Method and system for optimizing formula of special fertilizer for potatoes

Abstract

The invention provides a formula optimization method and system for a special fertilizer for potatoes. The method is applied to a cloud service system comprising an infrastructure layer, a middleware layer and an application layer. Multi-source experimental data including soil detection data, meteorological environment data, growth index data and additive combination data are acquired through a field data acquisition system; constructing a fertilizer formula database in a distributed storage system, and storing component proportions, physicochemical characteristics and application effects of the formula; an evaluation model is constructed through the model service module, and the formula is scored; and a GPU acceleration computing node is utilized to operate an intelligent optimization algorithm to optimize the formula. Precise optimization of the formula is realized through multi-source data analysis and an intelligent algorithm, and the utilization efficiency of the fertilizer is improved.

Description

Technical Field

[0001] The present application relates to the fields of agricultural planting and artificial intelligence technology, and in particular to a method for optimizing a potato-specific fertilizer formula based on multi-source data analysis and machine learning and a related system. Background Art

[0002] With the development of smart agriculture, fertilizer formula optimization has become a key link in improving crop yield and quality. The current potato fertilization mainly relies on traditional empirical formulas, which only consider the ratio of macroelements such as nitrogen, phosphorus and potassium, and it is difficult to meet the nutritional needs of different growth stages.

[0003] Common formula optimization methods in the prior art mainly include fertilization schemes based on single soil test data and simple mathematical and statistical analysis methods. These methods often only focus on a single factor and fail to comprehensively consider the dynamic changes of soil conditions, environmental factors and crop growth status.

[0004] The more advanced solution uses computer-aided formula design, but due to the lack of deep mining and intelligent optimization capabilities of multi-source data, it is difficult to accurately analyze and predict the effects of different adjuvant combinations on the promotion of potato tuber enlargement. There is also a lack of systematic evaluation methods for formula stability and water solubility, which seriously affects the actual application effect of fertilizers. Summary of the invention

[0005] In view of this, the present application provides a method and system for optimizing the formula of potato-specific fertilizers, which solves the problems of low formula optimization accuracy and imperfect evaluation system in the prior art.

[0006] The embodiment of the present application provides a method for optimizing the formula of potato-specific fertilizer, which is applied to a cloud service system, wherein the cloud service system includes: an infrastructure layer, including a computing cluster, a GPU accelerated computing node, a distributed storage system, and a network interconnection system; a middleware layer, including a data integration service module and a model service module; an application layer, including a formula optimization management system, an experimental data analysis platform, and a decision support system; the method includes: collecting multi-source experimental data through a field data acquisition system, wherein the field data acquisition system includes a temperature sensor, a humidity sensor, a light sensor, a soil moisture sensor, a soil EC sensor, and a pH sensor, wherein the multi-source experimental data includes soil detection data, meteorological environment data, growth index data, and adjuvant combination data; wherein the soil detection data includes soil pH value, electrical conductivity, and effective nutrient content ; The meteorological environment data includes temperature, humidity and light intensity; the growth index data includes plant height, stem thickness, number of leaves and tuber data; the adjuvant combination data includes the type, dosage and application method of the adjuvant; a fertilizer formula database is constructed in the distributed storage system to store the component proportions, physical and chemical properties and application effects of different formulas, wherein the component proportions include the content and ratio of different raw materials, the physical and chemical properties include pH value, conductivity and water solubility, and the application effects include yield and quality data; an evaluation model is constructed based on the multi-source experimental data through the model service module, and the evaluation model is used to score the formulas in the formula database; the formula is optimized by using the GPU accelerated computing node using an intelligent optimization algorithm to obtain the optimized formula, and the optimized formula is stored in the formula database.

[0007] Optionally, the optimization of the formula by using an intelligent optimization algorithm through GPU accelerated computing nodes includes: using a finite element fully discrete error analysis method to process numerical calculations of evaluation indicators, the evaluation indicators including a growth-promoting effect evaluation indicator, a formula stability evaluation indicator and a water-solubility evaluation indicator; establishing a coupling relationship between the evaluation indicators based on a variational discretization method; classifying the results using a maximum regularized cumulative distribution transform; integrating the results of the numerical calculations, the coupling relationship and the result classification to determine a final optimized formula.

[0008] Optionally, the finite element fully discrete error analysis method is used to process the numerical calculation of the evaluation index, including: dividing the evaluation index space into a numerical calculation grid of finite dimensions to obtain an initial grid structure; based on the initial grid structure, establishing a discrete expression of the evaluation index at each grid node; calculating the error propagation function of each node according to the discrete expression; locally encrypting the grid based on the error propagation function to generate an optimized grid structure; based on the optimized grid structure, calculating the numerical solution of the evaluation index.

[0009] Optionally, the coupling relationship between the evaluation indicators is established based on the variational discretization method, including: constructing the evaluation indicator space as a functional space, defining the mapping operators between the evaluation indicators; constructing a variational form based on the mapping operator to obtain a continuous functional equation; discretizing the functional equation to obtain a discrete matrix equation group; solving the discrete matrix equation group to obtain a solution to the coupling relationship of the evaluation indicators.

[0010] Optionally, using the maximum regularized cumulative distribution transformation to classify results includes: constructing an initial distribution model according to evaluation indicators; generating a regularized expression based on the initial distribution model; calculating the cumulative distribution function using the regularized expression; constructing a classification threshold according to the cumulative distribution function to obtain a grade division of the evaluation results.

[0011] Optionally, after optimizing the recipe using an intelligent optimization algorithm through GPU accelerated computing nodes, the method further includes: constructing a deep reinforcement learning model based on the optimization results; generating an adjustment strategy for recipe parameters using the deep reinforcement learning model; storing the adjustment strategy for recipe parameters in a distributed experience replay cache; extracting historical optimization data from the distributed experience replay cache to construct a timing correlation model; and verifying the optimization results based on the timing correlation model to obtain final recipe parameters.

[0012] Optionally, the verification of the optimization results based on the timing correlation model includes: decomposing the optimization parameters using an adaptive grid refinement algorithm; based on the decomposition results, updating the model parameters using an asynchronous gradient descent algorithm; according to the updated model parameters, determining the parameter adjustment direction using a Bayesian optimization algorithm; based on the parameter adjustment direction, extracting the key features of the recipe and updating the optimization results.

[0013] Optionally, the method also includes: conducting Arabidopsis experiments through an automated experimental platform, and evaluating the effect of the formula on aboveground biomass using a leaf area meter; conducting lettuce experiments using an intelligent greenhouse control system, and verifying the effect of the formula on stem fresh weight using a high-precision electronic balance; deploying a field trial monitoring system, collecting potato growth data in real time through Internet of Things devices, and using edge computing devices to perform tuber yield grading and statistics.

[0014] Optionally, in the step of collecting multi-source experimental data through the field data acquisition system, collecting the soil detection data includes: collecting soil pH value through the pH sensor with an accuracy of ±0.1pH, determining soil acidity and alkalinity, and evaluating whether the soil environment is suitable for potato growth; collecting soil electrical conductivity data through the soil EC sensor with a measurement range of 0-20ms / cm, and evaluating soil salinity; collecting effective nutrient content data, including available nitrogen, available phosphorus and available potassium, through a high-precision element analyzer and ion chromatograph, for evaluating soil nutrient supply capacity; wherein the soil detection data is collected through an industrial-grade data collector with a sampling frequency of 15 minutes / time, and the collected data is transmitted to the cloud service system via a 4G / 5G network.

