Medicinal and edible product screening method and system based on artificial intelligence

Abstract

The invention provides a medicinal and edible product screening method and system based on artificial intelligence, and the method comprises the steps: building a user physique and demand portrait through collecting user health condition data and food preference data; acquiring medicinal and edible raw material information and nutriology data based on the portrait, and constructing a medicinal and edible knowledge graph; analyzing market product information based on the knowledge graph, and generating a market insight report; generating an initial product formula scheme by using a large language model and a binary activation recurrent neural network; and outputting a mature product formula through trial production, user feedback and iterative optimization. According to the method, the problems that traditional medicinal and edible product development depends on expert experience and is lack of personalized and scientific basis are solved, intelligentization and data driving of the whole process of product development are realized, and the success rate and market adaptability of product development are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based method and system for screening medicinal and edible products, which are applied to the fields of healthy food development and personalized nutrition. Background Art

[0002] The development of food-and-medicine products is a key component of the modern health food industry. Combining traditional Chinese medicine theory with modern food science, the goal is to develop healthy foods that are both nutritious and beneficial for the body. With rising consumer health awareness, this sector is gaining market attention and experiencing rapid growth.

[0003] The development of traditional medicinal and edible products often relies on expert formula design. For example, some companies employ teams of Traditional Chinese Medicine experts to manually screen ingredients and design formulas, or they rely on traditional recipes based on historical experience. While these approaches can ensure basic product efficacy, they often struggle to adapt to the diverse needs of modern consumers.

[0004] Existing technologies for developing medicinal and edible products typically employ single-dimensional data analysis methods, such as screening ingredients solely based on ingredient databases or understanding user needs through simple questionnaires. While these technologies incorporate data analysis tools, their analytical dimensions are limited and they lack the ability to comprehensively process multi-source data. This makes it difficult for developed products to simultaneously strike a balance between efficacy, taste, and market demand.

[0005] The main problems with existing technologies include: First, the user constitution analysis and product matching lack scientificity and personalization, making it difficult to accurately match user health needs; second, the knowledge of the homology of medicine and food is insufficiently integrated with modern nutrition theory, resulting in a lack of scientific basis for product development; third, the product iteration and optimization process lacks data-driven methods, resulting in a long R&D cycle, high cost, and low success rate. Summary of the Invention

[0006] The purpose of the present invention is to provide an artificial intelligence-based method and system for screening medicinal and edible products, aiming to solve the problems in the existing technology of lack of personalization, insufficient scientific basis and low efficiency of iterative optimization in the development of medicinal and edible products.

[0007] To achieve the above objectives, the present invention provides an artificial intelligence-based method for screening medicinal and edible products, comprising: collecting health status data and food preference data of a user, analyzing the health status data and food preference data to obtain a user constitution and demand profile; based on the user constitution and demand profile, collecting medicinal and edible raw material information and nutritional data through automated collection technology, integrating the raw material information and nutritional data, and constructing a medicinal and edible knowledge graph; based on the medicinal and edible knowledge graph, collecting market product information, performing a multi-dimensional analysis on the market product information, and generating a market insight report; inputting the market insight report and the user constitution and demand profile into a trained large-scale language model, performing formula optimization through a binary-activated recurrent neural network, and generating an initial product formula plan; trial-producing the initial product formula plan, collecting user trial feedback data, iteratively optimizing the product formula plan based on the user trial feedback data, and outputting a mature product formula plan.

[0008] Collecting the user's health status data and food preference data, including: collecting the user's pulse, tongue and symptom information through an intelligent medical consultation system, classifying the pulse, tongue and symptom information based on the theory of traditional Chinese medicine differentiation of syndromes, and obtaining preliminary physical fitness assessment data; asking interactive questions based on the preliminary physical fitness assessment data, conducting in-depth research on the user's symptoms, collecting the user's preference information on food type, taste and form, and obtaining a user health status data set; designing a targeted food preference questionnaire based on the user's health status data set, collecting the user's multi-dimensional preference data on food, and obtaining food preference feature data; performing cluster analysis on the food preference feature data to generate the user's health status data and food preference data.

[0009] The health status data and food preference data are analyzed, including: based on the health status data, using a decision tree algorithm to classify and analyze user symptoms to obtain a user health feature vector; based on the food preference data, using a collaborative filtering algorithm to perform user group analysis to obtain a user food preference feature vector; based on the user health feature vector and the user food preference feature vector, using principal component analysis dimensionality reduction processing to generate a multi-dimensional user portrait; and performing cluster analysis on the multi-dimensional user portrait to obtain the user's physique and demand portrait.

[0010] The raw material information and nutritional data are integrated, including: performing data cleaning and standardization on the raw material information and nutritional data to obtain a standardized nutritional data set; designing an ontology model based on the standardized nutritional data set, defining the efficacy association, nutritional component association and indication association between materials, and constructing an initial knowledge graph; using an entity linking algorithm to fuse synonymous entities in the initial knowledge graph, and using a relational reasoning algorithm to supplement the implicit association between entities to obtain a complete entity relationship network; based on the complete entity relationship network, constructing an index structure that supports multi-dimensional retrieval, and establishing the medicine and food homology knowledge graph.

[0011] A multi-dimensional analysis is performed on the market product information, including: using image recognition and text analysis technology to extract product formulas, specifications and evaluation information, constructing a multi-dimensional scoring system that includes taste, price, nutritional value and health benefits, and obtaining market product scoring results; performing cluster analysis based on the market product scoring results to identify the type distribution of market products and obtain a competitive product analysis report; using association rules to mine the relationship between product features and user evaluations, combining the user physique and demand portraits to analyze the product preferences of different user groups and identify market gaps; based on the competitive product analysis report and market gaps, using time series analysis technology to establish a sales trend forecasting model, predict the market potential of different formulas, and generate the market insight report.

[0012] Formula optimization is performed through a binary-activated recurrent neural network, including: performing natural language processing on the market insight report to extract efficacy requirements, taste preferences and market positioning characteristics, and establishing a parameterized demand model; using domain adaptive fine-tuning technology to train professional knowledge on a large language model to enable it to acquire the ability to design medicine-food formulas and obtain a professional language model; based on the parameterized demand model and the professional language model, using Monte Carlo tree search technology to discretely optimize the formula to generate multiple candidate formulas; performing stability evaluation and multi-dimensional scoring on the multiple candidate formulas, using Pareto optimal solution theory to screen the optimal formula, and generating the initial product formula plan.

