtained at the forecast time and input together with historical feature sequences.

[0078] Specifically, the meteorological forecast features include at least predicted values ​​of temperature, humidity, wind speed, and radiation for a predetermined number of days in the future, arranged in a forecast feature sequence in the order of "the first forecast day, the second forecast day, ... the predicted day of the predetermined number of days in the future". When using combined input, the historical feature sequence is used as the historical input, and the meteorological forecast features are used as the forward input, enabling the forecast model to simultaneously utilize historical variation patterns and future external driving conditions to output water demand forecast results for a predetermined number of days in the future.

[0079] The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation.

[0080] Specifically, the bidirectional temporal network encodes the feature sequence within the historical time window in both directions to obtain the hidden state sequence corresponding to each moment; the attention mechanism assigns weights to the hidden state sequence, enabling the model to give higher weights to moments that contribute more significantly to changes in water demand; and the hidden states are aggregated based on the weights to obtain a temporal representation that characterizes the relationship between historical water change patterns and environmental drivers.

[0081] The temporal representation serves as one of the main inputs to the subsequent neural fuzzy inference network.

[0082] When predicting water demand for a predetermined number of days in the future, the time series characteristics are combined with future weather forecast features to form a composite feature.

[0083] Specifically, for each forecast day, the meteorological forecast features corresponding to that forecast day are combined with the time-series representation to obtain the spliced ​​features for that forecast day; or the meteorological forecast features for a preset number of future days are combined with the time-series representation to form a unified spliced ​​feature, and the corresponding results are output according to the forecast day in the subsequent inference network. The spliced ​​features are used to simultaneously express historical pattern information and future meteorological driving information, improving the sensitivity of water demand forecasts to future changes.

[0084] The time-series representation and its concatenated features are input into a neural fuzzy inference network, which outputs the predicted future water demand and the corresponding prediction interval. The neural fuzzy inference network includes an input fuzzification layer, a rule activation layer, a normalization weighting layer, and a defuzzification layer.

[0085] Input fuzzification layer: Based on trainable membership functions, the temporal representation and its spliced ​​features are fuzzified, and continuous features are mapped into multiple fuzzy membership components to form a fuzzy input expression that can participate in rule reasoning; the membership function parameters can be adaptively adjusted during the training phase according to historical irrigation data and water demand labels;

[0086] Rule activation layer: Based on a preset or trainable fuzzy rule library, rule matching is performed on each fuzzy input component to obtain the activation intensity of each rule; among them, the rule expression is used to characterize the relationship between historical water status, future meteorological drivers, and water demand changes;

[0087] Normalized weighted layer: The activation intensity of each rule is normalized and the rule outputs are weighted and fused to obtain an intermediate quantity of the comprehensive reasoning output, so as to reduce the impact of differences in scale of different rules on the output;

[0088] Defuzzing layer: Defuzzifies the intermediate values ​​and outputs the predicted future water demand, the lower bound of the prediction interval, and the upper bound of the prediction interval, respectively; where the lower bound of the prediction interval and the upper bound of the prediction interval together form the corresponding prediction interval.

[0089] When the prediction target is a preset number of days in the future, the neural fuzzy inference network can output the corresponding predicted value, lower bound value and upper bound value for each prediction day, thereby obtaining the water demand prediction sequence and its interval sequence for the preset number of days in the future.

[0090] After obtaining the prediction interval, a confidence score is generated from it. Specifically, the prediction interval width is calculated based on the upper and lower bounds of the prediction interval. A monotonic mapping is performed on the prediction interval width to ensure it falls within a preset comparable scale. The reciprocal of the mapped interval width is then taken and normalized to obtain the confidence score, ensuring it falls within a preset range. In this way, the confidence score decreases as the prediction interval width increases and increases as the prediction interval width decreases, thus providing a quantitative input for subsequent irrigation decisions regarding prediction uncertainty.

[0091] At each decision-making moment, an input vector for the first layer of fuzzy inference is constructed based on the output of the prediction module and the root zone moisture state. The input includes at least:

[0092] Water demand gap: Constructed by the difference between the predicted future water demand and the available soil water content. Specifically, the available soil water content at the current moment is read and compared with the predicted future water demand to obtain the difference representing the "predicted water demand gap". The larger the difference, the more insufficient the available water in the root zone at present.