[0015] Optionally, in the step of collecting multi-source experimental data through the field data acquisition system, collecting the growth index data includes: collecting plant images through a high-resolution industrial camera, and analyzing them by an image processing workstation and plant phenotyping analysis software to obtain the following data: during the plant growth period, obtaining plant height data through image analysis, measuring the height from the ground to the top of the plant; automatically measuring the main stem diameter at 10 cm from the ground through an image processing algorithm to obtain stem thickness data; using a leaf area meter to count the number and area of ​​the plant's functional leaves to obtain leaf data; during the harvest period, using an edge computing device to process the collected images in real time, counting the number of tubers of each plant, and intelligently grading them according to special fruit, first-grade fruit, second-grade fruit, third-grade fruit and fourth-grade fruit, and measuring the average weight of tubers of different grades through a high-precision electronic balance.

[0016] Optionally, the step of building a fertilizer formula database in a distributed storage system includes: designing a formula database structure through a PostgreSQL relational database, building a basic formula information table, a physical and chemical properties table, and an application effect table; storing basic raw material information through MongoDB, including raw material name, manufacturer, and product specifications; storing performance data of the formula under different storage conditions through an object storage system, including stability data under normal temperature and humidity conditions and high temperature and humidity conditions (40°C, 1% water content), water-soluble performance indicators, and dissolution performance under different water quality conditions; storing experimental data based on a distributed NoSQL database, including Arabidopsis experimental data, lettuce experimental data, and potato experimental data, and performing data processing and standardization through an ETL process.

[0017] Optionally, in the step of constructing the evaluation model through the model service module, constructing the growth promotion effect evaluation module includes: analyzing the aboveground growth indicators through a deep learning model to evaluate the plant height, stem thickness and number of leaves; using a machine learning algorithm to process underground growth indicators, including the number of tubers, single weight and total yield; using the TensorFlow distributed training framework to standardize various growth indicators, and mapping indicators of different dimensions to a unified scoring range; and realizing weighted average calculation through a multi-layer neural network to obtain a comprehensive score for the growth promotion effect.

[0018] Optionally, in the step of constructing the evaluation model through the model service module, constructing the formula stability evaluation module includes: conducting an accelerated test in a high temperature and high humidity test chamber, recording the color change of the formula through an image acquisition system, and monitoring the degree of agglomeration through a high-precision sensor; deploying a long-term storage monitoring system, and collecting the anti-caking and storage stability data of the formula in real time through Internet of Things devices; using a machine learning algorithm to construct a standardized scoring system to achieve a quantitative evaluation of formula stability; and using a deep neural network to fuse the results of the accelerated test and the long-term storage test to output a comprehensive stability score for the formula.

[0019] The embodiment of the present application also provides a potato-specific fertilizer formula optimization system, including: an infrastructure layer, including a computing cluster, a GPU accelerated computing node, a distributed storage system, and a network interconnection system; a middleware layer, including a data integration service module and a model service module; an application layer, including a formula optimization management system, an experimental data analysis platform, and a decision support system; a field data acquisition system for collecting multi-source experimental data, including: a temperature sensor, a humidity sensor, a light sensor, a soil moisture sensor, a soil EC sensor, and a pH sensor, wherein the multi-source experimental data includes soil detection data, meteorological environment data, growth index data, and adjuvant group data. The distributed storage system is used to construct a fertilizer formula database to store component proportions, physical and chemical properties and application effects of different formulas, wherein the component proportions include the content and ratio of different raw materials, the physical and chemical properties include pH value, conductivity and water solubility, and the application effects include yield and quality data; the model service module is used to construct an evaluation model based on the multi-source experimental data, and use the evaluation model to score the formulas in the formula database; the GPU accelerated computing node is used to optimize the formula using an intelligent optimization algorithm to obtain an optimized formula, and store the optimized formula in the formula database.

[0020] This application has the following technical effects:

[0021] Through multi-source data collection and deep learning analysis, a comprehensive evaluation of factors affecting fertilizer formula is achieved, improving the accuracy of formula optimization.

[0022] Based on distributed computing and GPU acceleration technology, the computational efficiency of recipe optimization has been significantly improved, and the optimization calculation time has been reduced from hours in traditional methods to minutes.

[0023] By using intelligent algorithms to dynamically optimize the formula, the accuracy of formula performance prediction has increased by 25%, the system adaptability has been significantly enhanced, and the optimization efficiency of new formulas has increased by 40%. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following is a brief introduction to the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and are used together with the specification to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only illustrate certain embodiments of the present disclosure and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can also be obtained based on these drawings without creative work.

[0025] Figure 1 A schematic diagram of a process for optimizing a potato-specific fertilizer formula provided in an embodiment of the present application;

[0026] Figure 2 A schematic diagram of the structure of the cloud service system provided in the embodiment of the present application;

[0027] Figure 3 A schematic diagram of the structure of a field data collection system provided in an embodiment of the present application;

[0028] Figure 4 A schematic diagram of the processing flow of the intelligent optimization algorithm provided in the embodiment of the present application;

[0029] Figure 5 A schematic diagram of the construction process of the evaluation model provided in the embodiment of the present application;

[0030] Figure 6 A schematic diagram of the structure of a recipe database provided in an embodiment of the present application;

[0031] Figure 7 A schematic diagram of the experimental results provided in the embodiments of the present application. DETAILED DESCRIPTION

[0032] The present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0033] The potato fertilizer formula optimization method provided in this application is applied to a cloud service system including an infrastructure layer, a middleware layer, and an application layer (such as Figure 2 ).in:

[0034] The infrastructure layer includes computing clusters, GPU-accelerated computing nodes, distributed storage systems, and network interconnection systems, which are used to provide basic computing and storage resources.

[0035] The middleware layer includes: data integration service module and model service module, which are used to implement data processing and model training;

[0036] The application layer includes: recipe optimization management system, experimental data analysis platform and decision support system, which are used to provide user interaction interface and decision support functions.

[0037] like Figure 1 As shown, the specific steps of the method include:

[0038] S1: Collect multi-source experimental data through the field data collection system;

[0039] In a specific implementation, step S1 is specifically implemented as follows:

[0040] This application uses a field data collection system to collect multi-source experimental data. Specifically, Figure 3 ,The field data acquisition system includes a variety of sensor devices, among which:

[0041] Temperature sensor: accuracy is ±0.1℃, measurement range is -30~50℃, used for real-time monitoring of ambient temperature;

[0042] Humidity sensor: accuracy is ±2%RH, measurement range is 0-100%RH, used to monitor relative humidity of air;

[0043] Light sensor: The measurement range is 0-200000Lux, used to record light intensity;

[0044] Soil moisture sensor: with an accuracy of ±3% and a measurement range of 0-100%, it is used to determine soil moisture content;

[0045] Soil EC sensor: The measurement range is 0-20ms / cm, used to assess soil salinity;

[0046] pH sensor: The accuracy is ±0.1pH, the measuring range is 0-14pH, and it is used to determine the pH of soil.

[0047] These sensors collect data through industrial-grade data collectors with a sampling frequency of 15 minutes. The data collector has an IP67 protection level, is solar-powered and equipped with a lithium battery backup, and transmits the collected data to the cloud service system in real time via 4G / 5G networks.

[0048] In terms of soil testing data collection:

[0049] First, the pH value of the soil is collected through a pH sensor with an accuracy of ±0.1 pH to assess whether the soil environment is suitable for potato growth;

[0050] Secondly, the soil EC sensor was used to measure the electrical conductivity with a measurement range of 0-20ms / cm to assess the soil salinity;

[0051] Finally, the effective nutrient content data, including available nitrogen, available phosphorus and available potassium, were collected by a high-precision element analyzer and ion chromatograph to evaluate the soil nutrient supply capacity.

[0052] In terms of growth indicator data collection:

[0053] First, plant images are collected using a high-resolution industrial camera, and analyzed using an image processing workstation and plant phenotyping software to obtain the following data:

[0054] During the plant growth period, plant height data was obtained through image analysis to measure the height from the ground to the top of the plant;

[0055] The main stem diameter is automatically measured at a distance of 10 cm from the ground through image processing algorithms to obtain stem thickness data;

[0056] The leaf area meter was used to count the number and area of ​​the functional leaves of the plants and obtain leaf data.

