Jinxiu sweet tea whole industry chain optimization method based on artificial intelligence

Abstract

The invention relates to a Jinxiu sweet tea whole industry chain optimization method based on artificial intelligence, and belongs to the crossing field of agricultural biotechnology and artificial intelligence, the method is realized in sequence by constructing five AI subsystems and a distributed data center: AI directional breeding, gene and character association based on an improved random forest algorithm, and optimization of the whole industry chain of Jinxiu sweet tea. The breeding period is shortened to be within 18 months by combining reinforcement learning; aI group embryo seedling culture: improving the survival rate to 90% or above through the Internet of Things and gradient improvement regression optimization environment; aI planting management and protection are carried out, AI-GIS site selection, LSTM water and fertilizer decision, YOLOv8 pest and disease damage identification and near infrared spectrum harvesting decision are fused, and precise planting is achieved; aI deep processing is carried out, an extraction process is optimized by utilizing reinforcement learning, and product research and development are accelerated based on online spectrum and GAN; and AI supply chain collaboration, and precise supply and demand matching is realized through LSTM demand prediction, intelligent recommendation and block chain traceability. According to the method, collaborative optimization of the whole industrial chain and comprehensive improvement of efficiency are realized.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of agricultural biotechnology and artificial intelligence, specifically to an artificial intelligence-based method for optimizing the entire industrial chain of Jinxiu sweet tea. Background Technology

[0002] Jinxiu sweet tea, belonging to the genus Rubus of the Rosaceae family, is a unique medicinal and edible plant native to Jinxiu Yao Autonomous County, Guangxi. Its leaves are rich in glycosides, which are characterized by high sweetness and low calories, making it a highly promising natural non-sugar sweetener that can be used as a new food ingredient in sugar-free beverages and health foods. However, the current Jinxiu sweet tea industry faces several technological bottlenecks across the entire chain, from variety selection to market supply, hindering its standardization, large-scale production, and industrialization.

[0003] First, targeted breeding of varieties is inefficient. Traditional breeding methods heavily rely on phenotypic selection and artificial hybridization, with breeding cycles lasting 3 to 5 years. Furthermore, they struggle to accurately aggregate complex traits such as high sucralose content (target ≥98%) and high stress resistance (e.g., resistance to felt disease), failing to meet the downstream food industry's rigid demands for uniform raw material composition and stable quality. Second, the level of intelligence in production and processing is insufficient. The survival rate of embryo culture seedlings is generally only 60%-70%, and water and fertilizer management in field planting relies on experience, resulting in resource waste exceeding 30%. Optimization of sucralose extraction processes relies on trial and error, with extraction rates below 80%, large fluctuations in product purity, and a defect rate often exceeding 5%. Finally, data is fragmented across the industry chain, leading to poor coordination. There is a lack of effective data linkage mechanisms between breeding, planting, processing, and end-market demand. Production parameters cannot be dynamically adjusted based on the specific formulation needs of downstream enterprises (e.g., specific taste compatibility, purity standards), resulting in frequent problems such as slow supply chain response, inventory backlogs, or raw material shortages.

[0004] In existing technologies, the application of artificial intelligence in agriculture is mostly focused on localized optimization of single stages, such as image recognition-based pest and disease diagnosis and simulation-based assisted breeding. While these solutions can improve efficiency in specific stages, they lack a systematic design for the entire Jinxiu sweet tea industry chain. In particular, they fail to construct dedicated AI models that take into account its unique genetic traits, tissue culture physiological requirements, and the physicochemical properties of sweet tea glycosides, thus failing to systematically address the aforementioned industry pain points. In conclusion, there is an urgent need for an integrated artificial intelligence optimization method that can span the entire process of "biological breeding—intelligent seedling raising—precision planting—efficient extraction—supply chain collaboration." Summary of the Invention

[0005] In order to solve the problems of the prior art, this invention provides an artificial intelligence-based optimization method for the entire industrial chain of Jinxiu sweet tea.

