Food formulation quantitative evaluation and innovation analysis method and system

Abstract

This invention relates to the interdisciplinary field of artificial intelligence and the food industry. The method and system for quantitative evaluation and innovative analysis of food formulas include S1. acquiring target formula data; S2. multi-dimensional feature extraction; S3. calculating dimensional similarity; S4. calculating comprehensive similarity score; and S5. generating an innovative analysis report. This invention achieves in-depth, objective, and interpretable quantitative evaluation of formula similarity by extracting features and fusing similarity in three dimensions: raw materials, proportions, and functions.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and the food industry, particularly to methods and systems for quantitative evaluation and innovative analysis of food formulations. Background Technology

[0002] The development of new food products is highly dependent on the design and optimization of the formula. In the early stages of R&D, it is crucial to conduct a comprehensive similarity assessment and innovation analysis of candidate formulas, which helps to mitigate intellectual property risks and identify product advantages.

[0003] Currently, this process mainly relies on the human experience of R&D personnel, subjectively comparing data by consulting patents, literature, and product labels, which is inefficient and inconsistent. Although some general text similarity algorithms (such as models based on TF-IDF or BERT) can be applied, they have fundamental flaws: First, they cannot effectively parse and understand the structured information of the formula, for example, they cannot distinguish the fact that the raw materials "90% water, 10% salt" and "10% salt, 90% water" have the same ratio vector despite the textual difference; second, they lack expertise in the field of food science and cannot understand the functional similarities and differences between "baking powder" and "baking soda"; finally, their output is usually just an abstract similarity score, unable to indicate in which specific aspects are similar or different, let alone assess the innovative value of the differences, thus making it difficult to provide direct and effective guidance for R&D decisions.

[0004] Therefore, existing technologies suffer from technical bottlenecks such as low analytical efficiency, high subjectivity, inability to conduct in-depth quantification, and a complete lack of ability to proactively identify and value-judge innovative points. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system that can automatically, efficiently and accurately quantify the similarity of food formulations, and can proactively identify innovative differences, assess their value and generate research and development inspiration.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The core of this method for quantitative evaluation and innovation analysis of food formulations lies in constructing an analytical framework incorporating domain knowledge to transform formulation information into computable multi-dimensional feature vectors, and then utilizing an improved fusion algorithm for comprehensive evaluation and innovation discovery. This method includes the following steps:

[0008] S1. Obtain target formula data; receive text or structured data of the target formula through a user interface or data interface;

[0009] S2. Multi-dimensional feature extraction: Based on a predefined food ingredient knowledge base, extract at least two dimensions of features from the target formula data. These dimensions include ingredient identity features, ingredient ratio features, and functional combination features; specifically:

[0010] Raw material identification characteristics ( ): Extract the set of all raw materials in the formula, and standardize (e.g., unify synonyms) and classify functions through a knowledge base, and finally represent it as a multi-hot encoded vector or embedding vector.

[0011] Raw material proportion characteristics ( ): Extract the amount or proportion of each raw material, form a numerical vector, and perform normalization (e.g., L2 normalization) to eliminate the influence of the total amount.

[0012] Functional combination features ( Based on the functional roles of ingredients defined in the food ingredient knowledge base, the ingredients in the target formula are aggregated into one or more functional categories they fulfill, forming a function-ingredient mapping relationship, and a functional combination feature vector is generated based on this mapping relationship. For example, all ingredients that fulfill the "sweetness" function are aggregated into a single functional module.

[0013] S3. Calculate dimensional similarity: Calculate the similarity between the features of at least two dimensions of the target formulation and the features of the corresponding dimensions of the reference formulation. For each feature dimension, calculate using a similarity metric suitable for its feature representation:

[0014] For features in vector form (such as matching features and functional combination features), cosine similarity calculation is preferred:

[0015]

[0016] Where · denotes the dot product operation of vectors. This represents the L2 norm of a vector.

[0017] : No. The cosine similarity score on each feature dimension typically ranges from 1 to 10. When all elements of the feature vector are non-negative, the value range is: ;

[0018] The new formula to be evaluated is in the first stage. Feature vectors in each feature dimension;

[0019] The reference formula is in the [section / section]. Feature vectors in each feature dimension;

[0020] Vector dot product operation, defined as

[0021] The L2 norm (Euclidean norm) of the eigenvectors of the new recipe is defined as follows:

[0022] The L2 norm of the reference formula eigenvector is defined as follows:

[0023] For features in the form of sets (such as raw material identity features), Jaccard similarity calculation is preferred:

[0024]

[0025] in, The cardinality of a set;

[0026] : No. Jaccard similarity scores on each feature dimension, with values ​​ranging from 1 to 10. ;

[0027] The new formula to be evaluated is in the [number]th [year]. A set of features across a feature dimension;

[0028] The reference formula is in [section number]. A set of features across a feature dimension;

[0029] The intersection of the feature sets of the new formula and the reference formula;

[0030] The union of the feature sets of the new formula and the reference formula;

[0031] The cardinality of a set is the number of elements in the set.

[0032] S4. Calculate the overall similarity score. Based on the similarity of the at least two dimensions and the weights configured for each dimension, calculate the overall similarity score between the target formulation and the reference formulation. The weights are adaptive, and their values ​​are dynamically adjusted according to the attributes of the target formulation and / or the reference formulation, or the analysis focus specified by the user. Overall Similarity Score Calculated using the following formula:

[0033]

[0034] in, The number of feature dimensions, For the first Adaptive weights for each dimension For the first Similarity across multiple dimensions.

[0035] The adaptive dynamic adjustment of weights refers to the weight coefficients. It is not fixed, but rather calculated in real time by a weighted decision function f based on the input parameters. Its core mechanism is as follows:

[0036] 1. Input signals for weighted decision-making

[0037] The weighted decision function f receives one or more of the following input signals:

[0038] Signal A: Recipe Type Attribute (T) The system assigns a type label to a recipe by analyzing its ingredient composition or user-specified information, such as T = {Beverages, Baked Goods, Meat Products, Health Foods, Seasonings...}. Different types of recipes have different core competitive dimensions.