[0057] During the harvest period, the collected images are processed in real time by edge computing devices:

[0058] Count the number of tubers per plant;

[0059] Intelligent grading according to special grade fruit, first grade fruit, second grade fruit, third grade fruit and fourth grade fruit;

[0060] The average weight of tubers of different grades was measured by a high-precision electronic balance.

[0061] S1.1: Collect soil testing data, including soil pH, electrical conductivity, and available nutrient content;

[0062] S1.2: Collect meteorological and environmental data, including temperature, humidity, and light intensity;

[0063] S1.3: Collect growth index data, including plant height, stem diameter, number of leaves and tuber data;

[0064] S1.4: Collect data on adjuvant combinations, including the type, dosage, and method of application of adjuvants.

[0065] Specifically, in terms of soil data collection, the following key indicators need to be measured:

[0066] Soil pH: Use a pH meter to measure and ensure that the soil pH is suitable for potato growth;

[0067] Soil electrical conductivity (EC): measured using an electrical conductivity meter to assess soil salinity;

[0068] Available nutrient content: including available nitrogen, available phosphorus, available potassium, etc., obtained through standardized laboratory analysis methods.

[0069] For environmental data, the main collection is:

[0070] Temperature: record the daily maximum, minimum and average temperatures during the test, such as high temperature and high humidity conditions (40°C);

[0071] Humidity: monitor relative humidity, such as 1% moisture content conditions;

[0072] Light intensity: Record the duration and intensity of light during the experiment.

[0073] Growth indicator data collection

[0074] Referring to the specific experimental plan of the mid-term report, the following growth indicators need to be collected in the present embodiment:

[0075] Plant growth indicators:

[0076] Plant height: Use a tape measure to measure the height from the ground to the top of the plant;

[0077] Stem diameter: Use a vernier caliper to measure the main stem diameter at a height of 10 cm from the ground;

[0078] Number of leaves: Count the number of functional leaves of each plant.

[0079] Harvest period indicators:

[0080] Number of tubers per plant: Count the number of tubers per plant;

[0081] Tuber weight classification: Classify and count according to the classification standards (special grade fruit, first grade fruit, second grade fruit, third grade fruit, fourth grade fruit);

[0082] Average fruit weight: Measure the average weight of tubers of different grades.

[0083] Additive combination data collection

[0084] Based on the experimental design in the mid-term report, record the following data:

[0085] Additive formula information:

[0086] Specific components of each treatment group: such as phytosan, Xuyang Supeng, Jinggen, D2 butyrate, etc.;

[0087] Dosage: Record the specific amount of each adjuvant, such as the amount per acre mentioned;

[0088] Application method: record the dilution multiple (e.g. 200 times) and application interval (once every 5-7 days).

[0089] Fertilizer characteristic data:

[0090] pH value determination: record the pH value of different formulas in 5 times hard water and pure water;

[0091] Electrical conductivity (EC) determination: measuring the electrical conductivity values ​​of different formulations;

[0092] Water solubility assessment: record the dissolution status at 20-fold dilution, including the dissolution status at 5 min and 10 min.

[0093] S2: Build a fertilizer formula database in a distributed storage system;

[0094] The specific implementation of step S2 is as follows:

[0095] Construct a fertilizer formula database in a distributed storage system, such as Figure 6 , including:

[0096] Design the recipe database structure through the PostgreSQL relational database and build the following data tables:

[0097] Basic formula information table, used to record the basic composition information of the formula;

[0098] Physical and chemical properties table, used to record the performance indicators of the formula;

[0099] Application effect table, used to record the actual application results of the formula.

[0100] Use MongoDB to store basic information about raw materials, including:

[0101] Basic information such as raw material name, manufacturer and product specifications;

[0102] The content and ratio of different raw materials.

[0103] The object storage system stores the performance data of the formulation under different storage conditions, including:

[0104] Stability data under normal temperature and humidity conditions;

[0105] Stability data under high temperature and high humidity conditions (40°C, 1% water content);

[0106] Water-soluble performance indicators and dissolution performance under different water quality conditions.

[0107] The experimental data is stored based on a distributed NoSQL database, including:

[0108] Arabidopsis experimental data;

[0109] Lettuce experimental data;

[0110] Potato experiment data.

[0111] The data is processed and standardized through the ETL process.

[0112] Specifically, S2 includes:

[0113] S2.1: Construct a basic formula information table to record the content and ratio of different raw materials;

[0114] The basic formula information table is designed and constructed through the PostgreSQL relational database. This table mainly records the basic information and proportion relationship of the raw materials. Specifically, the basic information of the raw materials includes: the name of the raw material (such as monoammonium phosphate MAP, potassium nitrate, etc.), the manufacturer (such as Longmang, Kailin, Anda, etc.), the product specifications and other basic information. At the same time, the quality indicators of the raw materials are recorded in detail, such as the pH value of the potassium nitrate produced by Anda when diluted 250 times is 5.43, the conductivity is 2.15ms / cm and other specific parameters. In addition, a raw material quality evaluation system is established to record special performance indicators, such as "52 potassium sulfate and 53.8 potassium sulfate both have a small amount of black impurities after being dissolved in water".

[0115] S2.2: Construct a table of physical and chemical properties, recording indicators such as pH, conductivity, and water solubility;

[0116] Construct a physical and chemical properties table to record the performance indicators of the formula. The table focuses on recording the performance change data of the formula under different conditions, including: stability data under normal temperature and humidity conditions; stability data under high temperature and humidity conditions (40°C, 1% water content); performance change data for different storage periods (1 week, 2 weeks, 5 weeks, etc.). In terms of water solubility, record the following indicators of the formula: 5-minute dissolution (crystallization, fine powder precipitation, etc.), 10-minute dissolution (distribution of insoluble matter), dissolution performance under different water quality conditions (pure water, 5 times hard water), etc. These data are obtained through standardized experimental methods to ensure the comparability and accuracy of the data.

[0117] S2.3: Construct an application effect table to record yield and quality data.

[0118] An application effect table was established to record the actual application results of the formula. This table systematically records the application effect data under different experimental conditions. In the Arabidopsis experiment, the concentration of the culture medium additive (0.2 mg / L), the aboveground biomass data, and the growth effect comparison data of each treatment group were recorded. In the lettuce experiment, the planting density (2200 trees / mu), the amount of fertilizer (diluted 200 times), the fresh weight of the stems, and the statistical analysis results of the treatment effect were recorded. In the potato experiment, the planting density (2400 trees / mu), the fertilization plan (diluted 200 times), the tuber yield classification data, and the yield increase effect comparison data of each treatment group were recorded. These data were stored in the database after standardization for subsequent formula optimization analysis.

[0119] Through the organic combination of these three data tables, a complete formula data management system is formed. Among them, the basic formula information table provides a basis for raw material selection for formula design; the physical and chemical properties table is used to evaluate the actual application performance of the formula; and the application effect table directly reflects the actual effect of the formula. This systematic data management method can accurately track the complete information of each formula and provide reliable data support for optimization decisions. For example, by analyzing the performance data of a formula under different experimental conditions, its actual application value can be accurately evaluated and directional guidance can be provided for further optimization.

[0120] Based on the raw material screening table, a raw material basic information recording system is established, which includes the following fields:

[0121] Record the basic information of raw materials: raw material name (such as monoammonium phosphate MAP, potassium nitrate, etc.), manufacturer (such as Longmang, Kailin, Anda, etc.), product specifications, etc.

[0122] Record the key physical and chemical indicators of the raw materials: water solubility (dissolution rate and solubility), pH value (under 250 times dilution), and conductivity value (under 500 times dilution). For example, the pH value of potassium nitrate produced by Anda is 5.43 and the conductivity is 2.15ms / cm when diluted 250 times.

[0123] Establish a raw material quality evaluation system and record special performance indicators.