[0013] The Monte Carlo tree search technology is used to discretely optimize the formula, including: constructing a multi-objective optimization function based on the parameterized demand model, setting the efficacy index, taste index and cost index as optimization targets, and generating an optimization objective function; setting ingredient dosage restrictions and process feasibility constraints based on the optimization objective function, and using the Gumbel-Softmax method to discretely sample the formula ingredients to obtain initial formula parameters; iteratively adjusting the initial formula parameters using a proximal strategy optimization algorithm, performing formula search in a manner that maximizes the expected reward value, and outputting the optimized formula parameters; calculating the comprehensive score of the formula based on the optimized formula parameters and the multi-objective optimization function, and generating the multiple candidate formulas.

[0014] The initial product formula scheme is trial-produced and iteratively optimized, including: designing a standardized trial-production process based on the initial product formula scheme, using digital twin technology to simulate the production process, and generating product samples; designing an A / B testing plan, recruiting target users for product trials, collecting user feedback through structured questionnaires and in-depth interviews, and obtaining a user experience evaluation report; calculating the contribution value of each formula ingredient to user satisfaction based on the user experience evaluation report, determining the ingredients with satisfaction contribution values ​​lower than the threshold as objects to be optimized, and adjusting the formula parameters using a genetic algorithm; conducting market testing, using a survival analysis method to evaluate the product's market performance, and making final formula adjustments based on market feedback to obtain the mature product formula scheme.

[0015] A genetic algorithm is used to adjust the formula parameters, including: taking user satisfaction, efficacy evaluation and market potential as evaluation indicators, constructing a fitness function based on weighted summation, and setting the weight coefficient of each indicator; discretely encoding the dosage range of the ingredients to be optimized in the formula, designing a single-point crossover operator and a Gaussian mutation operator, and generating an initial population; calculating the fitness value of each individual in the initial population, selecting high-quality individuals according to a roulette wheel strategy, performing crossover and mutation operations on the high-quality individuals, and generating a new generation population; performing fitness evaluation and individual selection on the new generation population, stopping the iteration when the fitness value of the optimal individual has not improved for three consecutive generations, and outputting the formula parameters with the highest fitness.

[0016] The present invention also provides a device for screening products of medicinal and edible origin, for implementing the above method, the device comprising: a user portrait generation module for collecting and analyzing user health status data and food preference data; a knowledge graph construction module for constructing a knowledge graph of medicinal and edible origin; a market analysis module for generating market insight reports; a formula optimization module for generating an initial product formula plan; and a product iteration module for generating a mature product formula plan.

[0017] The artificial intelligence-based method and system for screening medicinal and edible products provided by the present invention have the following beneficial effects:

[0018] 1) By building a multi-dimensional user physique and needs profile, we achieve a precise match between medicinal and edible products and user health needs, improving the level of product personalization;

[0019] 2) By constructing a knowledge graph of medicinal and edible homology, integrating traditional Chinese medicine knowledge with modern nutrition data, a systematic scientific basis is provided for product formula design;

[0020] 3) The recipe optimization system using binary activation RNN breaks through the traditional reliance on experience in recipe design and can generate scientific and reasonable product recipes under multiple constraints;

[0021] 4) By establishing a data-driven product iteration and optimization closed-loop system, the success rate of product development and market adaptability have been significantly improved, and the R&D cycle has been shortened. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. The drawings herein are incorporated into and constitute a part of the specification. These drawings illustrate embodiments consistent with the present disclosure and, together with the specification, are used 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 those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without inventive effort.

[0023] Figure 1 A schematic flow chart of a method for screening medicinal and edible products based on artificial intelligence according to an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of the user data collection and physical fitness assessment system provided by an embodiment of the present invention;

[0025] Figure 3 A schematic diagram of the process of constructing a knowledge graph of medicinal and edible homology provided by an embodiment of the present invention;

[0026] Figure 4 A schematic diagram of a product recipe generation process based on a large language model provided by an embodiment of the present invention;

[0027] Figure 5 A schematic diagram of the product iterative optimization and feedback closed-loop process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. The components of the embodiments of the present disclosure generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the drawings is not intended to limit the scope of the disclosure for which protection is sought, but merely represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present disclosure.

[0029] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0030] The term "and / or" herein simply describes an association relationship, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, and the existence of B alone. In addition, the term "at least one" herein refers to any combination of at least two of any one or more of a plurality of items. For example, "at least one of A, B, and C" can represent any one or more elements selected from the set consisting of A, B, and C.

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0032] Example 1

[0033] like Figure 1 As shown, the embodiment of the present invention provides a method for screening medicinal and edible products based on artificial intelligence, comprising the following steps:

[0034] Step S1: collecting the user's health status data and food preference data, analyzing the health status data and food preference data to obtain a user's physique and demand profile;

[0035] Step S2: Based on the user's physique and demand profile, collect information on raw materials and nutritional data of medicine and food with the same origin through automated collection technology, integrate the raw material information and nutritional data, and construct a knowledge graph of medicine and food with the same origin;

[0036] Step S3: Based on the medicine and food homology knowledge graph, collect market product information, perform multi-dimensional analysis on the market product information, and generate a market insight report;

[0037] Step S4: Input the market insight report and user physique and demand profile into the trained large language model, perform formula optimization through a binary activated recurrent neural network, and generate an initial product formula solution;

[0038] Step S5: trial-produce the initial product formula scheme, collect user trial feedback data, iteratively optimize the product formula scheme based on the user trial feedback data, and output a mature product formula scheme.

[0039] like Figure 2 As shown, in step S1, collecting the user's health status data and food preference data specifically includes:

[0040] Step S1.1: The intelligent consultation system collects the user's pulse, tongue, and symptom information, classifies it based on Traditional Chinese Medicine (TCM) syndrome differentiation theory, and generates preliminary physical fitness assessment data. Specifically, based on classical TCM theory and analysis of modern medical literature, a multi-dimensional intelligent consultation question bank encompassing pulse, tongue, and symptom data is constructed. Incorporating natural language processing technology, an interactive consultation process is developed to achieve an intelligent and personalized consultation experience, ultimately outputting preliminary physical fitness assessment data.

[0041] To ensure the scientific nature and effectiveness of the questionnaire database, this system utilizes a three-tiered validation mechanism for continuous optimization. First, a systematic review of ancient Chinese medical texts and modern research literature using bibliometric methods extracts high-frequency symptom descriptors and diagnostic key points to construct an initial questionnaire framework. Second, a team of Chinese medicine experts conducts multiple rounds of review using the Delphi method to identify key issues and optimize the questionnaire logic, ensuring the professionalism of the questionnaire database. Third, an active learning algorithm is applied to continuously optimize the questionnaire database. The system automatically identifies ambiguities and redundant information in user feedback and dynamically adjusts the structure and presentation of questions. Furthermore, a medical data interoperability interface is designed. With user authorization, the system securely accesses medical history, test reports, and medication records from electronic health records (EHRs). This information is then integrated with real-time symptom data for analysis, significantly improving the accuracy and personalization of physical fitness assessments. To address the differences in user expression habits, the system also integrates a semantic understanding module that can identify synonymous symptom expressions and dialect expressions, ensuring the accuracy and comprehensiveness of information collection.