[0006] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: Firstly, an artificial intelligence-based method for optimizing the entire industrial chain of Jinxiu sweet tea, comprising the following steps: S1. Construct an AI-driven seed breeding system: Collect gene sequences and trait data of Jinxiu sweet tea and construct a gene-trait database; use an improved random forest algorithm to train a gene-trait association model and locate high-value gene fragments; predict the traits of hybrid offspring through an AI breeding simulation system and dynamically adjust breeding parameters by combining reinforcement learning to cultivate specialized varieties with a sweet tea glycoside purity ≥98% and a resistance to felt disease level ≤2. S2. Construct an AI-powered intelligent control subsystem for tissue culture seedling cultivation: Deploy IoT sensors in the tissue culture room to collect environmental parameters in real time; use a gradient boosting regression model to predict the survival rate of test seedlings and optimize the combination of environmental parameters through grid search; combine machine vision and near-infrared spectroscopy to automatically screen robust seedlings. S3. Construct a standardized AI planting and management subsystem: Integrate environmental data of Jinxiu area, and use AIGIS spatial analysis model to screen planting plots with a suitability score of ≥80; deploy drones and soil sensors, and generate precise water and fertilizer solutions based on LSTM time series prediction model; use improved YOLOv8 algorithm combined with meteorological data to construct a pest and disease control system; and use near-infrared spectroscopy to detect the content of sweet tea glycosides in leaves to trigger intelligent harvesting instructions. S4. Construct an AI-based subsystem for optimizing the deep processing of sucrose glycosides: Build a reinforcement learning environment based on the physicochemical properties of sucrose glycosides, train a process optimization model using the PPO algorithm; deploy online spectral detection equipment in the extraction and purification stages, and use AI to adjust process parameters in real time; construct a database of sucrose glycoside applications, and use generative adversarial networks to simulate product taste; AI automatically integrates data to generate new food ingredient application documents. S5. Construct an AI supply chain collaboration subsystem: Analyze downstream enterprise order data using an LSTM time series model to generate raw material procurement and inventory plans; based on enterprise product types and target audiences, push customized solutions through AI recommendation algorithms; use AI and blockchain technologies to record data throughout the entire process and generate a unique traceability QR code. S6. Build a distributed data hub, using a MySQL+Redis architecture to achieve real-time synchronization of data from the five subsystems, and establish data cleaning rules and permission management modules to ensure data quality and security.

[0007] In one specific implementation of the first aspect, in step S1, the improved random forest algorithm uses feature out-of-bag error to screen gene fragment features, and adjusts the number and maximum depth of decision trees through grid search, so that model R... 2 Coefficient ≥ 0.92, confidence level ≥ 90%.

[0008] In one specific implementation of the first aspect, the input features of the LSTM time-series prediction model in step S3 include plant NDVI value, soil moisture content, nitrogen, phosphorus and potassium content, average temperature and precipitation over the past 7 days, and the output is the watering time and fertilization plan for the next 7 days.

[0009] In one specific implementation of the first aspect, the reward function calculation formula for the reinforcement learning environment in step S4 is as follows: Reward = 0.4 × Extraction rate + 0.5 × Purity − 0.1 × (Energy consumption / 2) Extraction rate and purity are expressed as percentages, and energy consumption is expressed in kW.

[0010] In one specific implementation of the first aspect, the blockchain technology in step S5 adopts a consortium blockchain architecture, with nodes including planting bases, deep processing workshops, downstream enterprises and regulatory agencies, and data modification requires consensus verification by at least 3 nodes.

[0011] In one specific implementation of the first aspect, the data cleaning rules in step S6 include the use of the 3σ principle for sensor outlier removal, the standardization of gene sequence format to FASTQ format, and the use of the K-nearest neighbor algorithm for filling missing trait data values.

[0012] Secondly, a system for optimizing the entire industrial chain of Jinxiu sweet tea based on artificial intelligence includes: Gene sequencing equipment used to collect the sweet tea gene sequence; Internet of Things (IoT) sensors are used to detect light intensity, temperature, and soil parameters. Multispectral drones are used for monitoring plant growth. Online spectrometer for detecting the purity of catechins; The server is used to deploy the five major AI subsystems and the distributed data hub. The client is used by users to view data, receive alerts, and issue commands.

[0013] In one specific implementation of the second aspect, the server hardware configuration includes: CPU ≥ Intel Xeon Gold 6330, memory ≥ 64GB, hard disk ≥ 2TB SSD, GPU ≥ NVIDIA A100; the client supports Windows 10 / 11 and macOS 12+ operating systems, and the browser supports Chrome 90+ and Edge 90+.

[0014] Thirdly, a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an artificial intelligence-based optimization method for the entire industrial chain of Jinxiu sweet tea.

[0015] 10. The computer-readable storage medium according to claim 9, wherein the storage medium includes at least one of a USB flash drive, a portable hard drive, ROM, RAM, and SSD, and the computer program includes an AI model training module, a data processing module, and a user interaction module.