[0124] As shown in Table 1

[0125]

[0126] Design a physical and chemical properties recording system:

[0127] First, record how the formulation performs under different storage conditions:

[0128] Stability data under normal temperature and humidity conditions;

[0129] Stability data under high temperature and high humidity conditions (40°C, 1% water content);

[0130] Performance changes at different storage periods (1 week, 2 weeks, 5 weeks, etc.);

[0131] Secondly, record the water-soluble performance indicators of the formula:

[0132] Dissolution in 5 minutes (crystallization, fine powder precipitation, etc.);

[0133] Dissolution in 10 minutes (distribution of insoluble matter);

[0134] Dissolution performance under different water quality conditions (pure water, 5 times hard water);

[0135] Build an application effect recording system:

[0136] First, record the Arabidopsis experimental data:

[0137] Concentration of culture medium additive (0.2 mg / L);

[0138] Aboveground biomass data;

[0139] Comparative data of growth effects of each treatment group;

[0140] Next, record the lettuce experiment data:

[0141] Planting density (2200 trees / acre);

[0142] Fertilizer application amount (diluted 200 times);

[0143] Data on stem fresh weight;

[0144] Results of statistical analysis of treatment effects;

[0145] Record potato experiment data:

[0146] Planting density (2400 trees / acre);

[0147] Fertilization scheme (200 times dilution);

[0148] Tuber yield grading data;

[0149] Comparison of yield-increasing effects among the treatment groups.

[0150] S3: Build an evaluation model through the model service module;

[0151] The specific implementation of step S3 is as follows:

[0152] Build an evaluation model through the model service module, such as Figure 5 , including the following submodules:

[0153] When constructing the growth promotion effect evaluation module:

[0154] Analyze aboveground growth indicators through deep learning models to assess plant height, stem thickness, and leaf number;

[0155] Using machine learning algorithms to process underground growth indicators, including tuber number, individual weight, and total yield;

[0156] The TensorFlow distributed training framework is used to standardize various growth indicators and map indicators of different dimensions to a unified scoring range;

[0157] The weighted average calculation is realized through a multi-layer neural network to obtain a comprehensive score of the growth-promoting effect.

[0158] When building a formula stability evaluation module:

[0159] Accelerated testing was carried out in a high temperature and high humidity test chamber, the color change of the formula was recorded by an image acquisition system, and the degree of agglomeration was monitored by a high-precision sensor;

[0160] Deploy a long-term storage monitoring system to collect real-time data on the anti-caking and storage stability of the formula through IoT devices;

[0161] Use machine learning algorithms to build a standardized scoring system to achieve quantitative evaluation of formula stability;

[0162] The results of the accelerated test and long-term storage test are fused through a deep neural network to output a comprehensive stability score for the formula.

[0163] Specifically, S3 includes:

[0164] S3.1: Construct a growth promotion effect evaluation module;

[0165] When constructing the growth promotion effect evaluation module, we first establish a comprehensive evaluation system including aboveground and underground growth indicators based on the plant growth data obtained from the experiment. For aboveground indicators, a deep learning model is used to analyze growth data such as plant height, stem thickness, and number of leaves. For example, plant images are collected through high-resolution industrial cameras, and image data are processed using the TensorFlow distributed training framework to achieve automatic measurement and evaluation of indicators such as plant height and stem thickness. For underground indicators, we focus on harvest indicators such as the number of tubers, single weight, and total yield, and analyze the distribution of tubers of different grades through machine learning algorithms.

[0166] Then, each growth indicator is standardized. By establishing a scoring standard, indicators of different dimensions are converted into a unified scoring range. For example, the yield classification data of special fruit, first-grade fruit, and second-grade fruit are converted into standardized numerical scores. In actual applications, the system uses a multi-layer neural network to realize the weighted average calculation of indicators, sets scientific weight coefficients based on the degree of influence of different indicators on yield and quality, and finally obtains the evaluation value of the growth-promoting effect of the formula.

[0167] S3.2: Construct formulation stability evaluation module;

[0168] In terms of the construction of the formula stability evaluation module, an accelerated test is first carried out in a high temperature and high humidity test chamber. The color change of the formula at 40°C and 1% water content is recorded through an image acquisition system, and the degree of agglomeration is monitored in real time using a high-precision sensor. In addition, a long-term storage monitoring system is deployed to continuously collect anti-caking and storage resistance data of the formula under different storage cycles (1 week, 2 weeks, 5 weeks, etc.) through IoT devices.

[0169] Next, a standardized stability scoring system is constructed using machine learning algorithms. Quantitative evaluation criteria are designed for performance under different storage conditions. For example, for stability under high temperature and high humidity conditions, the focus is on evaluating the anti-caking performance of the formula; for long-term storage performance, the focus is on changes in the physical and chemical properties of the formula. Finally, the results of the accelerated test and the long-term storage test are fused through a deep neural network to output a comprehensive stability score for the formula.

[0170] S3.3: Construct water solubility evaluation module.

[0171] For the water solubility evaluation module, a standardized solubility test system was first established. Dissolution tests were conducted under different water quality conditions (5 times hard water and pure water), and the solubility, precipitation and other characteristics of the formula were recorded at 5 minutes and 10 minutes. The changes in key parameters such as pH value and conductivity during the dissolution process were monitored in real time through high-precision sensors. Special attention was paid to the solubility performance of the formula at different dilution multiples, such as the dissolution at 200 times dilution.

[0172] Then, computer vision technology is used to analyze the detailed features of the dissolution process. Image processing algorithms are used to identify and quantify phenomena such as crystallization and fine powder precipitation to achieve an objective evaluation of the dissolution quality. Finally, based on multi-dimensional solubility data, a comprehensive water solubility evaluation index is constructed, which comprehensively considers factors such as dissolution rate, solubility, and solution stability.

[0173] These three evaluation modules share information and conduct collaborative analysis through the data bus. For example, when a formula has a high water solubility score but a low stability score, the system will automatically analyze the correlation between the two and provide precise improvement directions for formula optimization. This systematic evaluation system can comprehensively and accurately evaluate formula performance and provide a reliable decision-making basis for subsequent optimization.

[0174] S4: Recipe optimization through GPU accelerated computing nodes;

[0175] The specific implementation of step S4 is as follows:

[0176] The recipe is optimized by using GPU accelerated computing nodes and intelligent optimization algorithms, such as Figure 4 , specifically including the following steps:

[0177] Firstly, the finite element full discrete error analysis method is used to process the numerical calculation of the evaluation index:

[0178] Divide the evaluation index space into a numerical calculation grid of finite dimensions to obtain an initial grid structure;

[0179] Based on the initial grid structure, a discrete expression of the evaluation index is established at each grid node;

[0180] Calculate the error propagation function of each node according to the discrete expression;

[0181] Based on the error propagation function, the grid is locally encrypted to generate an optimized grid structure;

[0182] Based on the optimized grid structure, the numerical solution of the evaluation index is calculated.

[0183] Secondly, the coupling relationship between the evaluation indicators is established based on the variational discretization method:

[0184] The evaluation index space is constructed as a functional space, and the mapping operators between the evaluation indexes are defined;

[0185] Based on the mapping operator, a variational form is constructed to obtain a continuous functional equation;

[0186] Discretize the functional equations to obtain a discrete matrix equation system;

[0187] Solve the discrete matrix equations to obtain the coupling relationship solution of the evaluation indicators.

[0188] Then, the maximum regularized cumulative distribution transform is used to classify the results:

[0189] Construct an initial distribution model based on the evaluation indicators;

[0190] Generate a regularized expression based on the initial distribution model;

[0191] Compute the cumulative distribution function using regularized expressions;

[0192] The classification threshold is constructed according to the cumulative distribution function to obtain the grade division of the evaluation results.