[0093] Water demand trend: Constructed from the trend of future water demand forecasts over a predetermined number of days. Specifically, based on the future water demand forecast sequence, the direction and intensity of change between adjacent forecast days are extracted to form a trend, which reflects the direction of future water consumption increase or decrease.

[0094] Confidence level: The confidence level, obtained by normalizing the inverse of the prediction interval width, is used as a measure of uncertainty. The lower the confidence level, the wider the prediction interval and the stronger the prediction uncertainty.

[0095] Crop stress quantity: The stress quantity obtained by standardizing the difference between canopy infrared temperature and meteorological temperature is used to characterize the degree of crop water shortage stress;

[0096] After aligning the above inputs at the same decision time, a first-layer fuzzy inference input vector is formed and passed to the first-layer fuzzy inference engine.

[0097] The first-layer fuzzy inference engine fuzzifies each input variable, mapping continuous variables to several membership degree sets. Specifically, it sets membership intervals of "small gap, medium gap, large gap" for water demand difference, "decreasing, stable, increasing" for water demand trend, "high, medium, low" for confidence level, and "light, moderate, heavy" for crop stress, enabling inputs to participate in rule matching in the form of membership degrees. The boundaries of the membership intervals can be configured based on historical operating data and crop type, and can be self-tuned during the update phase.

[0098] The first-layer fuzzy inference engine calculates the basic irrigation duration based on the rule base. The rule base contains at least the following types of rule logic:

[0099] Water demand gap dominant rule: When the water demand difference corresponds to "large gap", the crop stress corresponds to "severe", and the water demand trend corresponds to "rising", the rule of increasing the duration of basic irrigation is triggered to prioritize the mitigation of water shortage risk;

[0100] Trend correction rule: When the water demand trend corresponds to "decreasing" or "stable", the basic irrigation duration is reduced to avoid over-irrigation when water consumption decreases in the future.

[0101] Confidence suppression rule: When the confidence level corresponds to "low", the conservative strategy rule is triggered to suppress the basic irrigation duration, so that the basic irrigation duration is reduced relative to the conventional inference result, in order to suppress the risk of over-irrigation caused by prediction uncertainty.

[0102] The first-layer fuzzy inference engine performs weighted aggregation of the outputs of all triggering rules and performs defuzzification to obtain the basic irrigation duration as the first-layer output.

[0103] When the confidence level is below a preset threshold or is determined to be dominated by "low" membership, in addition to reducing the basic irrigation duration, the basic irrigation duration is further divided into multiple segments for execution. Specifically, the basic irrigation duration is divided into at least two segments, and a preset interval is set for adjacent segments. After each segment is executed, the short-term response of soil water potential and soil moisture content is read. If the response shows that the root zone water replenishment has reached the target range, the execution of the subsequent segment is terminated; if the response is still insufficient, the next segment is executed. Through this segmented strategy, when the prediction uncertainty is large, the target water replenishment is approached segment by segment in a "small steps, quick progress" manner, reducing the probability of deep seepage and over-irrigation caused by a single long-duration irrigation.

[0104] The basic irrigation duration is used as the output of the first-layer fuzzy inference, along with a conservative flag related to the confidence level and a segmented execution flag. The basic irrigation duration and flag are then passed to the second-layer fuzzy inference and instruction fusion module, which combines soil water potential and post-irrigation feedback to generate valve fine-tuning amounts and form the final irrigation instruction. The segmented execution flag guides the instruction fusion module in generating the segmented valve opening sequence and the start and end times.

[0105] In each irrigation control cycle, the input variables for the second-layer fuzzy inference are acquired, including the real-time value of the root zone soil water potential and post-irrigation feedback. The real-time value of the root zone soil water potential is the measurement value collected by the soil water potential sensor within the current control cycle, aggregated by representative points or multiple points in the root zone. Post-irrigation feedback is defined as the rate of increase of soil moisture content within a preset time window after the last irrigation. Specifically, the end time of the last irrigation is used as the starting point of the time window. Soil moisture content sequences are continuously collected within the preset time window, and the ratio of the increase in soil moisture content relative to the starting point to the time length is calculated to obtain the rate of increase. A lower rate of increase indicates a weaker infiltration response after irrigation, while a higher rate of increase indicates a stronger response.