[0042] Step S1.2: Interactively ask questions based on the preliminary physical fitness assessment data, conduct in-depth research on user symptoms, collect user preference information on food types, tastes, and forms, and obtain a user health status data set; specifically, based on the preliminary physical fitness assessment data, classify and analyze user symptoms through a decision tree algorithm, combine with Bayesian network reasoning to understand the user's potential health status, and through multiple rounds of interactive questioning, deeply explore the user's health needs and output a structured user health status data set.

[0043] Step S1.3: Design a targeted food preference questionnaire based on the user health data set to collect multi-dimensional user preference data on food and obtain food preference feature data. Specifically, design a targeted food preference questionnaire based on the user health data set to collect multi-dimensional user preference data on food type, taste, form, etc. A collaborative filtering algorithm is used to analyze user groups and output a user food preference feature vector.

[0044] Step S1.4: Cluster analysis is performed on the food preference feature data to generate user health status data and food preference data. Specifically, the user health status dataset and food preference feature vectors are integrated, and user groups are divided using a clustering algorithm to construct a multi-dimensional user profile model. Principal component analysis is applied to reduce the dimensionality of complex features, ultimately generating a comprehensive user profile that encompasses physical characteristics, health needs, and food preferences.

[0045] In step S1, the health status data and food preference data are analyzed, including:

[0046] Based on the health data, a decision tree algorithm is used to classify and analyze user symptoms to obtain a user health feature vector. Specifically, the collected health data is first preprocessed, including missing value filling, outlier processing, and feature standardization. A decision tree algorithm, such as C4.5 or CART, is then used to classify and analyze the symptom data to identify the association rules between user symptoms and physical constitution types. The leaf nodes of the decision tree represent the probability distribution of different physical constitution types, and these probability values ​​are combined into a vector to serve as the user health feature vector. The decision tree is trained using a cross-validation method to optimize model parameters and ensure classification accuracy.

[0047] Based on the food preference data, a collaborative filtering algorithm is used to analyze user groups and obtain a user food preference feature vector. Specifically, a user-food preference matrix is ​​constructed to record users' ratings or selection frequencies for different food types, flavors, and forms. An item-based collaborative filtering algorithm is used to calculate the similarity between food items, recommend similar foods to each user, and record the recommendation strength. Simultaneously, matrix decomposition techniques (such as singular value decomposition or non-negative matrix decomposition) are used to decompose the high-dimensional and sparse user-food preference matrix into a low-dimensional dense matrix, extracting implicit features that constitute the user food preference feature vector.

[0048] Based on the user health feature vector and the user food preference feature vector, principal component analysis (PCA) is used for dimensionality reduction to generate a multidimensional user profile. Specifically, the user health feature vector and the food preference feature vector are concatenated into a high-dimensional feature vector, and principal component analysis (PCA) is used for dimensionality reduction. First, the feature covariance matrix is ​​calculated, its eigenvalues ​​and eigenvectors are solved, and the top k principal components with the largest contribution are selected as the new feature space. The original high-dimensional feature vector is mapped to this low-dimensional feature space to obtain a more compact user representation and form a multidimensional user profile.

[0049] Perform cluster analysis on the multi-dimensional user portrait to obtain the user's physique and needs portrait. Specifically, apply K-means or hierarchical clustering algorithm to perform cluster analysis on the user portrait after dimensionality reduction, and divide the users into multiple groups with similar characteristics. Use evaluation indicators such as silhouette coefficient or trough rate to determine the optimal number of clusters. Perform feature analysis on each cluster center to identify the typical characteristics of the group, such as the main physique type, common health needs and food preference patterns. Based on these characteristics, generate a structured user physique and needs portrait, including physique classification, health need priority and food preference labels.

[0050] like Figure 3 As shown, in step S2, the raw material information and nutritional data are integrated, including:

[0051] Step S2.1: Clean and standardize the raw material information and nutritional data to obtain a standardized nutritional dataset. Specifically, based on the initial medicinal and edible raw material dataset, connect to authoritative data sources such as the National Food Composition Database and the Nutrition Research Database through API interfaces to obtain scientific data such as the nutritional composition and functional factor content of each raw material. Apply data cleaning and standardization algorithms to eliminate data redundancy and inconsistencies, and output a standardized nutritional dataset.

[0052] Step S2.2: Design an ontology model based on the standardized nutrition dataset, define the efficacy, nutritional component, and indication associations between materials, and construct an initial knowledge graph. Specifically, design the ontology model to define the relationship types between materials, including relationship types such as "has efficacy," "contains ingredients," and "is suitable for symptoms." Define entity types such as materials, efficacy, ingredients, and symptoms, and their attributes, and formally describe the ontology model using OWL (Web Ontology Language). Based on the defined ontology model, convert the standardized nutrition dataset into a triple form (subject-relationship-object) to construct the initial knowledge graph.

[0053] Step S2.3: An entity linking algorithm is used to fuse synonymous entities in the initial knowledge graph, and a relational reasoning algorithm is used to supplement the implicit associations between entities, resulting in a complete entity relationship network. Specifically, string matching and semantic similarity calculations are used to identify synonymous entities, such as "wolfberry" and "wolfberry." The entity linking algorithm is applied to fuse synonymous entities, retaining the most standardized entity representation. Then, based on the existing explicit relationships, rule-based reasoning and statistical relationship prediction methods are used to infer implicit relationships between entities. For example, if ingredient A contains ingredient B, and ingredient B has efficacy C, then it is inferred that ingredient A is likely to have efficacy C. A confidence score is assigned to each inferred relationship.

[0054] Step S2.4: Based on the complete entity-relationship network, construct an index structure that supports multi-dimensional retrieval and establish the knowledge graph for the homology of medicine and food. Specifically, design a multi-level index structure to support retrieval by multiple dimensions, such as efficacy, ingredients, and indications. Create an inverted index for entities and relationships to accelerate query operations. Implement a similar material recommendation function based on graph embedding, mapping the entities and relationships in the graph into a low-dimensional vector space and discovering similar materials by calculating vector similarity. Ultimately, a complete knowledge graph for the homology of medicine and food is constructed, providing rich retrieval and reasoning capabilities.