[0193] Furthermore, a deep reinforcement learning model is constructed based on the optimization results:

[0194] Build a deep reinforcement learning model based on the optimization results;

[0195] Generate adjustment strategies for recipe parameters using deep reinforcement learning models;

[0196] Storing the adjustment strategy of recipe parameters in the distributed experience replay cache;

[0197] Extract historical optimization data from the distributed experience replay cache and build a time series correlation model.

[0198] Finally, the optimization results are verified based on the timing correlation model:

[0199] Decomposition of optimization parameters using adaptive mesh refinement algorithm;

[0200] Based on the decomposition results, the model parameters are updated using the asynchronous gradient descent algorithm;

[0201] According to the updated model parameters, the parameter adjustment direction is determined by using the Bayesian optimization algorithm;

[0202] Based on the parameter adjustment direction, the key features of the recipe are extracted and the optimization results are updated.

[0203] In one embodiment, S4 specifically includes:

[0204] S4.1: Use the finite element full discrete error analysis method to process the evaluation indicators;

[0205] Firstly, the numerical calculation process of the evaluation indexes by using the finite element full discrete error analysis method is explained in detail. In the specific implementation process, the optimization of the G group formula in potato fertilizer is taken as an example. The basic composition of the formula is 12-0-42+2MgO, and adjuvants such as phytosan, Xuyang rapid expansion, and Jinggen are added.

[0206] Before numerical calculation, the evaluation index space needs to be preprocessed. Specifically, the indicators of different dimensions, such as the growth-promoting effect evaluation index, the formula stability evaluation index, and the water solubility evaluation index, are first mapped to the same calculation space. For example, for the growth-promoting effect evaluation index, it is necessary to consider growth indicators such as plant height, stem thickness, and number of leaves, as well as yield indicators such as the number of tubers and single fruit weight.

[0207] Next, the evaluation index space is divided into a numerical calculation grid of finite dimensions. Taking the G group recipe as an example, a 500×600-dimensional calculation grid is established, in which the row dimension corresponds to different evaluation indicators and the column dimension corresponds to different parameter combinations. This grid division method can effectively control the calculation complexity while ensuring the calculation accuracy.

[0208] After the initial grid structure is established, it is necessary to construct a discrete expression of the evaluation index at each grid node. For example, for the water solubility evaluation index, the dissolution of the formula under 5 times hard water and pure water at 5 minutes and 10 minutes is recorded at each node. These data are converted into standardized numerical expressions for subsequent calculations.

[0209] Then, the error propagation function of each node is calculated based on the discrete expression. In practice, it is found that the error propagation characteristics of different evaluation indicators are significantly different. For example, the error propagation speed of the stability indicator under high temperature and high humidity conditions (40°C, 1% water content) is significantly faster than that under normal temperature conditions, and special attention should be paid to the numerical accuracy control under such conditions.

[0210] In view of the error propagation characteristics, the system will automatically perform local mesh encryption. During the optimization process of the G group formula, the system specially encrypted the mesh area related to the stability index, and refined the original 500×600-dimensional mesh to 2000×2400-dimensional in this area, significantly improving the calculation accuracy. This adaptive mesh encryption strategy not only ensures the calculation accuracy of key areas, but also avoids the waste of computing resources.

[0211] Finally, based on the optimized grid structure, the numerical solution of the evaluation index is calculated. Taking the G group formula as an example, the numerical solution of the complete evaluation index system including growth-promoting effect, stability and water solubility is obtained through finite element full discrete error analysis. Among them, the calculation error of each index is controlled within the order of 10^-6, providing a reliable data basis for subsequent optimization decisions.

[0212] Through this systematic numerical calculation method, the accurate evaluation of the G group formula was successfully achieved. This method not only ensures high calculation accuracy, but also significantly improves the calculation efficiency through adaptive grid optimization. Especially when dealing with stability evaluation under high temperature and high humidity conditions, this method shows obvious advantages and provides reliable technical support for further optimization of the formula.

[0213] S4.2: Establish the coupling relationship of evaluation indicators based on the variational discretization method;

[0214] When establishing the coupling relationship between the evaluation indicators based on the variational separation method, the embodiment of the present application takes the optimization of the G group formula as a detailed example to illustrate the specific implementation process. The formula contains the basic component (12-0-42+2MgO) and a variety of additives (such as phytonadione, Xuyang speed expansion, Jinggen, etc.), involving a complex coupling relationship between multiple evaluation indicators.

[0215] First, the evaluation index space needs to be constructed as a functional space. Specifically, the embodiment of the present application maps the three types of evaluation indicators, namely, growth-promoting effect, stability and water solubility, to different function spaces. Taking the growth-promoting effect indicator as an example, a multidimensional function space including above-ground indicators (plant height, stem thickness, number of leaves) and underground indicators (number of tubers, grade distribution, average weight) is established. Each indicator is represented as a continuously differentiable functional.

[0216] Then, a mapping operator is defined between the evaluation indicators. For example, when analyzing the formula of group G, the embodiment of the present application found that there was a significant correlation between the water solubility index and the stability index. By constructing a mapping operator, the embodiment of the present application can accurately describe how the change in water solubility of the formula affects its storage stability under high temperature conditions of 40°C. This mapping relationship is expressed as a nonlinear operator to characterize the coupling effect between indicators.

[0217] Based on the defined mapping operator, the next step is to construct a variational form. In the analysis of the G group formula, the embodiment of the present application expresses the coupling relationship of the three types of evaluation indicators as a continuous functional equation system. For example, when the dissolution time of the formula in 5 times hard water is extended from 5 minutes to 10 minutes, the embodiment of the present application accurately describes the influence of this time change on the stability of the formula through a variational form.

[0218] Then the functional equation needs to be discretized. In practice, the present embodiment adopts piecewise linear finite element method to convert the continuous functional equation into a discrete matrix equation system. For the optimization problem of the additive ratio in the G group formula, the present embodiment establishes a large-scale matrix equation system containing thousands of degrees of freedom to describe the coupling relationship between the evaluation indicators.

[0219] Finally, the discrete matrix equations are solved to obtain the coupling relationship solutions of the evaluation indicators. In the optimization process of the G group formula, the embodiment of the present application uses the multi-grid method to solve the large-scale matrix equations. Through this method, the embodiment of the present application not only obtains the quantitative relationship between the various evaluation indicators of the formula, but also finds some important coupling characteristics. For example, the embodiment of the present application found that when the adjuvant dosage is diluted 200 times and applied, there is an optimal balance between its effect on promoting the yield and the stability of the formula.

[0220] This method based on variable dispersion has shown significant advantages in practical applications. For example, in the process of optimizing the G group formula, the embodiment of the present application successfully predicted the stability change trend of the formula under high temperature and high humidity conditions (40°C, 1% water content), and optimized the auxiliary agent ratio accordingly, so that the formula not only maintains good water solubility, but also significantly improves storage stability.

[0221] Through this systematic coupling analysis method, the embodiment of the present application can deeply understand the mutual influence between different evaluation indicators and provide a reliable theoretical basis for formula optimization. Especially when dealing with complex multi-index optimization problems, this method shows strong adaptability and effectiveness. With this method, the embodiment of the present application successfully improved the comprehensive performance of the G group formula by more than 25%, while maintaining a good balance of various indicators.

[0222] S4.3: Classify the results using the maximum regularized cumulative distribution transform;

[0223] When using the maximum regularized cumulative distribution transformation to classify the results, the embodiment of the present application still takes the optimization process of the G group formula as a specific example. After the numerical calculation and coupling analysis of the preliminary evaluation indicators, the optimization results need to be further scientifically classified and evaluated.

[0224] First, an initial distribution model is constructed based on the evaluation index. When processing the evaluation data of the G group formula, the embodiment of the present application collected a complete evaluation data set including growth-promoting effect, stability and water solubility. For example, for the growth-promoting effect, the embodiment of the present application recorded the growth data under the standard planting density (2400 trees / mu), including the above-ground indicators (plant height, stem thickness, number of leaves) and underground indicators (number of tubers, grade distribution, single fruit weight, etc.) of the crop. These data preliminarily show the characteristics of an approximate normal distribution.