[0106] The second-layer fuzzy inference engine fuzzifies both the real-time soil water potential in the root zone and the post-irrigation feedback. Specifically, it sets "slightly dry," "moderately wet," and "slightly moist" membership intervals for the real-time soil water potential in the root zone to characterize the water suction level; and sets "weak response," "medium response," and "strong response" membership intervals for the post-irrigation feedback to characterize the actual arrival efficiency of irrigation water replenishment. The boundaries of each membership interval can be preset according to soil type, irrigation method, and crop root zone characteristics, and can be adjusted in subsequent self-tuning stages.

[0107] The second-layer fuzzy inference engine outputs valve fine-tuning amounts based on the rule base. These valve fine-tuning amounts represent the adjustment magnitude of the valve opening sequence. The rule base includes at least the following rule types:

[0108] Increased opening rule for drought and weak response: When the soil water potential in the root zone corresponds to "dry" and the post-irrigation feedback corresponds to "weak response", output positive valve fine adjustment amount to increase valve opening or extend effective water supply time to compensate for insufficient infiltration efficiency or water supply not reaching the root zone.

[0109] Reduce opening rule for slightly wet or nearly saturated soil: When the soil water potential in the root zone corresponds to "slightly wet", output a negative valve fine adjustment amount to reduce the valve opening and avoid over-irrigation and deep seepage.

[0110] Strong response suppression rule: When the post-irrigation feedback corresponds to "strong response", even if the soil water potential in the root zone is "moderate" or "slightly dry", a small positive fine adjustment or a negative fine adjustment will still be output to avoid repeated high-intensity watering in the short term.

[0111] The second-layer fuzzy inference engine aggregates and defuzzifies the output of the triggering rules to obtain the valve fine-tuning amount.

[0112] The valve fine-tuning amount is converted into an adjustment coefficient, which is then used to obtain the final irrigation duration. Specifically, the valve fine-tuning amount is converted into an adjustment coefficient through a preset monotonic mapping, such that the adjustment coefficient increases as the valve fine-tuning amount increases and decreases as the valve fine-tuning amount decreases; the adjustment coefficient is used to characterize the amplification or reduction of the basic irrigation duration.

[0113] After obtaining the adjustment coefficient, the basic irrigation duration is combined with the adjustment coefficient to generate the final irrigation duration, which can be adaptively corrected according to the root zone water potential and post-irrigation response.

[0114] A valve opening sequence is generated based on the basic opening sequence and the valve fine-tuning amount, forming an irrigation command. Specifically, the basic opening sequence is obtained by dividing the control cycle corresponding to the basic irrigation duration, representing the basic valve opening within each control cycle; the valve fine-tuning amount is superimposed on the basic opening sequence as the opening adjustment range to obtain the valve opening sequence. The irrigation command includes at least the irrigation start time, the final irrigation duration, and the valve opening sequence, and can carry a "fine-tuning source flag" and a "safety protection flag" for execution-end backtracking and fault diagnosis.

[0115] During the execution of the valve opening sequence, the real-time value of the root zone soil water potential is continuously monitored and compared with a preset safety threshold. When the root zone soil water potential reaches the preset safety threshold, a protective action is triggered. Specifically, if the root zone soil water potential is detected to reach or exceed the preset safety threshold, the valve opening in subsequent control cycles is immediately reduced, and a valve closing command is issued if necessary to terminate irrigation and prevent over-irrigation; at the same time, the trigger time, the trigger water potential value, and the valve opening at that time are recorded for subsequent effect scoring and rule self-tuning.

[0116] The second-layer output includes valve fine-tuning amount, adjustment coefficient, final irrigation duration and valve opening sequence, which are sent to the execution equipment as the core parameters of the final irrigation command. After the execution is completed, the soil moisture content within the post-irrigation time window is collected and the post-irrigation feedback is calculated for the next round of second-layer inference. At the same time, the water potential trajectory and the triggering of protection actions in this execution are written into the operation log to provide a basis for subsequent effect scoring and model rule updates.