[0039] Signal B: User-specified analysis focus (U)

[0040] Users can select the focus of this analysis from the drop-down menu or checkbox, such as U = {Focus on cost, focus on functionality, focus on cleanliness label, focus on taste...}.

[0041] Signal C: Domain Prior Knowledge (D) This is a predefined rule base that reflects the consensus of food science and specifies the basic importance of each dimension of various formulations when there are no special requirements.

[0042] 2. The core mechanism of weighted decision-making

[0043] The weight decision function f dynamically generates a weight set based on the above input through one or more of the following mechanisms: ,W2,W3}:

[0044] 1) Table lookup method based on recipe type template

[0045] A weight configuration lookup table is pre-stored within the system.

[0046] Workflow: When the target formula is identified as "beverage", the system automatically loads the corresponding weight set (0.2, 0.5, 0.3).

[0047] 2) Rule mapping based on user focus

[0048] This mechanism allows users to override default type templates. A set of user focus mapping rules is pre-stored in the system.

[0049] Rule U1: IF "Focus on costs" THEN { }

[0050] Logic: Costs are primarily determined by the raw materials themselves (their identity), therefore their weight is increased; functionality is not a primary consideration for the time being.

[0051] Rule U2: IF "Focus on functionality" THEN { }

[0052] Logic: Focus all efforts on functional innovation of the formula, and maximize the weight of functional dimensions.

[0053] Rule U3: IF "Pay attention to cleaning labels" THEN { }

[0054] Logic: Cleanliness labels focus on "whether it contains unfriendly ingredients," which falls under the category of ingredient identification.

[0055] Workflow: Based on the "Beverage" weight template (0.2, 0.5, 0.3), if the user selects "Focus on Functionality", then rule U2 is executed to adjust the weight to (0.2, 0.4, 0.6).

[0056] 3) Feedback-based fine-tuning based on similarity results

[0057] This mechanism enables the system to possess a certain degree of "thinking" ability. After calculating the basic similarity, it performs a secondary fine-tuning of the weights.

[0058] Rule F1: IF THEN .

[0059] If two formulas are already highly similar in terms of "raw material identity", then reduce the weight of this dimension, while increasing the weight of "ratio" and "function" to amplify their differences in key details and avoid the conclusion of "high similarity" from obscuring important micro-innovations.

[0060] Rule F2: IF THEN .

[0061] Logic: If two formulas are completely different in "function", this is a very strong signal of differentiation. Therefore, the weight of this dimension should be increased significantly to ensure that the overall score can significantly reflect this fundamental difference.

[0062] 4) Dynamically Adjusted Comprehensive Execution Process

[0063] In step S4, the system performs weight adjustments in the following order:

[0064] Initialization: Initialize the basic weights from the lookup table based on the recipe type (T) and domain prior (D). .

[0065] User coverage: If the user specifies the analysis focus (U), then the analysis is performed according to the user focus mapping rules. After making the first adjustment, we get

[0066] Feedback fine-tuning (optional): After calculating the initial dimensional similarity... Then, based on the feedback-based fine-tuning rules, A second fine-tuning was performed to obtain the final result used for calculation. weight .

[0067] Normalization: Ensure .

[0068] S5. Generate an innovation analysis report based on the differences between the target formulation and the reference formulation.

[0069] Furthermore, the generation of the innovation analysis report in step S5 specifically includes:

[0070] S51. Identify at least one difference between the target formulation and the reference formulation. ;

[0071] S52. By querying the knowledge graph, assess the innovation contribution of the aforementioned points of difference and obtain a contribution score. ;

[0072] S53. Based on the contribution score, generate an analysis report that includes interpretations of innovation points and / or R&D insights.

[0073] Furthermore, the knowledge graph construction and query scheme in step 52.

[0074] The knowledge graph, a food innovation knowledge graph stored in a graph structure (hereinafter referred to as the "graph"), is used for deep semantic understanding and value assessment of differences.

[0075] 1. Definition of the graph structure of a knowledge graph: The graph consists of "nodes" and "edges".

[0076] Node types include:

[0077] Raw materials: Corresponding to the standard names in the knowledge base.

[0078] Function: Corresponds to the functional roles in the knowledge base.

[0079] Technical effects: such as "improving stability", "enhancing satiety", and "improving emulsification".

[0080] Market trends include "clean label", "plant-based", "high protein", and "reduced sugar".

[0081] Potential risks: such as "may cause gastrointestinal discomfort", "poor thermal stability", "regulated limits".

[0082] The relationship types of edges include:

[0083] Function: Connects raw material nodes and functional nodes.

[0084] Achieved effect: Connects raw material or functional nodes with technical effect nodes.

[0085] Align with trends: Connect raw materials, functional or technological effects with market trend nodes.

[0086] Risk exists: Connecting raw material nodes with potentially risky nodes.

[0087] Antagonistic / synergistic effects exist: connect two raw material nodes and describe their interaction.

[0088] 2. Knowledge Graph Construction Method: The knowledge graph is constructed in the following way:

[0089] Structured data import: Import the relationships (such as raw material-function) from the aforementioned "Food Raw Material Knowledge Base" into the graph as the basic data.

[0090] Unstructured data mining: Using text mining and entity relationship extraction techniques, nodes such as technical effects, market trends, and potential risks are automatically or semi-automatically extracted from a large number of food science papers, patent texts, and industry reports, as well as their correlation with raw materials and functions, and added to the graph.

[0091] Expert verification: Domain experts review and correct the automatically extracted relationships to ensure the accuracy of the graph knowledge.

[0092] 3. Knowledge Graph Query and Contribution Evaluation Process: In step S52, the system evaluates the differences through the following process. Innovation contribution score ):

[0093] Difference Analysis: This involves analyzing the identified differences. (e.g., "New ingredient: probiotic powder") is converted into a graph query request.

[0094] Graph traversal query: The system executes a predefined graph traversal algorithm. Taking the differences mentioned above as an example, the query will:

[0095] Locate the raw material node: "probiotic powder".