[0225] Next, a regularized expression is generated based on the initial distribution model. Taking into account the performance differences of the G group formula under different environmental conditions, the embodiment of the present application introduces a regularization term to control the complexity of the model. Specifically, the embodiment of the present application adopts the L2 regularization method, focusing on the performance of the formula under high temperature and high humidity conditions (40°C, 1% water content). Through regularization processing, the volatility of the evaluation results is effectively reduced, making the classification results more stable and reliable.

[0226] Then, the regularized expression is used to calculate the cumulative distribution function. In the evaluation of the G group formula, the embodiment of the present application calculated the cumulative distribution functions of the three dimensions of growth-promoting effect, stability and water solubility. For example, for the evaluation of water solubility, the embodiment of the present application counted the dissolution of the formula under two conditions, 5 times hard water and pure water, at 5 minutes and 10 minutes, and the water solubility distribution characteristics of the formula were clearly shown through the cumulative distribution function.

[0227] Finally, a classification threshold is constructed according to the cumulative distribution function to obtain the grade division of the evaluation results. For the G group formula, the embodiment of the present application sets a scientific classification standard. For example, in the yield evaluation, based on the characteristics of the cumulative distribution function, the tuber yield is divided into five levels: special fruit, first-class fruit, second-class fruit, third-class fruit and fourth-class fruit. Each level has a clear numerical interval, and the division of these intervals fully considers the distribution characteristics of the data.

[0228] In specific applications, this classification method based on maximum regularized cumulative distribution transformation showed significant advantages. Taking the G group formula as an example, when it was applied under 200-fold dilution conditions, this method accurately identified its excellent growth-promoting effect, especially during the tuber expansion period. Statistical results showed that the proportion of special-grade and first-grade fruits in the treated crop group was significantly higher than that in the control group, and the yield-increasing effect was very significant.

[0229] At the same time, this method also effectively identifies the stability characteristics of the formula. By analyzing the performance changes of the G group formula in different storage cycles (1 week, 2 weeks, and 5 weeks), the system rated its stability as the highest level, indicating that the formula has excellent storage performance. In the water solubility evaluation, this method also showed a good classification effect, accurately reflecting the solubility characteristics of the formula under different water quality conditions.

[0230] Through this scientific classification method, the embodiments of the present application not only achieve accurate evaluation of the performance of the formula, but also provide clear guidance for the subsequent optimization direction. For example, based on the classification results, the embodiments of the present application found that the stability of the G group formula under high temperature conditions still has room for improvement, which provides a clear goal for further optimization. The application of this method significantly improves the accuracy of formula optimization, and the reliability of the evaluation results is greatly improved.

[0231] S4.4: Generate recipe parameter adjustment strategy based on deep reinforcement learning model.

[0232] The specific implementation process of building a deep reinforcement learning model based on the optimization results is also explained in detail using the optimization of group G recipe as an example. This step aims to achieve automatic tuning of recipe parameters through deep learning technology to further improve the optimization effect.

[0233] First, a deep reinforcement learning model based on the Actor-Critic architecture is constructed. When processing the G group formula, the Actor network is responsible for generating adjustment strategies for the formula parameters, while the Critic network evaluates the value of the adjustment behavior. For example, when it is detected that the potassium nitrate content in the formula is 889g / kg, the Actor network will generate adjustment suggestions based on historical data, such as suggesting that the content be adjusted to 875g / kg to increase water solubility. At the same time, the Critic network will evaluate the potential benefits of this adjustment and decide whether to adopt this adjustment suggestion by calculating the expected performance improvement value (such as a 5% increase in water solubility).

[0234] Then, the adjustment strategy of the formula parameters is stored in the distributed experience playback cache. In the process of optimizing the G group formula, the system stores the experience data of each parameter adjustment in the distributed cache. These experience data include experimental results under different conditions, such as "under high temperature conditions of 40°C, when the potassium nitrate content is 889g / kg, the 5-minute dissolution rate of the formula is 98%" and other specific data. The system continuously learns these experiences through the asynchronous gradient descent algorithm to optimize the formula parameters.

[0235] Next, historical optimization data is extracted from the distributed experience playback cache to build a time series correlation model. For the G group formula, the embodiment of the present application uses an LSTM recurrent neural network to analyze the performance change trend of the formula under different storage periods (1 week, 2 weeks, and 5 weeks) to predict long-term stability. At the same time, multiple evaluation indicators are processed through a random forest model, such as considering the yield improvement effect of the formula (such as the total fruit weight of potatoes after G group treatment is significantly higher than that of the control group) and water solubility performance (such as no crystallization in 5-minute dissolution).

[0236] In the verification phase, the optimization parameters are decomposed through the adaptive mesh refinement algorithm. Taking the adjustment of the dosage of additives in the G group formula as an example, the system first meshes the basic dosage of 3kg / ton, and then dynamically adjusts the mesh density according to the actual effect, focusing on areas that are sensitive to performance changes. The model parameters are continuously updated through the asynchronous gradient descent algorithm to ensure the convergence of the optimization process.

[0237] Finally, the Bayesian optimization algorithm is used to determine the direction of parameter adjustment. When optimizing the G group formula, the system explores new formula parameter combinations (such as adjusting the amount of adjuvants) through dynamic balance and uses known good formula characteristics to find the optimal parameter adjustment direction. For example, when optimizing the application interval, the system determines through Bayesian optimization that an application interval of 5-7 days can achieve the best effect.

[0238] This optimization method based on deep reinforcement learning has achieved remarkable results in practical applications. Taking the G group formula as an example, after optimization by this method, the comprehensive performance of the formula has been significantly improved: while maintaining the effect of increasing the total output, the water solubility and stability of the formula have been improved by 15%, and the optimization process only takes 1 / 4 of the time of the original experimental method. This result fully demonstrates the advantages of this method in improving the efficiency and accuracy of formula optimization.

[0239] Through this systematic deep learning optimization method, the embodiment of the present application achieves precise control of formulation parameters. This method can not only handle complex multi-parameter optimization problems, but also continuously improve the optimization strategy based on real-time feedback, providing strong technical support for the development of efficient potato fertilizers. In particular, when dealing with multi-objective optimization problems, this method shows excellent adaptability and scalability.

[0240] In summary, in this step S4, the embodiment of the present application deeply uses three core intelligent algorithms to achieve accurate optimization of the formula. First, the finite element fully discrete error analysis method is used to construct an optimized numerical model. Specifically, the embodiment of the present application divides the formula parameter space into a finite number of discrete units, and establishes a local optimization objective function on each unit. This method can accurately capture the impact of parameter changes on formula performance, while effectively controlling the error of numerical calculation. For example, when adjusting the content of potassium nitrate, the embodiment of the present application can accurately calculate the impact of this change on overall performance.

[0241] The embodiment of the present application introduces a variational discretization method to handle complex optimization problems. This method first constructs a variational form of formulation optimization, converting a continuous optimization problem into a discrete numerical solution process. In practical applications, this means that the embodiment of the present application can represent the formulation optimization problem as a series of discrete numerical equations, which accurately describe the relationship between formulation parameters and performance indicators. By solving these equations, the embodiment of the present application can find the optimal parameter combination.

[0242] The embodiment of the present application uses the maximum regularized cumulative distribution transformation method to verify and refine the optimization results. This method is mainly used to deal with the uncertainty in the optimization process. By establishing a probability distribution model for parameter adjustment, the embodiment of the present application can more accurately predict the impact of parameter changes. Specifically, the embodiment of the present application establishes a cumulative distribution function of experimental data, eliminates the impact of random fluctuations through regularization processing, and finally obtains a stable and reliable optimization solution.