[0117] After reasoning about the basic irrigation duration and valve fine-tuning, the instruction fusion stage begins. Specifically, the current decision time is used as the candidate irrigation start time, and the final irrigation duration is obtained by combining the adjustment coefficients corresponding to the basic irrigation duration and valve fine-tuning. Simultaneously, a valve opening sequence is generated based on the basic opening sequence and valve fine-tuning. The irrigation instruction includes at least: irrigation start time, final irrigation duration, and valve opening sequence; it may also include field number, valve number, control cycle length, instruction version number, and safety protection flags for execution end identification and log traceability. The generated irrigation instruction is written to the edge-side cache and simultaneously uploaded to cloud storage.

[0118] The irrigation command is issued to the execution equipment, which includes at least a valve actuator and a water pump or fertigation equipment linked to it. During execution, the execution equipment adjusts the valve opening according to the valve opening sequence in each control cycle, aligns the execution timeline with the irrigation start time, and continues execution until the final irrigation duration ends or is prematurely terminated by a protection action.

[0119] During execution, irrigation process data such as flow rate, pressure, valve opening confirmation, and valve opening / closing status are collected synchronously, and execution deviation records are generated by comparing the commanded opening degree with the confirmed opening degree. When valve action is not in place, pressure is abnormal, or flow rate is abnormal, the abnormality is marked and written into the irrigation log for subsequent explanation and correction during performance evaluation.

[0120] The end of irrigation is taken as the starting point of the post-irrigation time window. Within the preset time window, soil moisture content at multiple depths, soil water potential in the root zone, and crop stress (obtained by standardization of the difference between canopy infrared temperature and meteorological temperature) are continuously collected.

[0121] The collected post-irrigation response sequence is used to calculate the adequacy of water replenishment, the uniformity of water distribution, and the mitigation trend of crop stress. It is also stored in association with the irrigation command and execution process data to form a complete "irrigation-response" closed-loop sample.

[0122] An effectiveness score is also included, which is determined by the adequacy of water replenishment, the uniformity of water distribution, and the trend of crop stress relief within a preset time window after irrigation.

[0123] Water replenishment adequacy: Based on the increase in root zone soil moisture content relative to the end of irrigation within the post-irrigation time window, combined with the target water replenishment range, it is determined whether the water replenishment has achieved the expected results; the closer the water replenishment increment is to the target range, the higher the adequacy.

[0124] Moisture distribution uniformity: determined based on the degree of difference in soil moisture content at multiple depths or points within the post-irrigation time window. The smaller the difference, the more uniform the moisture distribution in the root zone, and the higher the uniformity.

[0125] Crop stress mitigation trend: Based on the direction and magnitude of the decrease in crop stress within the post-irrigation time window, the mitigation trend is better when the stress level continues to decrease and the magnitude of the decrease is more significant.

[0126] The above three results are combined to obtain the effect score of this irrigation. The effect score, along with the corresponding anomaly markers and execution deviation records, are written into the sample record for subsequent learning and updates.

[0127] A rolling sample set is constructed based on the effect score. Specifically, each irrigation cycle generates "historical feature sequence, weather forecast features, water demand prediction output, confidence level, irrigation instructions, execution process data, post-irrigation response sequence, and effect score" as a sample and writes it into the sample pool. The sample pool uses a rolling strategy to retain samples from the most recent cycles or the most recent number of irrigations, and adds time labels, crop growth stage labels, and soil type labels to each sample to support scenario-based updates and retrospective analysis.

[0128] When preset trigger conditions are met (such as the number of new samples in the sample pool reaching a threshold, the effect score being consistently low, or the execution deviation being significant), the prediction model is incrementally updated.

[0129] Specifically, the incremental update takes the rolling sample set as input, uses historical feature sequences and meteorological forecast features as model input, and uses the water demand error or prediction deviation obtained by back-calculation of post-irrigation response as the basis for update. The parameters of the attention-based bidirectional time series network and the neural fuzzy inference network are incrementally adjusted so that the model can gradually adapt to the drift in water demand relationship caused by changes in soil infiltration, changes in facility status and changes in growth stage.