[0096] Traverse its functional edges to find its function (such as "gut health").

[0097] Traverse the edges of the achieved effects to find the relevant technical effects (such as "regulating gut microbiota").

[0098] Traverse the edges that match the trend to find the relevant market trends (such as "digestive health" and "immune support").

[0099] Check for potential risks and assess potential issues (such as "requires cold chain storage and transportation").

[0100] Contribution Measurement: The system calculates a contribution score (Score) based on the number and weight of associated nodes returned by the query, and a predefined scoring rule base. ).

[0101] Example of a rating rule library:

[0102] IF (related to the "high-heat" market trend) THEN +N+N points.

[0103] IF (achieved a "novel" technical effect) THEN +M+M points.

[0104] IF (High level of risk exists) point.

[0105] Ultimately, Score ( The score is the weighted algebraic sum of all these positive and negative scores. This score collectively reflects the net value of the difference across the technology, market, and risk dimensions.

[0106] Furthermore, in step S52, the sign and magnitude of the contribution score are determined based on predefined rules that associate the differences with their potential impact on the formulation's technical effectiveness, market value, or stability.

[0107] Furthermore, the contribution score "determined based on predefined rules" in step S52 is achieved through a structured and computable "innovation contribution scoring rule base." This rule base encodes food industry expertise, market insights, and regulatory requirements into explicit IF-THEN logic and quantitative indicators, enabling the identification of points of difference. The evaluation process is automated and made objective.

[0108] 1. Composition of the scoring rule base

[0109] The rule base consists of multiple scoring dimensions and specific rules under them. Each rule includes triggering conditions and scoring operations.

[0110] Dimension 1: Technological Novelty and Advancement

[0111] Rule T1 (Technological Scarcity):

[0112] If the core ingredients or technology combination involved in the difference point appears less than 5% of the reference formula library for the target product category (such as "yogurt"), then THEN Score = Score + 8.

[0113] If the frequency of IF is between 5% and 20%, then THEN Score = Score + 4.

[0114] If the frequency of IF is higher than 20%, then THEN Score = Score + 0.

[0115] Rule T2 (Technical Effect Intensity):

[0116] If the knowledge graph is linked to a "disruptive" technological effect (such as "achieving heatless processing"), then THEN Score = Score + 10.

[0117] If the IF is associated with a “significant improvement” effect (such as “50% increase in shelf life”), then THEN Score = Score + 6.

[0118] If it relates to a "general" effect (such as "taste adjustment"), then THEN Score = Score + 2.

[0119] Dimension Two: Alignment with Market Trends

[0120] Rule M1 (Trend Popularity):

[0121] If the differences are linked to market trends in the "introduction phase" (such as "precision nutrition") through the knowledge graph, THENScore = Score + 3.

[0122] If it relates to a market trend in a "rapid growth phase" (such as "plant-based"), then THEN Score = Score + 8.

[0123] If the trend is related to a "mature" market trend (such as "low-fat"), then THEN Score = Score + 4.

[0124] Rule M2 (Consumer Group Coverage):

[0125] If the difference point can meet the needs of multiple consumer groups at the same time (such as "suitable for both vegetarians and lactose intolerant people"), then THEN Score = Score + 5.

[0126] Dimension Three: Health and Regulatory Attributes

[0127] Rule H1 (Health Gain):

[0128] The IF difference point achieves "core nutrient optimization" (such as "30% reduction in sugar" or "increased dietary fiber"), THENScore = Score + 7.

[0129] The IF difference point achieves "reduction or elimination of unhealthy components" (such as "zero trans fatty acids"), THEN Score = Score + 5.

[0130] Rule H2 (Regulatory Compliance):

[0131] If the raw materials used at the difference point are subject to regulatory restrictions or use beyond permitted scope in the target market region, (A veto-type punishment).

[0132] Dimension Four: Impact of Process and Cost

[0133] Rule P1 (Process Simplification):

[0134] If the difference point can simplify the production process (such as "eliminating a heating step"), then THEN Score = Score + 4.

[0135] Rule P2 (Cost Impact):

[0136] The IF difference point led to a significant increase in the cost of core raw materials (>15%). .

[0137] If the IF difference point leads to a significant decrease in the cost of core raw materials (>15%), then THEN Score = Score + 3.

[0138] 2. Implementation process of scoring rules

[0139] In step S52, the system executes the following process to calculate :

[0140] initialization: =0.

[0141] Feature extraction: from differences Extract features that can be recognized by the rule engine, including:

[0142] Difference types: Addition of raw materials, replacement of raw materials, deletion of raw materials, and changes in proportion.

[0143] Raw materials involved: Specific standard raw material names.

[0144] Range of change: The amount of numerical change in proportion to a change in ratio.

[0145] Rule triggering and execution:

[0146] The rule engine matches the extracted features with the IF part in the rule base.

[0147] All triggered rules, according to their THEN part Add or subtract.

[0148] Example: For the difference point δ = "Replace white sugar with stevia".

[0149] Triggering rule T1: Query the database. "Stevia" appears in the "beverage" category with a frequency of 15%, which is "less common". Therefore, Score = 0 + 4.

[0150] Triggering rule M1: The knowledge graph shows that "stevia" is strongly associated with the "sugar reduction / sugar-free" trend (rapid growth period), so Score=4+8.

[0151] Triggering rule H1: This replacement directly achieves "100% sugar reduction", which belongs to core nutrient optimization, so Score=12+7.

[0152] Triggering rule P2: Cost calculation shows that the cost of stevia is slightly higher, but the difference is less than 15%, so the rule is not triggered.

[0153] Final score: =19.

[0154] Score standardization (optional): To prevent a certain dimension from being scored too high, the final total score can be normalized, for example, by using the Sigmoid function to compress it to the range of -10 to +10.

[0155] 3. Construction and maintenance of the rule base

[0156] The rules in the rule base are derived from the summarization and quantification of food science literature, market analysis reports, consumer survey data, and regulatory documents.