[0055] In step S3, a multi-dimensional analysis is performed on the market product information, including:

[0056] Image recognition and text analysis techniques were used to extract product formulas, specifications, and review information, constructing a multidimensional scoring system encompassing taste, price, nutritional value, and health benefits, and generating market product ratings. Specifically, image recognition was first used to extract product formula information, including ingredient lists and nutritional information, from product packaging photos. Optical character recognition (OCR) was then applied to convert text within the images into processable text. Natural language processing was then used to analyze product descriptions and user reviews to extract keywords and phrases. A multidimensional scoring system was constructed, including a taste score (based on flavor descriptors and sensory-related content in user reviews), a price index (based on market positioning and price distribution of similar products), a nutritional value score (based on a comparison of nutrient content with recommended intakes), and a health benefit score (based on functional ingredient content and user feedback on perceived benefits). The analytic hierarchy process (AHP) was used to determine the weights of each dimension and output a comprehensive score.

[0057] Based on the market product ratings, cluster analysis is performed to identify the distribution of product types and generate a competitive product analysis report. Specifically, products are clustered based on their rating feature vectors, using algorithms such as K-means or DBSCAN to categorize them into different types. Common features and differences between product types are analyzed, including key target demographics, core selling points, and technological approaches. Market share, growth rate, and user rating distribution are calculated for each product type to identify leading brands and their competitive strategies. A competitive product analysis report is generated, including market segmentation and product positioning, competitive landscape analysis, and a SWOT analysis of key competitors.

[0058] Association rules are used to mine the relationship between product features and user reviews. Product preferences of different user groups are analyzed in combination with the user's physique and demand profiles to identify market gaps. Specifically, product review data, including ratings, text comments, and post-purchase behavior data, is collected to construct a product feature-user review dataset. Association rule mining algorithms such as Apriori or FP-Growth are used to discover association rules between specific product feature combinations and high reviews, and support, confidence, and lift are calculated. Users are grouped according to their physique and demand profiles, and the differences in preferences for product features among different user groups are analyzed to identify the unmet needs of specific user groups. By comparing the existing product distribution with the user demand distribution, areas of mismatch between product supply and user demand, i.e., market gaps, are discovered.

[0059] Based on the competitive product analysis report and market gaps, a sales trend forecasting model is established using time series analysis technology to predict the market potential of different formulas and generate the market insight report. Specifically, historical sales data and market trend data are collected, and time series analysis methods such as ARIMA or exponential smoothing are applied to predict sales trends of different product types. Combined with social media data analysis, sentiment analysis and topic modeling are used to identify emerging consumer trends and changes in user demand. Based on comprehensive competitive product analysis and consumer trends, the commercial potential of market gaps is evaluated, including potential market size, growth prospects and entry barriers analysis. The market potential prediction model is trained based on a machine learning algorithm to predict the potential market performance of different formula combinations, generate a comprehensive market insight report, and provide decision support for product formula design.

[0060] like Figure 4 As shown, in step S4, the recipe optimization is performed through a binary activated recurrent neural network, including:

[0061] Step S4.1: Natural language processing is performed on the market insight report to extract efficacy requirements, taste preferences, and market positioning characteristics, and a parameterized demand model is established. Specifically, the system receives as input a comprehensive market insight report and a comprehensive user profile. This data contains multi-dimensional information such as market trends, user constitution characteristics, health needs, and food preferences. During the parameter extraction phase, the system uses natural language processing technology to analyze the market report text, identifying keywords and phrases such as efficacy requirements such as "lowering blood pressure" and "improving sleep," taste preferences such as "sweet and sour" and "crisp," and market positioning characteristics such as "portable packaging" and "low-calorie." Furthermore, key information such as constitution type (e.g., "damp-heat constitution" and "qi deficiency constitution") and personal health goals is extracted from the user profile. The system uses a specially designed semantic mapping algorithm to convert the extracted qualitative descriptions into quantitative indicators. For example, "sour taste" is converted into a parameter constraint of a pH range of 5.0-6.0, and "improving sleep" is converted into a requirement for the content of specific functional ingredients (e.g., GABA and tryptophan). Next, the system constructs a multi-objective optimization function, which is a mathematical model used to express the multiple objectives that the product needs to meet simultaneously. This optimization function usually takes the form of a weighted sum: F = w1f1+w2f2+...+w n f n , where f1...f n Represents different objective functions (such as efficacy satisfaction, taste compatibility, cost control, etc.), w1...w nRepresents the weight of each objective. The weight distribution is automatically determined based on the importance analysis in the market insight report. At the same time, the system will also set a series of constraints, such as the maximum dosage limit of certain ingredients in the formula (based on safety considerations), the minimum content requirements of certain functional ingredients (based on effectiveness considerations), process feasibility constraints, etc. These constraints are expressed as mathematical inequalities: g1(x)≤0,g2(x)≤0,...,gm(x)≤0, where x represents the formula parameter vector. Ultimately, the system outputs a complete parameterized product demand model, including the optimization objective function, the constraint set, and the value range of each parameter. This model is stored in JSON or other structured formats as the mathematical basis for subsequent formula generation.

[0062] Step S4.2: Domain-adaptive fine-tuning techniques are used to train the large-scale language model with specialized knowledge, enabling it to acquire the ability to design food-drug homology formulas, thereby generating a specialized language model. Specifically, the system first selects a suitable pre-trained large-scale language model as a base model based on a parameterized product requirement model, such as GPT, LLaMA, or PaLM. These models already possess a broad knowledge base, but lack specialized knowledge in the field of food-drug homology. The system collects examples of food-drug homology formulas from multiple sources, including: published formula research in academic literature; formula information in patent databases; formula analysis of successful commercial products; and classic formulas recorded in traditional Chinese medicine texts. This raw data is structured and converted into pairs of "requirement descriptions" and "recipe designs" to construct a specialized domain dataset. Domain-adaptive fine-tuning techniques are used for fine-tuning. Specifically, unsupervised learning is used to adapt the model to the specialized terminology and expressions in the food-drug homology field. Then, supervised learning is used to train the model to understand the mapping between requirement parameters and recipe designs. Technically, methods such as low-rank adaptation (LoRA) or parameter-efficient fine-tuning (PEFT) are used to update only a small number of key parameters in the model, maintaining the general capabilities of the original model while acquiring professional domain knowledge and reducing computing resource requirements. To address the special challenges in the design of food-medicine formulas, such as interactions between ingredients and incompatibility taboos, the system has designed special training tasks. For example, through autoregressive prediction tasks, the model learns to predict suitable compatibility ingredients based on some existing formulas; through comparative learning tasks, the model distinguishes between reasonable and unreasonable formulas; and through multi-round dialogues, the formulator's thinking process is simulated. Model evaluation uses professional indicators, including: formula rationality score, in which domain experts judge the scientific nature of the generated formula; requirement satisfaction, which evaluates whether the formula meets the various requirements of the parameterized model; and innovation index, which measures the difference between the generated formula and existing solutions. Only models that meet all three requirements will be adopted.