[0130] Simultaneously, self-tuning is performed on the rule weights and membership function parameters of the first and second layer fuzzy inference. Specifically, using the effect score as a feedback signal: when the effect score is low and exhibits the characteristic of "insufficient water replenishment", the weight of rules related to "drought / weak response / severe stress" is increased or their membership function coverage is adjusted, making the inference output more inclined to increase irrigation; when the effect score is low and exhibits the characteristics of "over-irrigation / uneven distribution / frequent water potential triggering protection", the weight of rules related to "wetness / strong response / low confidence and conservatism" is increased or their membership function coverage is adjusted, making the inference output more inclined to reduce irrigation and decrease the opening degree.

[0131] This self-tuning mechanism enables the fuzzy inference engine to maintain a stable and controllable control style across different plots and stages.

[0132] Example 2:

[0133] Please see Figure 2 This exemplary precision irrigation system for crops based on algorithm fusion, used to implement the aforementioned precision irrigation method for crops based on algorithm fusion, includes: a processor, a memory, and a communication interface. The memory stores a program that can be executed by the processor, and the processor executes the program to implement the following functional units:

[0134] Data acquisition module: Collects time-series data on meteorology, soil, crop physiology, and irrigation processes;

[0135] Data processing module: performs anomaly removal, missing data completion, and standardization on time-series data, and constructs derived features;

[0136] Fusion prediction module: The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation. The input is then input into a neural fuzzy inference network to output the predicted value of future water demand and the corresponding prediction interval. The confidence level is obtained from the prediction interval. The confidence level is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing it.

[0137] Duration Decision Module: Input the water demand forecast, confidence level, and crop stress into the first-level fuzzy inference to obtain the basic irrigation duration;

[0138] Valve position fine-tuning module: The soil water potential and post-irrigation feedback are input into the second layer of fuzzy inference to obtain the valve fine-tuning amount. The post-irrigation feedback is the rate of increase of soil moisture content within a preset time window after the last irrigation.

[0139] Command-driven module: Integrates and generates irrigation commands and drives the execution equipment. The irrigation commands include at least the irrigation start time, the final irrigation duration, and the valve opening sequence.

[0140] It should be noted that the algorithm-fusion-based precision irrigation system for crops provided in the above embodiments and the algorithm-fusion-based precision irrigation method for crops provided in the above embodiments belong to the same concept. The specific methods by which each module and unit performs its operations have been described in detail in the method embodiments and will not be repeated here. In practical applications, the algorithm-fusion-based precision irrigation system for crops provided in the above embodiments can be configured to have different functional modules perform the functions as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0141] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A precision crop irrigation method based on algorithm fusion, characterized in that, include: Collect time-series data on meteorology, soil, crop physiology, and irrigation processes; The time-series data is subjected to anomaly removal, missing data completion, and standardization, and derived features are constructed. The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation. The input is then input into a neural fuzzy inference network to output the predicted value of future water demand and the corresponding prediction interval. The confidence level is obtained from the prediction interval. The confidence level is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing it. The water demand forecast, confidence level, and crop stress level are input into the first layer of fuzzy inference to obtain the basic irrigation duration; The soil water potential and post-irrigation feedback are input into the second layer of fuzzy inference to obtain the valve fine-tuning amount. The post-irrigation feedback is the rate of increase of soil moisture content within a preset time window after the last irrigation. The irrigation command is generated and the execution device is driven. The irrigation command includes at least the irrigation start time, the final irrigation duration, and the valve opening sequence.

2. The method for precision crop irrigation based on algorithm fusion according to claim 1, characterized in that, Meteorological data include temperature, humidity, wind speed, radiation, and precipitation; soil data includes soil moisture content at multiple depths, soil temperature, electrical conductivity, and soil water potential; crop physiological data includes canopy infrared temperature data and / or stem flow data; irrigation process data includes flow rate, pressure, valve opening and closing status, and valve opening degree.

3. The method for precision crop irrigation based on algorithm fusion according to claim 2, characterized in that, Anomaly removal includes joint detection based on constraints on the magnitude of changes in adjacent time intervals and outlier detection; missing data completion includes completion based on temporal interpolation or nearest neighbor interpolation; standardization processing includes dimensional unification based on sliding window statistics.