[0157] The rules and their weight coefficients in the rule base are configurable, allowing users to adjust them according to different evaluation focuses (such as prioritizing cost or health).

[0158] Machine learning optimization: In a preferred embodiment, the system can record user feedback on the evaluation results and use this feedback data to dynamically optimize the rule weights through a machine learning model (such as reinforcement learning) so that the scores are more in line with the acceptance of the actual market.

[0159] Through the aforementioned detailed and quantifiable rule base, this invention transforms the "innovation assessment" process, which originally relied on expert subjective experience, into a transparent, repeatable, and computable automated process. This is not only an improvement in efficiency but also a revolution in assessment methodology, ensuring the objectivity and consistency of the analysis results.

[0160] Furthermore, step S5 also includes: calculating the overall innovation index of the target formulation relative to the reference formulation based on the contribution scores of all identified differences.

[0161] Based on the contribution scores of all identified differences, the overall innovation index of the target formulation relative to the reference formulation is calculated. :

[0162]

[0163] in, The total number of differences. For the first The importance coefficient of each difference point A normalization function (such as the Sigmoid function) is used to map the weighted sum to a uniform interpretation range.

[0164] A food formulation analysis system for performing the above-described method includes one or more processors; and a memory for storing computer-executable instructions; when the computer-executable instructions are executed by the processors, the system is controlled to perform the steps of the method described above.

[0165] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0166] This invention relates to a method and system for quantitative evaluation and innovation analysis of food formulations. It possesses the following beneficial effects:

[0167] 1) By extracting features and fusing similarity across three dimensions—raw materials, proportions, and functions—a deep, objective, and interpretable quantitative assessment of formula similarity was achieved.

[0168] 2) By evaluating the contribution of differences based on knowledge graphs and calculating the overall innovation index, the innovative value of formulations can be proactively identified and quantified, providing a clear direction for research and development.

[0169] 3) The adaptive weighting mechanism and the built-in food domain knowledge base make the analysis process more intelligent and the results more professional and reliable.

[0170] 4) The generated integrated report contains similarity, innovation and inspiration information, which greatly improves the efficiency and scientific nature of R&D decision-making. Attached Figure Description

[0171] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0172] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0173] The following is in conjunction with the appendix Figure 1 The present invention will be described in further detail below.

[0174] The core of this method for quantitative evaluation and innovation analysis of food formulations lies in constructing an analytical framework incorporating domain knowledge to transform formulation information into computable multi-dimensional feature vectors, and then utilizing an improved fusion algorithm for comprehensive evaluation and innovation discovery. This method includes the following steps:

[0175] S1. Obtain target formula data; receive text or structured data of the target formula through a user interface or data interface;

[0176] S2. Multi-dimensional feature extraction: Based on a predefined food ingredient knowledge base, extract at least two dimensions of features from the target formula data. These dimensions include ingredient identity features, ingredient ratio features, and functional combination features; specifically:

[0177] Raw material identification characteristics ( ): Extract the set of all raw materials in the formula, and standardize (e.g., unify synonyms) and classify functions through a knowledge base, and finally represent it as a multi-hot encoded vector or embedding vector.

[0178] Raw material proportion characteristics ( ): Extract the amount or proportion of each raw material, form a numerical vector, and perform normalization (e.g., L2 normalization) to eliminate the influence of the total amount.

[0179] Functional combination features ( Based on the functional roles of ingredients defined in the food ingredient knowledge base, the ingredients in the target formula are aggregated into one or more functional categories they fulfill, forming a function-ingredient mapping relationship, and a functional combination feature vector is generated based on this mapping relationship. For example, all ingredients that fulfill the "sweetness" function are aggregated into a single functional module.

[0180] Further, step S2 extracts functional combination features, specifically including: according to the functional roles of raw materials defined in the food raw material knowledge base, aggregating the raw materials in the target formula to one or more functional categories they implement, forming a function-raw material mapping relationship, and generating a functional combination feature vector based on the mapping relationship.

[0181] Furthermore, the construction scheme of the predefined food ingredient knowledge base in step S2;

[0182] The predefined food ingredient knowledge base (hereinafter referred to as the "knowledge base") is a structured, machine-readable database, which serves as the foundation for multi-dimensional feature extraction and intelligent analysis. The construction scheme of this food ingredient knowledge base is as follows:

[0183] 1) Core Data Structure of the Food Raw Material Knowledge Base: The knowledge base shall contain at least the following core fields to establish a digital profile for each raw material:

[0184] Standard name: A unique identifier for the raw material, such as "ascorbic acid".

[0185] Synonym set: This includes the common name, trade name, and colloquial name of the raw material, used for normalization during text parsing. For example, synonyms for the standard name "ascorbic acid" include "vitamin C" and "VC".

[0186] Chemical and physical properties: including molecular formula, molecular weight, solubility, stability, etc.

[0187] Sensory attributes: such as taste, smell, and color.

[0188] Core: Functional Role Tag Set: This is crucial for extracting combined functional features. Each ingredient is assigned one or more predefined functional role tags. These functional roles are classified based on food science standards, for example:

[0189] Nutritional components: such as protein sources, carbohydrate sources, fat sources, and dietary fiber.

[0190] Technical functional ingredients: such as moisture retainers, leavening agents, emulsifiers, thickeners, gelling agents, stabilizers, sweeteners, and acidity regulators.

[0191] Sensory components: such as colorants, flavoring agents, and umami agents.

[0192] Bioactive ingredients: such as antioxidants and prebiotics.

[0193] 2) Construction Method and Process of Food Raw Material Knowledge Base: The knowledge base is constructed through the following steps:

[0194] Data collection: Raw material information is collected from authoritative food science textbooks, industry standards (such as GB2760 Food Additives Usage Standard), publicly released raw material supplier technical manuals, and labeled patents and academic literature.

[0195] Information structuring: The collected unstructured text information is extracted and populated into the core data structure mentioned above through natural language processing technology combined with manual verification.

[0196] Continuously updated: The knowledge base is designed as a scalable system, allowing administrators to add, delete, and modify the knowledge base based on the emergence of new raw materials and the progress of scientific research.