[0016] The beneficial effects of this invention are as follows: 1. This invention, by constructing a full-chain artificial intelligence optimization system covering breeding, seedling cultivation, planting, processing, and the supply chain, has produced significant and synergistic beneficial technical effects. In the core production stage, based on a gene-trait association model using an improved random forest algorithm and dynamic regulation using reinforcement learning, the precise location and expression optimization of key genes for stevia synthesis have been achieved, shortening the targeted breeding cycle from the traditional 3-5 years to less than 14 months, and stably cultivating specialized varieties with stevia purity ≥98% and disease resistance level ≤2. Through the integrated application of models such as IoT sensing and gradient boosting regression, LSTM time series prediction, and YOLOv8 visual recognition, the survival rate of embryo-based seedling cultivation has increased from 60%-70% to over 92%, water and fertilizer waste during the planting process has been reduced by 32%, the accuracy rate of pest and disease identification is ≥98% with early warning and control, and the use of chemical pesticides has been reduced by over 60%. At the same time, relying on near-infrared spectroscopy detection, the quality fluctuation of raw material harvesting is controlled to ≤1%.

[0017] 2. At the level of industrial collaboration and standardization, relying on reinforcement learning (PPO algorithm) process optimization models and online real-time spectral quality control, the extraction rate of sucralose has increased from less than 80% to 95.5%, product purity has stabilized at over 98%, and the defect rate has decreased to below 1%. Furthermore, generative adversarial networks (GANs) have shortened the downstream application formula development cycle by 80%. Through the system integration of LSTM demand forecasting models, AI recommendation algorithms, and consortium blockchain traceability technology, efficient matching of downstream order demand and upstream production plans has been achieved, reducing supply chain response time from 7 days to 3 days, increasing inventory turnover by 40%, and constructing a fully tamper-proof, credible data storage and transparent traceability system. In summary, the various subsystems achieve deep collaboration through a distributed data hub, significantly improving efficiency and quality at each stage while driving the Jinxiu sweet tea industry towards standardization, digitalization, and sustainability. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the Jinxiu Sweet Tea whole industry chain optimization system architecture based on artificial intelligence of the present invention.

[0019] Figure 2 This is a schematic diagram of the training process of the AI-directed breeding gene-trait association model of the present invention.

[0020] Figure 3 This is a schematic diagram of the reinforcement learning optimization curve of the AI-based sweet tea glycoside extraction process of the present invention.

[0021] Figure 4 This is a time-series diagram of data collaboration across the entire industry chain according to the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figures 1 to 4 This paper presents an AI-based optimization method for the entire industrial chain of Jinxiu sweet tea.

[0024] 1. Construction and Implementation of AI-Driven Targeted Seed Breeding System This subsystem aims to address the core pain points of traditional breeding, such as long cycles and poor targeting, by establishing an intelligent association between genotype and phenotype to achieve precise and rapid breeding of target traits.

[0025] Step S1.1: Multidimensional data collection and gene-trait database construction Sample Collection: Germplasm resources of *Rubus jinxiuensis* (family Rosaceae) were systematically collected within Jinxiu County to ensure genetic diversity. Specifically, this included: no fewer than 100 wild germplasm samples (covering an altitude gradient of 500-1500 meters) and no fewer than 50 existing cultivated varieties.

[0026] Genotype data acquisition: Whole-genome resequencing was performed on the above samples using a high-throughput sequencing platform (such as Illumina NovaSeq 6000), with an average sequencing depth of no less than 30×. The bioinformatics analysis workflow focuses on: Key functional genes: Completely obtain sequence variation information of the UGT72E gene family (such as UGT72E2, UGT72E3) in the biosynthetic pathway of stevia.

[0027] Adaptation-related genes: Analyzing the genetic polymorphism of photoperiod-responsive genes (such as PHYA) and disease resistance-related gene clusters.

[0028] Phenotypic data determination: Core quality traits: During the full leaf stage, the content (%) of stevia in the leaves was accurately determined by high performance liquid chromatography (HPLC).

[0029] Resistance traits: Resistance was identified by artificial inoculation with the pathogen of felt disease, and was classified into 1-5 levels (level 1 being highly resistant) based on the proportion of lesion area.

[0030] Agronomic traits: Standardized measurement of leaf length, leaf width, plant height, number of branches, etc.

[0031] Database integration: Gene sequences are stored in FASTQ format, phenotypic data are stored in structured CSV format, and association mapping is performed through unique sample IDs to construct a gene-trait association database for Jinxiu sweet tea with an initial capacity of no less than 100,000 records.

[0032] Step S1.2: Gene-trait association modeling based on improved random forest algorithm Feature engineering: Extract high-quality SNPs (single nucleotide polymorphisms) and InDel (insertion and deletion) markers from the resequencing data as the initial feature set, with the dimension controlled between 500-1000.

[0033] Model Training and Optimization: An improved Random Forest algorithm is used for modeling. The improvements are as follows: Out-of-Bag Error (OOB) estimation is used to assess feature importance and perform initial screening, effectively eliminating irrelevant or noisy labels.

[0034] The model hyperparameters were optimized using a grid search method, with the key parameters being: the number of decision trees (100-200) and the maximum depth of the trees (8-12 layers).