[0063] For example, this system adopts a refined hyperparameter optimization strategy in the training process of large-scale language models. In view of the particularity of the field of medicine and food homology, a three-stage progressive training scheme is designed. In the first stage, a larger learning rate (5e-5) and batch size (batch size = 64) are used to coarse-tune the model with a wide range of domain knowledge; in the second stage, the learning rate (1e-5) and batch size (batch size = 32) are reduced, and the gradient accumulation technology is introduced to focus on the fine-tuning of professional formula design capabilities; in the third stage, an extremely small learning rate (5e-6) and batch size (batch size = 16) are used, combined with targeted adversarial sample training, to improve the model's processing ability for rare formula combinations. In order to solve the problem of long-tail distribution in the data of medicine and food homology, the system adopts weighted sampling technology to give higher weights to rare efficacy types and rare materials to balance the data distribution. In terms of regularization, it combines weight decay (weight decay = 0.01), attention dropout (attention dropout = 0.1), and hidden layer dropout (hidden dropout = 0.2) techniques to effectively prevent overfitting. The model training adopts a dynamic early stopping strategy, and automatically stops training when the performance of the validation set has not significantly improved for 5 consecutive epochs. In addition, the system has also designed a special field indicator evaluation system, including the formula rationality score (FRS), material compatibility index (MCI) and innovation score (IS), to comprehensively evaluate the quality of the formula generated by the model, and incorporate these indicators into the model selection process to ensure that the final applied model not only performs well in general NLP indicators, but also has practical value in the professional field of medicine and food homology.

[0064] Step S4.3: Based on the parameterized demand model and professional language model, the recipe is discretely optimized using Monte Carlo tree search technology to generate multiple candidate recipes. Specifically, the core of the system is a specially designed recursive neural network (RNN), whose notable feature is the use of a binary activation function. Unlike continuous activation functions such as Sigmoid or Tanh used in traditional RNNs, binary activation functions only output two states, 0 or 1. This design is highly consistent with the essence of recipe decision-making - an ingredient is either selected into the recipe (1) or not selected into the recipe (0). Technically, the system uses a "straight-through estimator" to solve the problem of binary functions being non-differentiable, using a true binary function in forward propagation and an approximate differentiable function in backpropagation. The system also uses a dedicated discretization algorithm, which is a technical innovation of this step. Traditional recipe optimization often uses continuous optimization methods such as gradient descent, and then discretizes by rounding, which often leads to suboptimal solutions. This system searches directly in discrete space, employing combinatorial optimization techniques, specifically a differentiable discrete sampling method based on Gumbel-Softmax. This enables the network to directly output discrete decisions while maintaining trainability. When processing the ratios of ingredients in a recipe, the system uses hierarchical quantization to map continuous ratios into discrete steps, such as 0.1%, 0.5%, and 1%. The system employs a reinforcement learning framework to guide the RNN in generating feasible recipes while satisfying multiple constraints. The system views recipe generation as a sequential decision-making process: at each step, an ingredient and its dosage are selected to gradually build a complete recipe. The quality of the generated recipe is evaluated using a custom reward function based on a parameterized demand model that comprehensively considers factors such as efficacy satisfaction, taste compatibility, and cost. Constraint violations, such as undesirable interactions between ingredients, are addressed using a penalty function. Technically, the system employs the Proximal Policy Optimization (PPO) algorithm to balance exploring new recipes with leveraging known good recipes. To improve system efficiency, the Monte Carlo Tree Search (MCTS) technique is also incorporated into the RNN model to form a hybrid architecture. MCTS helps the system avoid obvious suboptimal choices and significantly reduces the search space by simulating the future outcomes of multi-step decisions.

[0065] Step S4.4: Perform stability evaluation and multi-dimensional scoring on the multiple candidate formulas, use the Pareto optimal solution theory to screen the optimal formula, and generate the initial product formula solution. Specifically, the system uses the Monte Carlo simulation method to generate a large number of variant formulas by randomly perturbing the formula parameters (such as fine-tuning of ingredient dosage, processing temperature fluctuations, etc.), and then evaluates the performance distribution of these variants on key indicators. This method can identify "fragile" formulas - those whose performance drops sharply when the parameters change slightly. The simulation process is built based on physical and chemical models and historical experience data, and can predict important characteristics of the formula under different conditions, such as stability, taste changes, and efficacy fluctuations. For example, the system will calculate the stability curves of the active ingredients in the formula under different temperature and pH conditions, and exclude those solutions that may fail under actual production and storage conditions. The system has designed a comprehensive scoring index system, including: efficacy indicators (calculating potential efficacy based on the content and ratio of functional ingredients in the formula); sensory indicators (predicting taste, aroma, etc. based on ingredient properties and interactions); safety indicators (assessing potential risk ingredients in the formula and their interactions); production indicators (considering process complexity, raw material availability, production costs, etc.); and market indicators (predicting commercial potential by combining user profiles and market trends). Each indicator is further subdivided into multiple dimensions, forming a multi-level scoring tree. Technically, the system uses the Analytic Hierarchy Process (AHP) to determine the weights of each indicator and employs fuzzy comprehensive evaluation methods to address the interactions between indicators. When selecting the optimal formula, the system does not use a simple weighted total score ranking, but instead applies the Pareto optimal solution theory. In multi-objective optimization problems, few solutions can perform optimally across all dimensions. A Pareto optimal solution is one in which no other solution is superior in some dimensions and not inferior in others. By calculating the Pareto front, the system identifies a set of high-quality formulas that balance various indicators, rather than blindly pursuing the maximization of a single indicator. The final output of the initial product formula plan not only includes a detailed ingredient ratio table, but also includes: formula design instructions, explaining the reasons for selecting and the function of each ingredient; expected effect analysis, predicting the performance characteristics of the product in various aspects; process suggestions, providing preliminary production process parameters; risk analysis, pointing out possible uncertainties in the formula and issues that need to be focused on.

[0066] When using Monte Carlo tree search technology to perform discrete optimization of the recipe, it specifically includes:

[0067] Based on the parameterized demand model, a multi-objective optimization function is constructed, and the efficacy index, taste index and cost index are set as optimization targets to generate an optimization target function. Specifically, each requirement in the parameterized demand model is first converted into a quantifiable target function. For the efficacy index, according to the efficacy-ingredient association relationship in the knowledge graph, an efficacy contribution coefficient is assigned to each ingredient, and an efficacy target function f1(x)=∑(w i *c i ), where w i is the efficacy contribution coefficient of ingredient i, c i is the amount of ingredient i. For taste index, based on the sensory characteristics data of the ingredients, a taste prediction model f2(x) is constructed, which can predict the consistency between the taste characteristics (such as sweetness, sourness, texture, etc.) of different ingredient combinations and the target taste. For cost index, according to the market price and processing cost of each ingredient, a cost objective function f3(x) = ∑(p i *c i ), where p i is the unit cost of component i. The weight coefficients α, β, and γ of each objective function are determined by the hierarchical analysis method, and the comprehensive optimization objective function F(x) = α·f1(x)+β·f2(x)-γ·f3(x) is constructed. Note that the cost function is preceded by a negative sign because the goal is to minimize the cost.