[0197] 3) Application of the food ingredient knowledge base in feature extraction: In step S2, the system uses this knowledge base to perform the following operations:

[0198] Regarding raw material identification characteristics: the raw material terms identified in the formula text are matched with the "synonym set" in the knowledge base, and finally uniformly mapped to standard names, thereby achieving the standardization of raw materials and ensuring the consistency of analysis.

[0199] For functional combination features: The system retrieves the corresponding functional role tag set from the knowledge base based on the standard name of the raw material. All raw materials in the formula are then aggregated according to their functional roles. For example, all raw materials labeled as sweeteners in the formula (such as white sugar, fructose syrup, and stevia) are identified and categorized into the "sweetening functional module." A functional combination feature vector is generated based on the distribution of all functional modules.

[0200] S3. Calculate dimensional similarity: Calculate the similarity between the features of at least two dimensions of the target formulation and the features of the corresponding dimensions of the reference formulation. For each feature dimension, a similarity metric suitable for its feature representation is used for calculation, taking into account external data such as a formulation database to provide a reference formulation.

[0201] For features in vector form (such as matching features and functional combination features), cosine similarity calculation is preferred:

[0202]

[0203] Where · denotes the dot product operation of vectors. This represents the L2 norm of a vector.

[0204] : No. The cosine similarity score on each feature dimension typically ranges from 1 to 10. When all elements of the feature vector are non-negative, the value range is: ;

[0205] The new formula to be evaluated is in the first stage. Feature vectors in each feature dimension;

[0206] The reference formula is in the [section / section]. Feature vectors in each feature dimension;

[0207] Vector dot product operation, defined as ;

[0208] The L2 norm (Euclidean norm) of the eigenvectors of the new recipe is defined as follows: ;

[0209] The L2 norm of the reference formula eigenvector is defined as follows: .

[0210] For features in the form of sets (such as raw material identity features), Jaccard similarity calculation is preferred:

[0211]

[0212] in, The cardinality of a set;

[0213] : No. Jaccard similarity scores on each feature dimension, with values ​​ranging from 1 to 10. ;

[0214] The new formula to be evaluated is in the [number]th [year]. A set of features across a feature dimension;

[0215] The reference formula is in [section number]. A set of features across a feature dimension;

[0216] The intersection of the feature sets of the new formula and the reference formula;

[0217] The union of the feature sets of the new formula and the reference formula;

[0218] The cardinality of a set is the number of elements in the set.

[0219] S4. Calculate the overall similarity score. Based on the similarity of the at least two dimensions and the weights configured for each dimension, calculate the overall similarity score between the target recipe and the reference recipe. The weights are adaptive, and their values ​​are dynamically adjusted according to the attributes of the target recipe and / or the reference recipe (e.g., recipe type: beverage, baked goods, etc.) or the user-specified analysis focus (e.g., more focus on cost or functionality). Overall Similarity Score Calculated using the following formula:

[0220]

[0221] in, The number of feature dimensions, For the first Adaptive weights for each dimension For the first Similarity across multiple dimensions.

[0222] The "adaptive dynamic weight adjustment" refers to the fact that the weight coefficients Wi are not fixed, but are calculated in real time by a weight decision function f based on the input parameters. Its core mechanism is as follows:

[0223] 1. Input signals for weighted decision-making

[0224] The weighted decision function f receives one or more of the following input signals:

[0225] Signal A: Recipe Type Attribute (T) The system assigns a type label to a recipe by analyzing its ingredient composition or user-specified information, such as T = {Beverages, Baked Goods, Meat Products, Health Foods, Seasonings...}. Different types of recipes have different core competitive dimensions.

[0226] Signal B: User-specified analysis focus (U) The user selects the focus of this analysis from the drop-down menu or checkbox, such as U = {Focus on cost, focus on functionality, focus on cleanliness label, focus on taste...}.

[0227] Signal C: Domain Prior Knowledge (D) This is a predefined rule base that reflects the consensus of food science and specifies the basic importance of each dimension of various formulations when there are no special requirements.

[0228] 2. The core mechanism of weighted decision-making

[0229] The weight decision function f dynamically generates the weight set {W1, W2, W3} based on the above input through one or more of the following mechanisms:

[0230] 1) Table lookup method based on recipe type template

[0231] A weight configuration lookup table is pre-stored within the system.

[0232] Workflow: When the target formula is identified as "beverage", the system automatically loads the corresponding weight set (0.2, 0.5, 0.3).

[0233] When the target recipe is identified as "baked food", the system automatically loads the corresponding weight set (0.3, 0.4, 0.3).

[0234] When the target formula is identified as "health food", the system automatically loads the corresponding weight set (0.3, 0.2, 0.5).

[0235] When the target formula is identified as a "compound seasoning", the system automatically loads the corresponding weight set (0.4, 0.3, 0.3).

[0236] 2) Rule mapping based on user focus

[0237] This mechanism allows users to override default type templates. A set of user focus mapping rules is pre-stored in the system.

[0238] Rule U1: IF "Focus on costs" THEN { }

[0239] Logic: Costs are primarily determined by the raw materials themselves (their identity), therefore their weight is increased; functionality is not a primary consideration for the time being.

[0240] Rule U2: IF "Focus on functionality" THEN { }

[0241] Logic: Focus all efforts on functional innovation of the formula, and maximize the weight of functional dimensions.

[0242] Rule U3: IF "Pay attention to cleaning labels" THEN { }

[0243] Logic: Cleanliness labels focus on "whether it contains unfriendly ingredients," which falls under the category of ingredient identification.

[0244] Workflow: Based on the "Beverage" weight template (0.2, 0.5, 0.3), if the user selects "Focus on Functionality", then rule U2 is executed to adjust the weight to (0.2, 0.4, 0.6).

[0245] 3) Feedback-based fine-tuning based on similarity results

[0246] This is a more advanced dynamic adjustment strategy that gives the system a certain degree of "thinking" ability. After calculating the basic similarity, this mechanism performs a secondary fine-tuning of the weights.