[0035] Model Validation and Application: Train the model using SNP / InDel markers as input and the target phenotype (e.g., glycoside content) as output. Model performance must meet the following requirement: coefficient of determination R0 2 ≥ 0.92, with a confidence level of ≥ 90% through 10-fold cross-validation. The model can output the top 10 gene markers by feature importance, clearly pointing to genomic regions significantly associated with "high saccharin content (≥12%)" and "high resistance to felt disease (grade ≤2)".

[0036] Step S1.3: Intelligent breeding simulation and dynamic control of the cultivation environment Virtual breeding design: Based on a trained association model, an AI breeding simulation module is developed. Users input the genotype information of preset hybrid parents, and the system can simulate and calculate tens of thousands of virtual offspring within 10 minutes, and output the predicted values ​​of their phenotypic distribution, including: stevia content (prediction error ±0.5%), disease resistance level, and optimal breeding cycle (12-18 months).

[0037] Adaptive optimization of the breeding process: During the tissue culture or greenhouse breeding stage, the physiological state of the plants (such as chlorophyll fluorescence parameters Fv / Fm and endogenous hormone levels) is monitored in real time using environmental sensors. Reinforcement learning algorithms (such as DeepQ-Network, DQN) are employed, using daily light duration (10-12 hours) and cytokinin concentration (such as 6-BA) in the culture medium (1.5-3.5 mg / L) as adjustable "actions." The algorithm uses plant biomass growth rate and the accumulation of stevia precursors as "rewards" to dynamically explore and lock in the optimal combination of environmental parameters, ensuring stable expression of the target traits. Ultimately, this led to the successful breeding of specialized new varieties such as "Jintian AI-1" (stevia purity ≥98%, resistance to felt disease ≤2).

[0038] 2. Construction and Implementation of AI-Powered Intelligent Regulation Subsystem for Embryo Tissue Culture and Seedling Raising This subsystem is dedicated to achieving intelligent and standardized production of tissue culture seedlings, significantly improving seedling efficiency and quality.

[0039] Step S2.1: IoT Sensing of All Elements of Tissue Culture Environment Deploying a high-precision, high-density sensor network within the tissue culture workshop: Light intensity sensor: range 0-10000 lux, accuracy ±5%.

[0040] Integrated temperature and humidity sensor: Temperature range 15-30℃, accuracy ±0.5℃; Humidity range 0-100%RH.

[0041] pH sensor: Implanted in the culture medium to monitor pH value in real time (range 5.5-6.5).

[0042] All sensors are networked via a wireless network (such as LoRaWAN) and upload data to the cloud data hub every 5 minutes.

[0043] Step S2.2: Survival rate prediction and parameter optimization based on gradient boosting regression Predictive Model Construction: Historical batch environmental parameter data and corresponding final seedling survival rate data were collected to train a Gradient Boosting Regression Tree (GBRT) model. This model, using real-time collected light, temperature, humidity, and pH values ​​as input, can accurately predict the survival rate of seedlings, with a required mean absolute error (MAE) of ≤3%.

[0044] Global optimization of environmental parameters: Using a grid search algorithm, an exhaustive search is performed within a preset feasible region (light: 3000-7000 lux, temperature: 20-26℃, pH: 5.8-6.2, hormone concentration: 2.0-3.0 mg / L) to find the combination of environmental parameters that makes the survival rate predicted by the GBRT model the highest. This combination serves as the gold standard parameter set for the tissue culture production of this variety, which can increase the average survival rate to over 90%.

[0045] Step S2.3: Intelligent screening of seedling conditions by integrating machine vision and spectral technology Dual-modal data acquisition: Conducted simultaneously before seedling hardening and transplanting. Machine vision imaging: Industrial cameras (1280×720 resolution) are used to capture images of the canopy of each seedling.

[0046] Near-infrared spectral scanning: Leaf spectral information was acquired non-destructively using a portable near-infrared spectrometer (wavelength range 700-1100 nm).

[0047] Intelligent analysis and decision-making: Morphological screening: YOLOv5 target detection algorithm was used to identify and measure the leaves. The qualified criteria were set as follows: leaf length ≥ 3cm and leaf width ≥ 1.5cm.

[0048] Physiological screening: Based on the established spectrum-chlorophyll content PLS regression model, chlorophyll content is quickly predicted, and the qualified threshold is ≥2.5 mg / g.

[0049] Combining the two results, substandard individuals (weak seedlings, diseased seedlings) were marked and removed by an automated production line, with an overall screening accuracy of ≥95%.

[0050] 3. Construction and Implementation of the AI-Based Standardized Planting and Management Subsystem This subsystem enables precise and standardized management of the entire process of field planting, from site selection to harvesting.