[0068] Based on the optimization objective function, the ingredient usage limit and process feasibility constraint conditions are set, and the formula ingredients are discretely sampled using the Gumbel-Softmax method to obtain the initial formula parameters; specifically, the constraint conditions are first set, including: ingredient usage range constraint g1(x):min i ≤c i ≤max i , based on security and effectiveness considerations; total amount constraint g2(x):∑c i =1, ensuring that the total formula is 100%; the incompatibility constraint g3(x) prohibits specific ingredient combinations based on the ingredient interaction information in the knowledge graph; the process feasibility constraint g4(x) ensures that the formula can be produced under existing technical conditions. Then, the Gumbel-Softmax method is used to discretely sample the formula ingredients. First, the set of candidate ingredients is represented as one-hot encoding. For each ingredient i, its selection probability is calculated: p i =exp((g i +logπ i ) / τ) / ∑exp((g j +logπ j ) / τ), where g i is the noise sampled from a Gumbel distribution, π iis the initial probability of selecting an ingredient, and τ is the temperature parameter. During training, the temperature parameter τ is gradually reduced so that the probability distribution gradually approaches a discrete one-hot distribution. For each selected ingredient, the dosage is determined by sampling from preset discrete dosage levels (e.g., 0.1%, 0.5%, 1%, 5%, etc.) to generate the initial recipe parameters.

[0069] The initial recipe parameters are iteratively adjusted using a proximal policy optimization algorithm, and the recipe search is performed in a manner that maximizes the expected reward value, and the optimized recipe parameters are output; specifically, the proximal policy optimization (PPO) algorithm is used to iteratively optimize the initial recipe. First, the state space S is defined as the state of the current partial recipe, the action space A is the decision to add a certain ingredient and its amount, and the reward function R is a comprehensive score based on the optimization objective function F(x) and the constraint g(x). For recipes that violate the constraints, a negative reward is given. RNN is used as the policy network π(a|s) to output the probability of selecting the next ingredient and amount a under the current partial recipe state s. At the same time, a value function network V(s) is trained to estimate the value of the current state. In each training iteration, multiple recipe samples are generated using the current strategy π, and the advantage function A(s,a)=R(s,a)+γV(s')-V(s) is calculated. According to the objective function of the PPO algorithm: L=E[min(r t (θ)A t ,clip(r t (θ),1-ε,1+ε)A t )], where r t (θ)=π θ (a t |s t ) / π θ old(a t |s t ) is the policy ratio, which updates the policy network parameters θ. Through multiple iterations, the policy network learns a recipe generation strategy that maximizes the reward while satisfying the constraints and outputs the optimized recipe parameters.

[0070] The comprehensive score of the formula is calculated based on the optimized formula parameters and the multi-objective optimization function, and the multiple candidate formulas are generated. Specifically, the multi-objective optimization function F(x) is applied to the multiple groups of formula parameters generated after the PPO algorithm is optimized to calculate their comprehensive scores. At the same time, for each formula, its score on each individual indicator (efficacy, taste, cost, etc.) is calculated to form a multi-dimensional scoring vector. In order to improve the diversity of the formula, a deterministic clustering algorithm is used to cluster the generated formulas into different types, and the highest-scoring formula representative is selected from each type. Finally, a group of diverse, high-quality candidate formulas are output, each formula containing a complete list of ingredients and their dosage, as well as scoring data for each dimension. These candidate formulas will enter the next step of stability assessment and multi-dimensional scoring.

[0071] like Figure 5 As shown, in step S5, the initial product formula is trial-produced and iteratively optimized, including:

[0072] Step S5.1: Design a standardized trial production process based on the initial product formulation, simulate the production process using digital twin technology, and generate product samples. Specifically, design a standardized trial production process and quality control system based on the initial product formulation. Use digital twin technology to simulate the production process, predict and resolve potential process issues, and generate product samples that meet food safety standards.

[0073] Step S5.2: Design an A / B testing plan, recruit target users for product trials, collect user feedback through structured questionnaires and in-depth interviews, and generate a user experience evaluation report. Specifically, design a scientific A / B testing plan based on product samples and recruit test users who meet the target profile. Collect user experience data through structured questionnaires and in-depth interviews, apply sentiment analysis techniques to process user feedback, and generate a user experience evaluation report that includes taste ratings, perceived efficacy, and improvement suggestions.

[0074] Step S5.3: Based on the user experience evaluation report, the contribution of each recipe component to user satisfaction is calculated. Components with satisfaction contribution values ​​below a threshold are identified as targets for optimization, and the recipe parameters are adjusted using a genetic algorithm. Specifically, based on the user experience evaluation report, the genetic algorithm is used to automatically adjust the recipe parameters. A sensitivity analysis is performed to identify key components that influence user satisfaction, and targeted optimization is performed to output an improved iterative recipe solution.

[0075] Step S5.4: Conduct market testing, using survival analysis to evaluate product market performance. Based on market feedback, final formulation adjustments are made to obtain the mature product formulation. Specifically, based on the iterative formulation, small-scale market testing is conducted to collect sales and user feedback data in a real-world market environment. Survival analysis is then applied to evaluate product market performance, and final formulation adjustments are made based on feedback. A mature product formulation is then output, complete with the complete formulation, process parameters, and quality standards.

[0076] When using genetic algorithms to adjust recipe parameters, specifically including:

[0077] Using user satisfaction, efficacy evaluation, and market potential as evaluation indicators, a fitness function based on weighted summation was constructed, with weight coefficients assigned to each indicator. Specifically, three key evaluation indicators were extracted based on user experience evaluation reports. User satisfaction indicators include overall satisfaction scores, willingness to repurchase, and willingness to recommend, obtained through questionnaire data; efficacy evaluation indicators include user-perceived product efficacy, obtained through a specific symptom improvement rating scale; and market potential indicators include user willingness to pay, expected purchase frequency, and perceived product differentiation. Based on market strategy and product positioning, weight coefficients α, β, and γ were assigned to each indicator (satisfying α + β + γ = 1). A fitness function F(x) = α·S(x) + β·E(x) + γ·M(x) was constructed, where S(x) is the user satisfaction score, E(x) is the efficacy evaluation score, and M(x) is the market potential score. x represents the formulation parameter vector.