[0247] Rule F1: IF THEN .

[0248] Logic: If two formulas are already highly similar in terms of "raw material identity", then reduce the weight of this dimension, while increasing the weight of "ratio" and "function" to amplify their differences in key details and avoid the conclusion of "high similarity" from obscuring important micro-innovations.

[0249] Rule F2: IF THEN .

[0250] Logic: If two formulas are completely different in "function", this is a very strong signal of differentiation. Therefore, the weight of this dimension should be increased significantly to ensure that the overall score can significantly reflect this fundamental difference.

[0251] 4) Dynamically Adjusted Comprehensive Execution Process

[0252] In step S4, the system performs weight adjustments in the following order:

[0253] Initialization: Initialize the basic weights from the lookup table based on the recipe type (T) and domain prior (D). .

[0254] User coverage: If the user specifies the analysis focus (U), then the analysis is performed according to the user focus mapping rules. After making the first adjustment, we get

[0255] Feedback fine-tuning (optional): After calculating the initial dimensional similarity... Then, based on the feedback-based fine-tuning rules, A second fine-tuning was performed to obtain the final result used for calculation. weight .

[0256] Normalization: Ensure .

[0257] Through the aforementioned multi-stage, multi-signal dynamic weight adjustment mechanism, similarity assessment is no longer a rigid calculation, but an intelligent process that understands the scenario, respects user intent, and focuses on key differences. This greatly improves the accuracy and practicality of the analysis results, making them closer to the flexible judgment of domain experts.

[0258] S5. Generate an innovation analysis report. Based on the differences between the target formulation and the reference formulation, generate an innovation analysis report. Check whether to compare with the next reference formulation. If not, output the report directly; if so, repeat steps S3-S5.

[0259] Furthermore, the generation of the innovation analysis report in step S5 specifically includes:

[0260] S51. Identify at least one difference between the target formulation and the reference formulation. ;

[0261] S52. By querying the knowledge graph, assess the innovation contribution of the aforementioned points of difference and obtain a contribution score. ;

[0262] S53. Based on the contribution score, generate an analysis report that includes interpretations of innovation points and / or R&D insights.

[0263] Furthermore, the knowledge graph construction and query scheme in step 52.

[0264] The knowledge graph, a food innovation knowledge graph stored in a graph structure (hereinafter referred to as the "graph"), is used for deep semantic understanding and value assessment of differences.

[0265] 1) Definition of the graph structure of a knowledge graph: The graph consists of “nodes” and “edges”.

[0266] Node types include:

[0267] Raw materials: Corresponding to the standard names in the knowledge base.

[0268] Function: Corresponds to the functional roles in the knowledge base.

[0269] Technical effects: such as "improving stability", "enhancing satiety", and "improving emulsification".

[0270] Market trends include "clean label", "plant-based", "high protein", and "reduced sugar".

[0271] Potential risks: such as "may cause gastrointestinal discomfort", "poor thermal stability", "regulated limits".

[0272] The relationship types of edges include:

[0273] Function: Connects raw material nodes and functional nodes.

[0274] Achieved effect: Connects raw material or functional nodes with technical effect nodes.

[0275] Align with trends: Connect raw materials, functional or technological effects with market trend nodes.

[0276] Risk exists: Connecting raw material nodes with potentially risky nodes.

[0277] Antagonistic / synergistic effects exist: connect two raw material nodes and describe their interaction.

[0278] 2) Knowledge graph construction method: The knowledge graph is constructed in the following way:

[0279] Structured data import: Import the relationships (such as raw material-function) from the aforementioned "Food Raw Material Knowledge Base" into the graph as the basic data.

[0280] Unstructured data mining: Using text mining and entity relationship extraction techniques, nodes such as technical effects, market trends, and potential risks are automatically or semi-automatically extracted from a large number of food science papers, patent texts, and industry reports, as well as their correlation with raw materials and functions, and added to the graph.

[0281] Expert verification: Domain experts review and correct the automatically extracted relationships to ensure the accuracy of the graph knowledge.

[0282] 3) Knowledge Graph Query and Contribution Evaluation Process: In step S52, the system evaluates the differences through the following process. Innovation contribution score ):

[0283] Difference Analysis: This involves analyzing the identified differences. (e.g., "New ingredient: probiotic powder") is converted into a graph query request.

[0284] Graph traversal query: The system executes a predefined graph traversal algorithm. Taking the differences mentioned above as an example, the query will:

[0285] Locate the raw material node: "probiotic powder".

[0286] Traverse its functional edges to find its function (such as "gut health").

[0287] Traverse the edges of the achieved effects to find the relevant technical effects (such as "regulating gut microbiota").

[0288] Traverse the edges that match the trend to find the relevant market trends (such as "digestive health" and "immune support").

[0289] Check for potential risks and assess potential issues (such as "requires cold chain storage and transportation").

[0290] Contribution Measurement: The system calculates a contribution score (Score) based on the number and weight of associated nodes returned by the query, and a predefined scoring rule base. ).

[0291] Example of a rating rule library:

[0292] IF (related to the "high-heat" market trend) THEN +N+N points.

[0293] IF (achieved a "novel" technical effect) THEN +M+M points.

[0294] IF (High level of risk exists) point.

[0295] Ultimately, Score ( The score is the weighted algebraic sum of all these positive and negative scores. This score collectively reflects the net value of the difference across the technology, market, and risk dimensions.

[0296] Furthermore, in step S52, the sign and magnitude of the contribution score are determined based on predefined rules that associate the differences with their potential impact on the formulation's technical effectiveness, market value, or stability.

[0297] Furthermore, the contribution score "determined based on predefined rules" in step S52 is achieved through a structured and computable "innovation contribution scoring rule base." This rule base encodes food industry expertise, market insights, and regulatory requirements into explicit IF-THEN logic and quantitative indicators, enabling the identification of points of difference. The evaluation process is automated and made objective.

[0298] 1) Composition of the Innovation Contribution Scoring Rule Base

[0299] The rule base consists of multiple scoring dimensions and specific rules under them. Each rule includes triggering conditions and scoring operations.