[0051] Step S3.1: Intelligent selection of planting plots based on AI-GIS model Multi-source data fusion: Integrating digital elevation model (DEM), soil type map, and multi-year meteorological raster data (annual average temperature and precipitation) of Jinxiu area.

[0052] Spatial Adaptability Modeling: An AI-enhanced Geographic Information System (AI-GIS) model was constructed to quantify the optimal ecological factors for Jinxiu sweet tea (altitude 500-1000 meters, yellow / brown soil, average annual temperature 18-22℃, annual precipitation 1500-2000 mm) into spatial evaluation indicators. Raster calculations and scoring (0-100 points) were performed on the entire region. The final output is a map of preferred planting areas with an adaptability score ≥80, achieving a site selection accuracy ≥90%.

[0053] Step S3.2: Integrated Sky-Ground Growth Monitoring and Intelligent Water and Fertilizer Decision Making Three-dimensional data acquisition: Aerial monitoring: Drone platforms such as DJI Matrice 300 RTK, equipped with multispectral cameras, fly once a week along preset routes to obtain the Normalized Difference Vegetation Index (NDVI) at the plot scale.

[0054] Ground-based sensor network: Deploy a wireless soil sensor network to monitor soil volumetric water content (target: 18%-22%), electrical conductivity, and available nitrogen, phosphorus, and potassium nutrient content in real time.

[0055] Intelligent irrigation and fertilization decision-making: A water and fertilizer requirement prediction model is constructed using a Long Short-Term Memory (LSTM) network. Input time-series data includes: historical and real-time NDVI, soil moisture and fertility, and a refined 7-day weather forecast. The model outputs the daily irrigation time points (accurate to the hour) for the next week and the recommended application rates of each nutrient element (e.g., 20-30 kg / mu of pure nitrogen), achieving precise integrated water and fertilizer management and improving resource utilization by more than 30%.

[0056] Step S3.3: Intelligent and Green Prevention and Control of Pests and Diseases Based on Computer Vision and Occurrence Models Intelligent pest and disease identification: High-definition visible light images collected by drone patrols are analyzed in real time using the YOLOv8 target detection algorithm optimized for agricultural scenarios. The model's accuracy requirements for identifying major pests and diseases of Jinxiu sweet tea are: felt disease lesions ≥98%, and beet armyworm body count ≥96%.

[0057] Occurrence Probability Prediction and Early Warning: A logistic regression model for the occurrence probability of pests and diseases is constructed by combining real-time weather station data (temperature, relative humidity). When the system identifies pests and diseases, and environmental conditions (e.g., temperature ≥25℃, humidity ≥80%) meet the outbreak threshold, the model predicts the probability of occurrence over the next 3-5 days. If the probability of occurrence is ≥70%, the system immediately issues an early warning to the administrator through multiple channels (APP, SMS).

[0058] Automatic matching of green prevention and control solutions: The system has a built-in plant protection knowledge graph that automatically recommends and pushes the optimal green prevention and control solutions based on the identified pest and disease types. For example, for aphids, it is recommended to "release 5,000 aphid wasps per acre within 3 days"; for felt disease, it is recommended to "spray 1% matrine soluble solution, diluted 800 times". This system can reduce the use of chemical pesticides by more than 60%.

[0059] Step S3.4: Quality-oriented intelligent harvesting based on near-infrared spectroscopy Rapid field quality testing: During the harvest season, a handheld high-resolution near-infrared spectrometer (1 nm resolution) is used to conduct sampling tests in the field to quickly and non-destructively determine the content of glycosides in the leaves.

[0060] Automatic harvesting command triggering: When the system detects that the average content of sucralose in a certain plot or batch is consistently stable at ≥12%, the AI ​​maturity judgment model automatically determines that the optimal process maturity period has been reached, and then issues a harvesting command containing plot number, area and expected yield to the corresponding harvesting team's smart terminal (such as PDA, tablet).

[0061] This instruction can guide manual or intelligent harvesting equipment to operate, ensuring the uniformity of quality within batches of raw materials and controlling the fluctuation of glycoside content to ≤1%.

[0062] 4. Construction and Implementation of AI-Based Optimization Subsystem for Deep Processing of Stevia Glutamate This subsystem optimizes the entire chain from extraction and purification to product application development, improving efficiency and quality.

[0063] Step S4.1: Autonomous optimization of extraction process based on reinforcement learning Markov decision process modeling: Modeling the water extraction process of betaine as a reinforcement learning environment.

[0064] Status (S): Current process parameters (temperature, liquid-to-material ratio, time) and spectral characteristics of raw material batches.

[0065] Action (A): Adjust the three core parameters: extraction temperature (50-90℃), solvent (water) to raw material ratio (1:5-1:15), and extraction time (30-120 minutes).