4. The precision crop irrigation method based on algorithm fusion according to claim 1, characterized in that, Derived characteristics include reference evapotranspiration, soil available water content, crop growth process characteristics, and crop stress; among them, reference evapotranspiration is calculated from temperature, humidity, wind speed, and radiation; soil available water content is calculated from soil moisture content combined with soil field capacity and wilting coefficient; and crop stress is obtained by standardizing the difference between canopy infrared temperature and meteorological temperature.

5. The method for precision crop irrigation based on algorithm fusion according to claim 4, characterized in that, The historical feature sequence is composed of multi-source derived features within a preset historical time window, aligned by time. When outputting the future water demand forecast and the corresponding forecast interval, the meteorological forecast features for the next preset number of days are input together with the historical feature sequence to generate the water demand forecast and the corresponding forecast interval for the next preset number of days.

6. The method for precision crop irrigation based on algorithm fusion according to claim 1, characterized in that, The neural fuzzy inference network includes an input fuzzification layer, a rule activation layer, a normalization weighting layer, and a defuzzification layer. The input fuzzification layer fuzzifies the temporal representation and its spliced ​​features based on a trainable membership function. The defuzzification layer outputs the predicted future water demand, the lower bound of the prediction interval, and the upper bound of the prediction interval, forming the corresponding prediction interval.

7. The method for precision crop irrigation based on algorithm fusion according to claim 4, characterized in that, The inputs to the first-layer fuzzy inference include at least the difference between the predicted future water demand and the available soil water content, the future water demand trend, the confidence level, and the crop stress level. The rule base of the first-layer fuzzy inference is set to reduce the basic irrigation duration and / or split the basic irrigation duration into multiple segments for execution when the confidence level decreases, in order to suppress the risk of over-irrigation caused by prediction uncertainty.

8. The method for precision crop irrigation based on algorithm fusion according to claim 1, characterized in that, The inputs to the second layer of fuzzy inference include the real-time value of the root zone soil water potential and post-irrigation feedback. The valve fine-tuning amount represents the adjustment range of the valve opening sequence. The final irrigation duration is obtained by combining the basic irrigation duration with the adjustment coefficient corresponding to the valve fine-tuning amount. The valve opening sequence is obtained by superimposing the valve fine-tuning amount on the basic opening sequence. When the soil water potential reaches the preset safety threshold, the protection action of reducing the opening or closing the valve is triggered.

9. The method for precision crop irrigation based on algorithm fusion according to claim 1, characterized in that, An effectiveness score is also set up, which is jointly determined by the sufficiency of water replenishment, the uniformity of water distribution, and the trend of crop stress relief within a preset time window after irrigation. A rolling sample set is constructed based on the effectiveness score, the prediction model is incrementally updated, and the weights and membership function parameters of the fuzzy rules are self-tuned.

10. A precision crop irrigation system based on algorithm fusion, used to implement the precision crop irrigation method based on algorithm fusion as described in any one of claims 1-9, characterized in that, include: The processor, memory, and communication interface are included. The memory stores programs that can be executed by the processor, and the processor executes the programs to implement the following functional units: Data acquisition module: Collects time-series data on meteorology, soil, crop physiology, and irrigation processes; Data processing module: performs anomaly removal, missing data completion, and standardization on time-series data, and constructs derived features; Fusion prediction module: The historical feature sequence is input into a bidirectional temporal network with attention to obtain a temporal representation. The input is then input into a neural fuzzy inference network to output the predicted value of future water demand and the corresponding prediction interval. The confidence level is obtained from the prediction interval. The confidence level is obtained by taking the reciprocal of the prediction interval width after monotonic mapping and normalizing it. Duration Decision Module: Input the water demand forecast, confidence level, and crop stress into the first-level fuzzy inference to obtain the basic irrigation duration; Valve position fine-tuning module: The soil water potential and post-irrigation feedback are input into the second layer of fuzzy inference to obtain the valve fine-tuning amount. The post-irrigation feedback is the rate of increase of soil moisture content within a preset time window after the last irrigation. Command-driven module: Integrates and generates irrigation commands and drives the execution equipment. The irrigation commands include at least the irrigation start time, the final irrigation duration, and the valve opening sequence.

Images