[0300] Dimension 1: Technological Novelty and Advancement

[0301] Rule T1 (Technological Scarcity):

[0302] If the core ingredients or technology combination involved in the difference point appears less than 5% of the reference formula library for the target product category (such as "yogurt"), then THEN Score = Score + 8.

[0303] If the frequency of IF is between 5% and 20%, then THEN Score = Score + 4.

[0304] If the frequency of IF is higher than 20%, then THEN Score = Score + 0.

[0305] Rule T2 (Technical Effect Intensity):

[0306] If the knowledge graph is linked to a "disruptive" technological effect (such as "achieving heatless processing"), then THEN Score = Score + 10.

[0307] If the IF is associated with a “significant improvement” effect (such as “50% increase in shelf life”), then THEN Score = Score + 6.

[0308] If it relates to a "general" effect (such as "taste adjustment"), then THEN Score = Score + 2.

[0309] Dimension Two: Alignment with Market Trends

[0310] Rule M1 (Trend Popularity):

[0311] If the differences are linked to market trends in the "introduction phase" (such as "precision nutrition") through the knowledge graph, THENScore = Score + 3.

[0312] If it relates to a market trend in a "rapid growth phase" (such as "plant-based"), then THEN Score = Score + 8.

[0313] If the trend is related to a "mature" market trend (such as "low-fat"), then THEN Score = Score + 4.

[0314] Rule M2 (Consumer Group Coverage):

[0315] If the difference point can meet the needs of multiple consumer groups at the same time (such as "suitable for both vegetarians and lactose intolerant people"), then THEN Score = Score + 5.

[0316] Dimension Three: Health and Regulatory Attributes

[0317] Rule H1 (Health Gain):

[0318] The IF difference point achieves "core nutrient optimization" (such as "30% reduction in sugar" or "increased dietary fiber"), THENScore = Score + 7.

[0319] The IF difference point achieves "reduction or elimination of unhealthy components" (such as "zero trans fatty acids"), THEN Score = Score + 5.

[0320] Rule H2 (Regulatory Compliance):

[0321] If the raw materials used at the difference point are subject to regulatory restrictions or use beyond permitted scope in the target market region, (A veto-type punishment).

[0322] Dimension Four: Impact of Process and Cost

[0323] Rule P1 (Process Simplification):

[0324] If the difference point can simplify the production process (such as "eliminating a heating step"), then THEN Score = Score + 4.

[0325] Rule P2 (Cost Impact):

[0326] The IF difference point led to a significant increase in the cost of core raw materials (>15%).

[0327] The IF difference point leads to a significant decrease in the cost of core raw materials (>15%), THEN Score = Score + 3.

[0328] 2) Execution process of scoring rules

[0329] In step S52, the system executes the following process to calculate :

[0330] initialization: =0.

[0331] Feature extraction: from differences Extract features that can be recognized by the rule engine, including:

[0332] Difference types: Addition of raw materials, replacement of raw materials, deletion of raw materials, and changes in proportion.

[0333] Raw materials involved: Specific standard raw material names.

[0334] Range of change: The amount of numerical change in proportion to a change in ratio.

[0335] Rule triggering and execution:

[0336] The rule engine matches the extracted features with the IF part in the rule base.

[0337] All triggered rules, according to their THEN part Add or subtract.

[0338] Example: For differences ="Use stevia instead of white sugar".

[0339] Triggering rule T1: Query the database. "Stevia" appears in the "beverage" category with a frequency of 15%, which is "less common". Therefore, Score = 0 + 4.

[0340] Triggering rule M1: The knowledge graph shows that "stevia" is strongly associated with the "sugar reduction / sugar-free" trend (rapid growth period), so Score=4+8.

[0341] Triggering rule H1: This replacement directly achieves "100% sugar reduction", which belongs to core nutrient optimization, so Score=12+7.

[0342] Triggering rule P2: Cost calculation shows that the cost of stevia is slightly higher, but the difference is less than 15%, so the rule is not triggered.

[0343] Final score: =19.

[0344] Score standardization (optional): To prevent a certain dimension from being scored too high, the final total score can be normalized, for example, by using the Sigmoid function to compress it to the range of -10 to +10.

[0345] 3) Construction and maintenance of the rule base

[0346] The rules in the rule base are derived from the summarization and quantification of food science literature, market analysis reports, consumer survey data, and regulatory documents.

[0347] The rules and their weight coefficients in the rule base are configurable, allowing users to adjust them according to different evaluation focuses (such as prioritizing cost or health).

[0348] Machine learning optimization: In a preferred embodiment, the system can record user feedback on the evaluation results and use this feedback data to dynamically optimize the rule weights through a machine learning model (such as reinforcement learning) so that the scores are more in line with the acceptance of the actual market.

[0349] Through the aforementioned detailed and quantifiable rule base, this invention transforms the "innovation assessment" process, which originally relied on expert subjective experience, into a transparent, repeatable, and computable automated process. This is not only an improvement in efficiency but also a revolution in assessment methodology, ensuring the objectivity and consistency of the analysis results.

[0350] Furthermore, step S5 also includes: calculating the overall innovation index of the target formulation relative to the reference formulation based on the contribution scores of all identified differences.

[0351] Based on the contribution scores of all identified differences, the overall innovation index of the target formulation relative to the reference formulation is calculated. :

[0352]

[0353] in, The total number of differences. For the first The importance coefficient of each difference point A normalization function (such as the Sigmoid function) is used to map the weighted sum to a uniform interpretation range.

[0354] A food formulation analysis system for performing the above-described method includes one or more processors; and a memory for storing computer-executable instructions; when the computer-executable instructions are executed by the processors, the system is controlled to perform the steps of the method described above.

[0355] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0356] Example 1: Evaluation of a novel plant-based protein beverage

[0357] S1. Obtain target formula data: The target formula is: "80% oat base, 8% pea protein, 5% cocoa powder, 4% dietary fiber, and 3% stevia".