[0066] Reward (R): A comprehensive reward function is used: R = 0.4 × Extraction Rate (%) + 0.5 × Purity (%) - 0.1 × (Energy Consumption (kW) / 2). This function balances yield, quality, and cost.

[0067] Strategy Optimization and Output: The Proximal Policy Optimization (PPO) algorithm was used to train the system in this environment. After approximately 100 training cycles (simulating production batches), the strategy network converged, outputting the optimal process parameter range: temperature 75-80℃, liquid-to-material ratio 1:10-1:12, and time 60-80 minutes. Under these parameters, the actual production extraction rate can reach over 95%, and the unit energy consumption is reduced by 20%.

[0068] Step S4.2: Online real-time spectral quality control and feedback adjustment Application of process analysis technology: An online ultraviolet-visible spectrometer (wavelength range 200-400 nm) is installed in the outlet flow path of the purification chromatography column to continuously monitor the effluent.

[0069] Real-time purity calculation and control: The AI ​​algorithm analyzes the spectrum in real time and calculates the real-time purity of stevia in the current fraction by comparing it with the standard spectral library (calculation error ±0.2%). If the purity is detected to be lower than the set target (e.g., 98%), the system automatically performs feedback adjustment, such as reducing the column flow rate (from 1.5 BV / h to 1.2 BV / h) or fine-tuning the eluent ethanol concentration (e.g., from 75% to 78%), to ensure that the purity of the final product is consistently up to standard and to control the defect rate to below 1%.

[0070] Step S4.3: Digital taste simulation and rapid recipe development for downstream applications Application Database Construction: Establish a database of interactions between "sweet tea glycosides and food matrix" to record the specific needs of downstream customers in sugar-free tea drinks, baking and other scenarios for sweetness, solubility, thermal stability (≥120℃), and taste (bitterness, aftertaste duration).

[0071] Generative taste prediction: Using conditional generative adversarial networks (cGANs), predictive sensory description profiles and flavor wheels are generated based on the purity, addition ratio, and target food type of stevia.

[0072] Virtual screening: R&D personnel can quickly evaluate and compare the sensory prediction results of different formulations on a digital platform, thereby identifying the optimal application solution within 3 days, shortening the R&D cycle by 80% compared to the traditional trial-and-error method.

[0073] Step S4.4: Intelligent integration and document generation of new food ingredient application data Embedded legal knowledge: The application requirements of the National "Administrative Measures for the Safety Review of New Food Raw Materials" are structured and templated.

[0074] Automatic capture and filling of multi-source data: The AI ​​program automatically captures the necessary evidence chain data from the data hub, including: variety identification and stability reports, standardized toxicological test results (such as acute oral toxicity LD50 > 5000 mg / kg bw), production process documents that comply with the HACCP system, etc., and intelligently fills them into the corresponding chapters of the application template.

[0075] The generated initial application documents have an error rate of ≤0.5%, and manual review is mainly conducted to ensure compliance, which shortens the overall application preparation cycle by 30%.

[0076] 5. Construction and Implementation of the AI ​​Supply Chain Collaboration Subsystem This subsystem breaks down information silos in the industry chain, enabling demand-driven agile supply.

[0077] Step S5.1: Intelligent Demand Forecasting and Inventory Optimization Based on LSTM Demand Analysis: Access historical order data from key downstream customers over the past three years to analyze their monthly fluctuation patterns and seasonal peak characteristics.

[0078] Intelligent forecasting: Using a Long Short-Term Memory (LSTM) time series model, combined with constraints such as the company's own planting cycle (12 months) and processing capacity (500 kg / day), the demand for sucralose powder in the next 6 months is predicted in a rolling manner, and the mean absolute percentage error (MAPE) is controlled within ±5%.

[0079] Inventory strategy optimization: Based on forecast results and safety stock model, the system automatically generates the optimal raw material procurement plan and finished product inventory plan, which increases the overall inventory turnover rate by 40%.

[0080] Step S5.2: Precise Recommendation of Customized Sugar Substitute Solutions Based on Collaborative Filtering Customer profiling: Create multi-dimensional profiles for downstream enterprises, including: main product types (such as sugar-free tea drinks, low-sugar biscuits), target consumer groups (those who control their blood sugar, children), and preferences for the specifications of glycosides purchased in the past.

[0081] Intelligent Recommendation: We have developed a collaborative filtering recommendation algorithm based on matrix factorization. When a company submits a new product development request, the system proactively recommends the most suitable betaine product specifications (purity 98%-99.5%) and a scientifically appropriate addition ratio range (0.5%-2.0%) based on the company's profile and successful cases of similar companies. The matching accuracy rate is ≥90%.