[0078] Discretely encode the dosage range of the ingredients to be optimized in the formula, design a single-point crossover operator and a Gaussian mutation operator, and generate an initial population; specifically, first determine the ingredients to be optimized through sensitivity analysis, calculate the contribution of each ingredient to user satisfaction, and select ingredients with a contribution lower than the preset threshold as optimization targets. For each ingredient to be optimized, discretize its possible dosage range into a finite range, such as dividing the 0-5% range into 6 levels of 0%, 0.5%, 1%, 2%, 3%, and 5%, and represent them with integers 0-5. Design the crossover operation of the genetic algorithm, using a single-point crossover operator: randomly select a cut point, exchange the dosage of the ingredients before and after the point of the two parent formulas, and generate two offspring formulas. Design the mutation operation, using a mutation operator based on Gaussian distribution: for the ingredient selected for mutation, add its dosage value to a Gaussian distribution N(0,σ 2 ) is then mapped to the nearest discrete gear. Based on the original recipe and its fine-tuned variants, an initial population containing multiple individuals is generated, and the population size is usually set to 30-50 individuals.

[0079] Calculate the fitness value of each individual in the initial population, select high-quality individuals according to the roulette wheel strategy, perform crossover and mutation operations on the high-quality individuals, and generate a new generation of population; Specifically, each formula individual in the initial population is converted into a product sample through small-batch trial production, organize small-scale user testing to collect evaluation data, and calculate the fitness value F(x) of each individual. The roulette wheel selection strategy is adopted, that is, the probability of an individual being selected is proportional to its fitness value: p(x i )=F(x i ) / ∑F(x jThrough multiple random samplings, high-quality individuals are selected as parents according to this probability distribution. The selected parents are randomly paired and a single-point crossover operation is performed with a preset crossover probability (usually 0.7-0.8) to generate offspring individuals. A Gaussian mutation operation is performed on all offspring individuals with a preset mutation probability (usually 0.1-0.2). The offspring individuals generated through crossover and mutation are merged to form the next generation population.

[0080] Perform fitness evaluation and individual selection on the new generation population. When the fitness value of the best individual has not been improved for three consecutive generations, stop the iteration and output the recipe parameters with the highest fitness. Specifically, perform fitness evaluation on each recipe individual in the new generation population again. Adopt an elite retention strategy to retain the few individuals with the highest fitness in the current generation (such as the top 10%) directly to the next generation. Compare the fitness value of the best individual of the current generation with the historical optimal value, and update the historical optimal record. Check the termination condition: If the fitness value of the best individual has not been improved for three consecutive generations, or the preset maximum number of iterations (usually 20-30 generations) has been reached, stop the iteration. Output the historical best individual as the final optimization result, which contains complete recipe parameters. Verify and test the optimal recipe to confirm its performance stability and repeatability.

[0081] Example 2

[0082] This embodiment provides a device for screening medicinal and edible products, which is used to implement the method described in Example 1. The device includes:

[0083] A user portrait generation module for collecting and analyzing user health status data and food preference data, configured to perform the function of step S1 in Example 1;

[0084] A knowledge graph construction module for constructing a knowledge graph of medicinal and edible homology, configured to perform the function of step S2 in Example 1;

[0085] A market analysis module for generating a market insight report, configured to perform the function of step S3 in embodiment 1;

[0086] A recipe optimization module for generating an initial product recipe, configured to perform the function of step S4 in Example 1;

[0087] The product iteration module for generating a mature product formula solution is configured to perform the function of step S5 in Example 1.

[0088] The above modules can be implemented by hardware or in the form of software functional units and stored in a storage medium. The above software functional units are stored in a storage medium and, when executed by a processor, perform the steps of the method described in the exemplary embodiment of the present invention. The above functional units can correspond to program codes that execute the corresponding steps in the method of the present invention, or the above functional units can correspond to software functional modules that execute the corresponding steps in the method of the present invention, and the software functional modules can be supported by corresponding hardware entities for operation.

[0089] The artificial intelligence-based method and system for screening medicinal and edible products provided by the present invention achieves precise matching of medicinal and edible products with user health needs by constructing a multi-dimensional user constitution and demand portrait, thereby improving the personalization level of products; by constructing a medicinal and edible knowledge graph, integrating traditional Chinese medicine knowledge with modern nutrition data, a systematic scientific basis is provided for product formula design; the formula optimization system using binary activated RNN breaks through the empirical dependence of traditional formula design and can generate scientific and reasonable product formulas under multiple constraints; by establishing a data-driven product iterative optimization closed-loop system, the success rate and market adaptability of product development are significantly improved, and the R&D cycle is shortened.

[0090] Example 3

[0091] The present invention also constructs an integrated development platform for medicinal and edible products. This platform integrates full-process functional modules such as user data collection, knowledge graph construction, formula optimization, and market feedback, providing an end-to-end product development solution. The platform adopts a microservices architecture design, with loosely coupled and highly cohesive functional modules, facilitating maintenance and expansion. The platform front-end adopts a responsive design and supports both PC and mobile access, providing differentiated interactive interfaces for developers, decision makers, and end users. In terms of operational convenience, the platform has designed an intelligent workflow engine that can automatically generate optimal workflows based on product type and development stage, significantly reducing operational complexity. In terms of data visualization, the platform integrates a multi-dimensional data analysis dashboard that supports real-time monitoring of product development progress, user feedback trends, and market performance indicators. Through an interactive knowledge graph visualization tool, it intuitively displays the efficacy correlations and compatibility patterns between raw materials. In terms of system security, the platform implements role-based access control and encrypted data storage to ensure the security of sensitive data. In addition, the platform provides an open API interface that supports docking with existing enterprise ERP, CRM, and other systems, enabling data interoperability and business process integration, significantly improving product development efficiency and collaboration.

[0092] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0093] It should be noted that those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. If these changes and modifications fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these changes and modifications.

[0094] The present disclosure also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program executes the steps of the method for screening medicinal and edible products based on artificial intelligence in the above-mentioned method embodiment. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0095] In addition, an embodiment of the present disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps of the method for screening medicinal and edible products based on artificial intelligence provided by any of the above embodiments of the present disclosure. For details, please refer to the above method embodiments and will not be repeated here.

[0096] The computer program product may be implemented in hardware, software, or a combination thereof. In one embodiment, the computer program product is implemented as a computer storage medium, which may be a volatile or non-volatile computer-readable storage medium. In another embodiment, the computer program product is implemented as a software product, such as a software development kit (SDK).