[0243] Figure 7This is a schematic diagram of the results of the mid-term (non-final) experiment on potatoes. The results of the potato experiment show that the total fruit weight of the G group after treatment was significantly higher than that of the CK group, and there was no significant difference between the other treatment groups and the CK group. Among them, the percentage of special-grade fruit was higher in the D and G groups; there was no significant difference between the treatment groups in the percentage of first-grade fruit, the percentage of second-grade fruit, and the percentage of fourth-grade fruit; the percentage of third-grade fruit in the D, F, and G groups was significantly lower than that in the A, B, C, and E groups.

[0244] In one embodiment, the method can be performed in a precise optimization system based on deep learning and large-scale distributed computing, as follows:

[0245] In this step, the embodiment of the present application designs an optimization system based on a deep learning framework, which runs on a distributed computing cluster and makes full use of GPU acceleration to handle large-scale recipe optimization calculations. First, the embodiment of the present application uses a finite element fully discrete error analysis method to construct a deep neural network model. The system uses the TensorFlow distributed training framework to map the recipe parameter space to a multidimensional tensor and perform parallel calculations on the CUDA core. This method achieves sub-millisecond optimization parameter calculations through an adaptive grid refinement algorithm, while controlling the numerical calculation error within the order of 10^-6.

[0246] The system introduces a deep reinforcement learning algorithm based on variational discretization. The embodiment of the present application uses PyTorch to build an end-to-end reinforcement learning model, which adopts the Actor-Critic architecture to transform the recipe optimization problem into a Markov decision process. The Actor network is responsible for generating the adjustment strategy of the recipe parameters, while the Critic network evaluates the value of the adjustment behavior. The system stores historical optimization data through the Distributed Experience Replay Buffer and uses the Asynchronous Gradient Descent algorithm to update the model parameters, thereby realizing real-time dynamic adjustment of recipe optimization.

[0247] The embodiment of the present application develops an integrated learning system based on maximum regularized cumulative distribution transformation. The system uses the Apache Spark distributed computing framework to achieve parallel processing of large-scale data. The optimization results are integrated and verified by building a random forest model, and the temporal correlation of the recipe parameters is captured using the LSTM recurrent neural network. The system also integrates an automatic feature engineering module (Automated Feature Engineering), which can automatically discover and extract key features that affect recipe performance. In order to deal with the uncertainty in the optimization process, the embodiment of the present application introduces a Bayesian optimization algorithm (Bayesian Optimization) to dynamically adjust the exploration-exploitation trade-off.

[0248] The following is a specific example of recipe optimization to illustrate this process:

[0249] Taking the optimization of formula G in potato fertilizer as an example, the reinforcement learning system operates as follows:

[0250] First, the specific application process of the Actor-Critic network is as follows: the Actor network receives the initial formula parameters (12-0-42+2MgO basic formula plus additives), and then predicts the direction of parameter adjustment. For example, when it is detected that the potassium nitrate content in the formula is 889g / kg, the Actor network will generate adjustment suggestions based on historical data, such as adjusting the content to 875g / kg to improve water solubility. At the same time, the Critic network will evaluate the potential benefits of this adjustment and decide whether to adopt this adjustment suggestion by calculating the expected performance improvement value (such as a 5% increase in water solubility).

[0251] In the actual optimization process, the distributed experience replay cache stores the data of all past formula experiments. For example, when the system processes a batch of new formula experiment data (such as the Arabidopsis, lettuce and potato experiment results in the mid-term report), these data will be stored in the distributed cache. To give a specific example, the cache contains experience data such as "under high temperature conditions of 40°C, when the potassium nitrate content is 889g / kg, the 5-minute dissolution rate of the formula is 98%." The system will continuously learn these experiences through the asynchronous gradient descent algorithm to optimize the formula parameters.

[0252] In the integrated learning system, the embodiment of the present application uses Apache Spark to parallel process experimental data under different conditions. The random forest model processes multiple evaluation indicators, such as considering the yield improvement effect of the formula (such as the total fruit weight of potatoes after G group treatment is significantly higher than that of the control group) and water solubility (such as no crystallization after 5 minutes of dissolution). The LSTM network specifically analyzes the performance change trend of the formula under different storage cycles (1 week, 2 weeks, 5 weeks) and predicts long-term stability.

[0253] The automatic feature engineering module found through analysis that, for example, the dosage of adjuvants (3kg / ton) and the application interval (once every 5-7 days) are key factors affecting the effect of the formula. Based on these known features, the Bayesian optimization algorithm determines the next optimization direction by balancing and exploring new formula parameter combinations (such as adjusting the dosage of adjuvants) and using known good formulas (such as Group G formula).

[0254] Through such a system, the embodiments of the present application achieve accurate optimization of the formula. For example, while maintaining the effect of increasing the total output, the water solubility and stability of the formula are increased by 15%, and the optimization process only takes 1 / 4 of the time of the original experimental method. This greatly improves the efficiency and accuracy of formula development.

[0255] This example shows how to apply advanced computer technology to practical agricultural problems and achieve precise formulation optimization through data-driven methods. The system can not only handle complex multi-parameter optimization problems, but also continuously improve the optimization strategy based on real-time feedback, providing strong technical support for the development of efficient potato fertilizers.

[0256] Through this implementation scheme based on modern computer technology, the embodiment of the present application successfully transforms the theoretical model into an efficient and practical engineering system, providing strong technical support for formula optimization. The system not only improves the optimization efficiency, but also enhances the reliability of the optimization results, providing a solid technical guarantee for the industrial optimization of potato-specific fertilizer formulas.

[0257] The following is a set of comparisons:

[0258] Table 2 Performance comparison before and after optimization of formula

[0259]

[0260]

[0261] By comparing the data, we can see that the optimized formula has significantly improved various performance indicators. Among them:

[0262] In terms of growth indicators, plant height, stem diameter and number of functional leaves increased by 20.1%, 19.4% and 23.0% respectively, indicating that the optimized formula significantly promoted plant growth and development;

[0263] In terms of yield indicators, the number of tubers per plant increased by 28.0%, the rate of premium fruit increased by 38.1%, and the average single fruit weight increased by 15.9%, which comprehensively reflects the significant improvement effect of the optimized formula on yield and quality.

[0264] In terms of stability and water solubility, high-temperature anti-caking and storage stability increased by 15.3% and 14.7% respectively, and the 5-minute dissolution rate increased to 98.4%, indicating that the optimized formula has better application performance;

[0265] In terms of system performance, the optimization calculation time was reduced from 48 hours to 12 hours, and the formula evaluation accuracy was improved by 23.3%, reflecting the efficiency and accuracy of the optimization method.

[0266] The specific implementation of the verification experiment is as follows:

[0267] Conduct multi-level verification experiments through the automated experiment platform:

[0268] First, conduct the Arabidopsis experiment:

[0269] Deployment of automated training systems;

[0270] The effect of the formulation on aboveground biomass was assessed using a leaf area meter;

[0271] The growth status changes were recorded by an image analysis system.

[0272] Secondly, lettuce experiments were conducted using the intelligent greenhouse control system:

[0273] Set standard planting density (2,200 trees / acre);

[0274] The effect of the formula on the fresh weight of the stem was verified by a high-precision electronic balance;

[0275] Record growth process data in real time.

[0276] Finally, deploy the field trial monitoring system for potato experiments:

[0277] Set standard planting density (2,400 trees / acre);

[0278] Collect potato growth data in real time through IoT devices;

[0279] Use edge computing devices to perform tuber yield grading and statistics;

[0280] Analyze the effects of different formulations on yield and quality.

[0281] Based on the above implementation, the present application may have other variations and improvements. For example, the configuration and sampling frequency of the sensor may be adjusted according to the specific application scenario, or the evaluation index system may be optimized for different crop types. These variations and improvements should fall within the scope of protection of the present application.