[0082] Step S5.3: Trusted traceability across the entire industry chain based on consortium blockchain Blockchain architecture deployment: It adopts a permissioned consortium blockchain architecture, with node members including growers, processors, logistics providers, brand owners, and regulatory authorities.

[0083] Key data is stored on the blockchain: Key quality data from all stages of the entire industry chain (breeding gene ID, planting plot number and agricultural operation log, processing batch process parameters, laboratory quality inspection report) are generated into hash values ​​and written into the blockchain.

[0084] Transparent consumer access: Each final product is assigned a unique QR code. Consumers can scan the code to access immutable, end-to-end traceability information on the blockchain, significantly enhancing consumer trust. Market validation shows this can increase product repurchase rates by 25%.

[0085] 6. Construction of a Distributed Data Hub Hybrid data storage architecture: MySQL relational database is used to manage highly structured business data (orders, product information); Redis in-memory database is used to cache frequently accessed real-time data (sensor stream data, model prediction results) and session state to support high concurrency and low latency access requirements.

[0086] Real-time data synchronization and stream processing: Data streams from various subsystems are received through the Apache Kafka message queue, and stream processing frameworks are used for real-time cleaning, transformation and distribution to ensure that the end-to-end latency from data generation to consumption is ≤100 milliseconds.

[0087] Data governance and security: Data cleaning rules: For IoT sensor data, the 3σ principle is applied to automatically remove significant outliers; for gene sequence data, the format is standardized to FASTQ; for a small number of missing values ​​in phenotypic data, the K-nearest neighbor algorithm (K=5) is used to fill them.

[0088] Access control: Implement strict role-based access control (RBAC) and attribute-based access control (ABAC). For example, downstream customers can only query traceability information related to their orders and customized application solutions, but cannot access core breeding gene data.

[0089] 7. System Anomaly Handling Plan To ensure the stable and reliable operation of the whole industry chain optimization system in complex production environments, this invention designs a targeted abnormal situation handling mechanism.

[0090] (1) Emergency handling of sensor data interruption When the data flow of IoT sensors in the tissue culture room or planting base is interrupted due to network failure, equipment damage or other reasons, the system automatically triggers the following backup plan to maintain the continuity of control: Data substitution and prediction: The system automatically calls the average historical environmental data (such as light intensity, temperature, and soil moisture content) of the interrupted sensor locations for the same period over the past 3 days as the basic input. At the same time, combined with the Long Short-Term Memory (LSTM) network time series prediction model, based on the data trends of the associated sensors, it predicts the parameter changes during the interruption period and generates temporary environmental control or water and fertilizer solutions.

[0091] Fault warning: While activating the backup plan, the system immediately pushes a level 3 alarm to the administrator, including a system interface pop-up and a mobile phone text message, clearly indicating the location of the faulty device and the expected impact.

[0092] Data recovery and model correction: After the sensor resumes communication, the system automatically compares the predicted data with the actual recorded data. If the deviation exceeds a set threshold (e.g., temperature deviation > 2℃), the system automatically uses this batch of data as a new sample to incrementally learn the LSTM prediction model, optimize subsequent prediction accuracy, and ensure the continued effectiveness of the control scheme.

[0093] (2) Detection and correction of AI model prediction bias When the prediction output of the core AI model deviates significantly from the actual detection value, the system initiates a self-check and optimization process: Breeding prediction bias handling: If the deviation between the predicted stevia content and the actual measured value by the AI ​​breeding simulation system is >1% (e.g., predicted 12.5%, actual 11.2%), the system will perform the following steps: Data traceability: Automatically trace back the original sequencing data of the gene sequences of the associated parents to check for sequencing errors or data contamination.

[0094] Model recalibration: If data issues are ruled out and the problem is determined to be model parameter drift, then the incremental learning algorithm is initiated. The system automatically collects phenotypic data from 50-100 new samples in the current breeding batch as a supplementary training set to quickly retrain the "gene-trait association model" so that the prediction bias is restored to the normal range of ≤0.5%.

[0095] Extraction process prediction deviation handling: If the deviation between the predicted value of the glycoside extraction rate reinforcement learning model and the actual production value is >2%, the system will execute the following steps: Cause analysis: The system automatically analyzes the incoming inspection data of this batch of raw materials, focusing on checking indicators such as leaf maturity and moisture content detected by near-infrared spectroscopy, to determine whether the model is inaccurate due to differences in the physical properties of the raw material batches.