[0358] S2. Multi-dimensional feature extraction:

[0359] Raw material identification characteristics Extract the set {oat-based, pea protein, cocoa powder, dietary fiber, stevia}.

[0360] Raw material proportioning characteristics Generate a normalized vector [0.80, 0.08, 0.05, 0.04, 0.03].

[0361] Functional combination features The knowledge base maps raw materials to functions, forming a structure like {Main ingredient: oat-based, Protein supplement: pea protein, Flavor and color: cocoa powder, Functional fiber: dietary fiber, Sweetener: stevia}, and encodes it as a feature vector.

[0362] S3. Calculate dimensional similarity;

[0363] The formula was compared with the most similar reference formula in the database (ingredients: 85% milk, 5% whey protein, 5% cocoa powder, and 5% white sugar).

[0364] Calculate dimensional similarity: (Jaccard, raw material identity) = 0.2, (Cosine, Balance) = 0.7, (Cosine, Function) = 0.6.

[0365] S4. Calculate the overall similarity:

[0366] Assign adaptive weights based on the "beverage" type: = 0.3, = 0.4, = 0.3.

[0367] Overall similarity: 52%.

[0368] S5. Innovation Analysis and Report Generation:

[0369] S51. Identify differences: Use "oat-based, pea protein, stevia, dietary fiber" instead of "milk, whey protein, white sugar".

[0370] S52. Contribution Assessment: Knowledge Graph Judgment It is considered "positive systemic innovation" because it aligns with multiple trends such as "plant-based", "high fiber", and "no added sugar", with a score (δ1) of +15.

[0371] Computational Innovation Index: sigmoid (assumption) .

[0372] S53. Report Generation: The report states: "This formula has a moderate similarity to milk-based formulas (52%), but it has undergone a complete plant-based modification, resulting in an extremely high innovation index (92). The core innovation lies in the complete plant protein and fiber system, and it is recommended to emphasize 'pure plant, high dietary fiber, and zero added sugar' as the core selling points." The report is output directly.

[0373] Through the above steps, the present invention achieves automated and intelligent in-depth analysis of food formulas, effectively solving many problems raised in the background art.

[0374] 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 quantitative evaluation and innovation analysis of food formulations, characterized in that, Includes the following steps: S1. Obtain the target formula data; S2. Multi-dimensional feature extraction: Based on a predefined food ingredient knowledge base, at least two dimensions of features are extracted from the data of the target formula. The dimensions include ingredient identity features, ingredient ratio features, and functional combination features. The ingredient identity features are extracted from the set of all ingredients in the formula and standardized through the knowledge base, such as by unifying synonyms and functional classifications, and finally represented as a multi-hot encoded vector or embedding vector. The raw material ratio characteristics are used to extract the amount or proportion of each raw material, form a numerical vector, and then normalize it to eliminate the influence of the total amount. The functional combination feature, based on the functional roles of raw materials defined in the food raw material knowledge base, aggregates the raw materials in the target formula under one or more functional categories they implement, forming a function-raw material mapping relationship, and generates a functional combination feature vector based on this mapping relationship; S3. Calculate dimensional similarity, calculate the similarity between the features of the target formulation in at least two dimensions and the features of the corresponding dimensions of the reference formulation; S4. Calculate the overall similarity score. Based on the similarity of the at least two dimensions and the weights configured for each dimension, calculate the overall similarity score between the target formula and the reference formula. The weights configured for each dimension in step S4 are adaptive, and their values ​​are dynamically adjusted according to the attributes of the target formula and / or the reference formula, or the analysis focus specified by the user. In S4, weight adjustment is performed in the following order: Initialization: Based on the recipe type T and domain prior D, the basic weights are initialized from the lookup table. , User Coverage: If the user specifies their analytics focus as cost, functionality, or cleaning tags, then the analysis will be performed according to the user focus mapping rules. After making the first adjustment, we get ; Feedback and fine-tuning: After calculating the initial dimensional similarity... Subsequently, according to the feedback-based fine-tuning rules, if the two formulas are already highly similar in terms of raw material identity, the weight of this dimension is reduced, while the weight of the proportions and functions is increased to amplify their differences in key details and avoid the conclusion of high similarity from obscuring important micro-innovations; if the two formulas are completely different in function, the weight of this dimension is significantly increased to ensure that the comprehensive score can significantly reflect this fundamental difference. right A second fine-tuning was performed to obtain the final result used for calculation. weight ; Normalization: Ensure ; S5. Generate an innovation analysis report based on the differences between the target formulation and the reference formulation; Step S5, which generates an innovation analysis report, specifically includes: S51. Identify at least one difference between the target formulation and the reference formulation; S52. By querying the knowledge graph, assess the innovation contribution of the differences and obtain a contribution score; S53. Based on the contribution score, generate an analysis report that includes interpretations of innovation points and / or R&D insights; In step S52, the sign and magnitude of the contribution score are determined based on predefined rules that associate the differences with their potential impact on the formulation's technical effectiveness, market value, or stability. The rules are implemented through a structured, computable base of rules for scoring innovation contributions; The innovation contribution scoring rule base consists of dimensions including technological novelty and advancement, market trend alignment, health and regulatory attributes, and process and cost impact. Technological novelty and advancement are comprised of rules on technological scarcity and the strength of technological effects. The market trend alignment dimension comprises rules on trend popularity and consumer group coverage. The health and regulatory attributes dimension comprises rules on health benefits and regulatory compliance. The process and cost impact dimension comprises rules on process simplification and cost impact. Step S5 further includes: calculating the overall innovation index of the target formulation relative to the reference formulation based on the contribution scores of all identified differences; the overall innovation index : in, The total number of differences. For the first The importance coefficient of each difference point A standardized sigmoid function is used to map the weighted sum to a uniform interpretation range.

2. A food formulation analysis system for performing the method of claim 1, characterized in that: include: One or more processors; A memory for storing computer-executable instructions; when the computer-executable instructions are executed by the processor, controlling the system to perform the steps of the method as described in any of claims 1.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the method as described in claim 1.

Images