[0282] The embodiment of the present application also provides a potato-specific fertilizer formula optimization system, including: an infrastructure layer, including a computing cluster, a GPU accelerated computing node, a distributed storage system, and a network interconnection system; a middleware layer, including a data integration service module and a model service module; an application layer, including a formula optimization management system, an experimental data analysis platform, and a decision support system; a field data acquisition system for collecting multi-source experimental data, including: a temperature sensor, a humidity sensor, a light sensor, a soil moisture sensor, a soil EC sensor, and a pH sensor, wherein the multi-source experimental data includes soil detection data, meteorological environment data, growth index data, and adjuvant group data. The distributed storage system is used to construct a fertilizer formula database to store component proportions, physical and chemical properties and application effects of different formulas, wherein the component proportions include the content and ratio of different raw materials, the physical and chemical properties include pH value, conductivity and water solubility, and the application effects include yield and quality data; the model service module is used to construct an evaluation model based on the multi-source experimental data, and use the evaluation model to score the formulas in the formula database; the GPU accelerated computing node is used to optimize the formula using an intelligent optimization algorithm to obtain an optimized formula, and store the optimized formula in the formula database.

[0283] The system of the embodiments of the present disclosure can execute the method provided by the embodiments of the present disclosure, and the implementation principles are similar. The actions performed by each module in the system of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed functional description of each module of the system, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0284] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

Claims

1. A method for optimizing a potato-specific fertilizer formula, characterized in that: The method is applied to a cloud service system, and the cloud service system comprises: The infrastructure layer includes computing clusters, GPU-accelerated computing nodes, distributed storage systems, and network interconnection systems; Middleware layer, including data integration service module and model service module; Application layer, including recipe optimization management system, experimental data analysis platform and decision support system; The method comprises: Multi-source experimental data are collected through a field data collection system, the field data collection system includes a temperature sensor, a humidity sensor, a light sensor, a soil moisture sensor, a soil EC sensor and a pH sensor, and the multi-source experimental data include soil detection data, meteorological environment data, growth index data and adjuvant combination data; wherein the soil detection data include soil pH value, electrical conductivity and effective nutrient content; the meteorological environment data include temperature, humidity and light intensity; the growth index data include plant height, stem thickness, number of leaves and tuber data; the adjuvant combination data include the type, dosage and application method of the adjuvant; A fertilizer formula database is constructed in the distributed storage system to store component ratios, physical and chemical properties, and application effects of different formulas, wherein the component ratios include the content and ratio of different raw materials, the physical and chemical properties include pH value, conductivity, and water solubility, and the application effects include yield and quality data; Building an evaluation model based on the multi-source experimental data through the model service module, and scoring the recipes in the recipe database using the evaluation model; The GPU is used to accelerate the computing node and utilize an intelligent optimization algorithm to optimize the recipe to obtain an optimized recipe, and the optimized recipe is stored in the recipe database.

2. The method according to claim 1, characterized in that The optimization of the recipe by using the intelligent optimization algorithm through the GPU accelerated computing node includes: The finite element full discrete error analysis method is used to process the numerical calculation of the evaluation index, wherein the evaluation index includes a growth promotion effect evaluation index, a formula stability evaluation index and a water solubility evaluation index; Establishing the coupling relationship between the evaluation indicators based on the variational discretization method; The results were classified using the maximum regularized cumulative distribution transform; The numerical calculation results, coupling relationships and result classification are integrated to determine the final optimization formula.

3. The method according to claim 2, characterized in that The finite element fully discrete error analysis method is used to process the numerical calculation of the evaluation index, including: Divide the evaluation index space into a numerical calculation grid of finite dimensions to obtain an initial grid structure; Based on the initial grid structure, a discrete expression of the evaluation index is established on each grid node; Calculating the error propagation function of each node according to the discrete expression; Locally encrypting the grid based on the error propagation function to generate an optimized grid structure; Based on the optimized grid structure, a numerical solution of the evaluation index is calculated.

4. The method according to claim 2, characterized in that: The coupling relationship between the evaluation indicators is established based on the variational discretization method, including: Constructing the evaluation index space as a functional space and defining a mapping operator between the evaluation indexes; constructing a variational form based on the mapping operator to obtain a continuous functional equation; Discretizing the functional equation to obtain a discrete matrix equation group; Solve the discrete matrix equations to obtain a coupling relationship solution of the evaluation index.

5. The method according to claim 2, characterized in that: Using the maximum regularized cumulative distribution transformation to classify the results includes: Construct an initial distribution model based on the evaluation indicators; generating a regularized expression based on the initial distribution model; Calculating a cumulative distribution function using the regularized expression; A classification threshold is constructed according to the cumulative distribution function to obtain a grade division of the evaluation results.

6. The method according to claim 2, characterized in that After optimizing the recipe by using the intelligent optimization algorithm through the GPU accelerated computing node, it also includes: Based on the optimization results, a deep reinforcement learning model is constructed; Generating an adjustment strategy for recipe parameters using the deep reinforcement learning model; Storing the adjustment strategy of the recipe parameters in a distributed experience replay cache; Extracting historical optimization data from the distributed experience replay cache and building a time series correlation model; The optimization results are verified based on the timing correlation model to obtain the final recipe parameters.

7. The method according to claim 6, characterized in that The verification of the optimization result based on the timing correlation model includes: Decomposition of optimization parameters using adaptive mesh refinement algorithm; Based on the decomposition result, an asynchronous gradient descent algorithm is used to update the model parameters; According to the updated model parameters, determining the parameter adjustment direction through a Bayesian optimization algorithm; Based on the parameters, the direction is adjusted, the key features of the recipe are extracted and the optimization results are updated.

8. The method according to claim 1, characterized in that: The method further comprises: The Arabidopsis experiment was conducted using an automated experimental platform, and the effect of the formula on aboveground biomass was evaluated using a leaf area meter; The lettuce experiment was conducted using an intelligent greenhouse control system, and the effect of the formula on the fresh weight of the stem was verified using a high-precision electronic balance; Deploy a field trial monitoring system, collect potato growth data in real time through IoT devices, and use edge computing devices to perform tuber yield grading and statistics.

9. The method according to claim 1, characterized in that: In the step of collecting multi-source experimental data through the field data collection system, collecting the soil detection data includes: The pH sensor is used to collect the soil pH value with an accuracy of ±0.1 pH, and the soil acidity and alkalinity are measured to evaluate whether the soil environment is suitable for potato growth; The soil EC sensor is used to collect soil conductivity data with a measurement range of 0-20ms / cm to evaluate soil salinity; The effective nutrient content data, including available nitrogen, available phosphorus and available potassium, are collected through high-precision element analyzers and ion chromatographs to assess the soil's nutrient supply capacity; Among them, the soil detection data is collected by an industrial-grade data collector with a sampling frequency of 15 minutes per time, and the collected data is transmitted to the cloud service system via the 4G / 5G network.

10. A potato-specific fertilizer formula optimization system, characterized in that: include: The infrastructure layer includes computing clusters, GPU-accelerated computing nodes, distributed storage systems, and network interconnection systems; Middleware layer, including data integration service module and model service module; Application layer, including recipe optimization management system, experimental data analysis platform and decision support system; A field data collection system for collecting multi-source experimental data, including: a temperature sensor, a humidity sensor, a light sensor, a soil moisture sensor, a soil EC sensor, and a pH sensor, wherein the multi-source experimental data includes soil detection data, meteorological environment data, growth index data, and adjuvant combination data; The distributed storage system is used to construct a fertilizer formula database to store component ratios, physical and chemical properties, and application effects of different formulas, wherein the component ratios include the content and ratio of different raw materials, the physical and chemical properties include pH value, conductivity, and water solubility, and the application effects include yield and quality data; The model service module is used to construct an evaluation model based on the multi-source experimental data, and use the evaluation model to score the recipes in the recipe database; The GPU accelerated computing node is used to optimize the recipe using an intelligent optimization algorithm to obtain an optimized recipe, and store the optimized recipe in the recipe database.

Images