[0096] Model optimization: After confirming the cause, the characteristic spectral data of this batch of raw materials were incorporated into the environmental state space of reinforcement learning, and the weight coefficients of various indicators in the reward function were fine-tuned to improve the model's generalization ability and adaptability to new batches of raw materials.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing the entire industrial chain of Jinxiu sweet tea based on artificial intelligence, characterized in that, Includes the following steps: S1. Construct an AI-driven seed breeding system: Collect gene sequences and trait data of Jinxiu sweet tea and construct a gene-trait database; use an improved random forest algorithm to train a gene-trait association model and locate high-value gene fragments; predict the traits of hybrid offspring through an AI breeding simulation system and dynamically adjust breeding parameters by combining reinforcement learning to cultivate specialized varieties with a sweet tea glycoside purity ≥98% and a resistance to felt disease level ≤2. S2. Construct an AI-powered intelligent control subsystem for tissue culture seedling cultivation: Deploy IoT sensors in the tissue culture room to collect environmental parameters in real time; use a gradient boosting regression model to predict the survival rate of test seedlings and optimize the combination of environmental parameters through grid search; combine machine vision and near-infrared spectroscopy to automatically screen robust seedlings. S3. Construct a standardized AI planting and management subsystem: Integrate environmental data of Jinxiu area, and use AIGIS spatial analysis model to screen planting plots with a suitability score of ≥80; deploy drones and soil sensors, and generate precise water and fertilizer solutions based on LSTM time series prediction model; use improved YOLOv8 algorithm combined with meteorological data to construct a pest and disease control system; and use near-infrared spectroscopy to detect the content of sweet tea glycosides in leaves to trigger intelligent harvesting instructions. S4. Construct an AI-based subsystem for optimizing the deep processing of sucrose glycosides: Build a reinforcement learning environment based on the physicochemical properties of sucrose glycosides, train a process optimization model using the PPO algorithm; deploy online spectral detection equipment in the extraction and purification stages, and use AI to adjust process parameters in real time; construct a database of sucrose glycoside applications, and use generative adversarial networks to simulate product taste; AI automatically integrates data to generate new food ingredient application documents. S5. Construct an AI supply chain collaboration subsystem: Analyze downstream enterprise order data using an LSTM time series model to generate raw material procurement and inventory plans; based on enterprise product types and target audiences, push customized solutions through AI recommendation algorithms; use AI and blockchain technologies to record data throughout the entire process and generate a unique traceability QR code. S6. Build a distributed data hub, using a MySQL+Redis architecture to achieve real-time synchronization of data from the five subsystems, and establish data cleaning rules and permission management modules to ensure data quality and security.

2. The method according to claim 1, characterized in that, In step S1, the improved random forest algorithm uses out-of-bag error to filter gene fragment features and adjusts the number and maximum depth of decision trees through grid search to improve model R. 2 Coefficient ≥ 0.92, confidence level ≥ 90%.

3. The method according to claim 1, characterized in that, The input features of the LSTM time-series prediction model in step S3 include plant NDVI value, soil moisture content, nitrogen, phosphorus and potassium content, average temperature and precipitation over the past 7 days, and the output is the watering time and fertilization plan for the next 7 days.

4. The method according to claim 1, characterized in that, The formula for calculating the reward function of the reinforcement learning environment in step S4 is as follows: Reward = 0.4 × Extraction rate + 0.5 × Purity − 0.1 × (Energy consumption / 2) Extraction rate and purity are expressed as percentages, and energy consumption is expressed in kW.

5. The method according to claim 1, characterized in that, In step S5, the blockchain technology adopts a consortium blockchain architecture, with nodes including planting bases, deep processing workshops, downstream enterprises, and regulatory agencies. Data modifications require consensus verification from at least three nodes.

6. The method according to claim 1, characterized in that, The data cleaning rules in step S6 include the 3σ principle for sensor outlier removal, the standardization of gene sequence format to FASTQ format, and the use of the K-nearest neighbor algorithm for filling missing trait data values.

7. A system for implementing the method of any one of claims 16, characterized in that, include: Gene sequencing equipment used to collect the sweet tea gene sequence; Internet of Things (IoT) sensors are used to detect light intensity, temperature, and soil parameters. Multispectral drones are used for monitoring plant growth. Online spectrometer for detecting the purity of catechins; The server is used to deploy the five major AI subsystems and the distributed data hub. The client is used by users to view data, receive alerts, and issue commands.

8. The system according to claim 7, characterized in that, The server's hardware configuration includes: CPU ≥ Intel Xeon Gold 6330, memory ≥ 64GB, hard disk ≥ 2TB SSD, GPU ≥ NVIDIA A100; the client supports Windows 10 / 11 and macOS 12+ operating systems, and browsers support Chrome 90+ and Edge 90+.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 16.

10. The computer-readable storage medium according to claim 9, characterized in that, The storage medium includes at least one of USB flash drive, portable hard drive, ROM, RAM, and SSD, and the computer program includes an AI model training module, a data processing module, and a user interaction module.

Images