[0097] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment and devices can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. In the several embodiments provided in the present disclosure, it should be understood that the disclosed equipment, devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0098] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0099] In addition, each functional unit in each embodiment of the present disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0100] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present disclosure, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present disclosure. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0101] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present disclosure, which are used to illustrate the technical solutions of the present disclosure, rather than to limit them. The scope of protection of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-mentioned embodiments within the technical scope disclosed in the present disclosure, or replace some of the technical features therein with equivalents. Such modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. A method for screening medicinal and edible products based on artificial intelligence, characterized in that: include: Collecting the user's health status data and food preference data, analyzing the health status data and food preference data to obtain a user's physique and demand profile; Based on the user's physique and demand profile, information on raw materials and nutritional data of medicinal and edible substances are collected through automated collection technology, and the raw material information and nutritional data are integrated to construct a knowledge graph of medicinal and edible substances; Based on the medicine and food homology knowledge graph, collect market product information, perform multi-dimensional analysis on the market product information, and generate a market insight report; Input the market insight report and user physique and demand profiles into the trained large-scale language model, perform formula optimization through a binary activated recurrent neural network, and generate an initial product formula solution; The initial product formula scheme is trial-produced, user trial feedback data is collected, the product formula scheme is iteratively optimized based on the user trial feedback data, and a mature product formula scheme is output.

2. The method according to claim 1, characterized in that Collect user health status data and food preference data, including: The intelligent medical consultation system collects the user's pulse, tongue and symptom information, and classifies the pulse, tongue and symptom information based on the theory of traditional Chinese medicine differentiation of syndromes to obtain preliminary physical fitness assessment data; Interactively ask questions based on the preliminary physical fitness assessment data, conduct in-depth research on user symptoms, collect user preference information on food types, tastes, and forms, and obtain a user health status dataset; Designing a targeted food preference questionnaire based on the user health status dataset to collect the user's multi-dimensional preference data on food and obtain food preference characteristic data; Cluster analysis is performed on the food preference characteristic data to generate user health status data and food preference data.

3. The method according to claim 1, characterized in that Analyzing the health status data and food preference data, including: Based on the health status data, a decision tree algorithm is used to classify and analyze the user's symptoms to obtain a user health feature vector; Based on the food preference data, a collaborative filtering algorithm is used to perform user group analysis to obtain a user food preference feature vector; Based on the user health feature vector and the user food preference feature vector, principal component analysis dimensionality reduction processing is used to generate a multi-dimensional user portrait; Perform cluster analysis on the multi-dimensional user portraits to obtain the user's physique and demand portraits.

4. The method according to claim 1, wherein Integrate the raw material information and nutritional data, including: Performing data cleaning and standardization on the raw material information and nutritional data to obtain a standardized nutritional data set; Designing an ontology model based on the standardized nutrition data set, defining the efficacy association, nutritional component association, and indication association among materials, and constructing an initial knowledge graph; An entity linking algorithm is used to fuse synonymous entities in the initial knowledge graph, and a relational reasoning algorithm is used to supplement implicit associations between entities to obtain a complete entity relationship network; Based on the complete entity relationship network, an index structure supporting multi-dimensional retrieval is constructed, and the medicine and food homology knowledge graph is established.

5. The method according to claim 1, wherein Conduct multi-dimensional analysis of the market product information, including: Image recognition and text analysis technologies are used to extract product formulas, specifications, and evaluation information, and a multi-dimensional scoring system is constructed that includes taste, price, nutritional value, and health benefits to obtain market product rating results. Perform cluster analysis based on the market product scoring results to identify the type distribution of market products and obtain a competitive product analysis report; Use association rules to mine the relationship between product features and user reviews, analyze product preferences of different user groups based on user characteristics and demand profiles, and identify market gaps; Based on the competitive product analysis report and market gaps, time series analysis technology is used to establish a sales trend forecasting model to predict the market potential of different formulas and generate the market insight report.

6. The method according to claim 1, characterized in that Recipe optimization via a binary activated recurrent neural network, including: Performing natural language processing on the market insight report to extract efficacy requirements, taste preferences, and market positioning characteristics, and establishing a parameterized demand model; Domain-adaptive fine-tuning technology is used to train large language models with professional knowledge, enabling them to design formulas for both medicinal and edible ingredients, thereby obtaining a professional language model. Based on the parameterized demand model and the professional language model, the recipe is discretely optimized using the Monte Carlo tree search technique to generate multiple candidate recipes; The multiple candidate formulations are subjected to stability evaluation and multi-dimensional scoring, and the optimal formulation is screened using Pareto optimal solution theory to generate the initial product formulation solution.

7. The method according to claim 6, characterized in that Monte Carlo tree search techniques are used to perform discrete optimization of the recipe, including: Constructing a multi-objective optimization function based on the parameterized demand model, setting the efficacy index, taste index and cost index as optimization targets, and generating an optimization objective function; Based on the optimization objective function, ingredient usage limits and process feasibility constraints are set, and the Gumbel-Softmax method is used to discretely sample the recipe ingredients to obtain initial recipe parameters; Iteratively adjusting the initial recipe parameters using a proximal strategy optimization algorithm, searching for a recipe in a manner that maximizes the expected reward value, and outputting optimized recipe parameters; A comprehensive score of the recipe is calculated based on the optimized recipe parameters and the multi-objective optimization function to generate the multiple candidate recipes.

8. The method according to claim 1, characterized in that Conduct trial production and iterative optimization of the initial product formulation, including: Design a standardized trial production process based on the initial product formula, use digital twin technology to simulate the production process, and generate product samples; Design A / B testing plans, recruit target users for product trials, collect user feedback through structured questionnaires and in-depth interviews, and generate user experience evaluation reports; Calculating the contribution value of each recipe component to user satisfaction based on the user experience evaluation report, determining the components with satisfaction contribution values ​​lower than a threshold as objects to be optimized, and adjusting the recipe parameters using a genetic algorithm; Conduct market testing, use survival analysis methods to evaluate product market performance, make final formula adjustments based on market feedback, and obtain the mature product formula solution.

9. The method according to claim 8, characterized in that Genetic algorithms are used to adjust the recipe parameters, including: Taking user satisfaction, efficacy evaluation and market potential as evaluation indicators, a fitness function based on weighted summation is constructed, and the weight coefficient of each indicator is set; Discretely encode the dosage range of the ingredients to be optimized in the formula, design a single-point crossover operator and a Gaussian mutation operator, and generate an initial population; Calculate the fitness value of each individual in the initial population, select high-quality individuals according to the roulette strategy, perform crossover and mutation operations on the high-quality individuals, and generate a new generation of population; The fitness of the new generation population is evaluated and individuals are selected. When the fitness value of the best individual does not improve for three consecutive generations, the iteration is stopped and the formula parameters with the highest fitness are output.

10. A device for screening medicinal and edible products, characterized in that: For implementing the method according to any one of claims 1 to 9, the device comprises: A user profile generation module for collecting and analyzing user health status data and food preference data; A knowledge graph construction module for constructing a knowledge graph of medicinal and edible homology; Market analysis module for generating market insight reports; A recipe optimization module for generating initial product recipe solutions; A product iteration module for generating mature product formula solutions.

Images