Intelligent generation method of industrial product design decision

Abstract

This application relates to an intelligent generation method for industrial product design decisions. It collects multi-source data and generates an evidence chain index for each data record, enabling a unified representation of the multi-source data and facilitating the traceability of original data records. Next, it extracts trend sets and demand sets, and performs multi-dimensional scaling and mapping to fuse them, unifying market trends and user needs into the same low-dimensional space. This achieves quantitative alignment and fusion of the two, allowing design solutions to simultaneously address market hotspots and user pain points. Then, it constructs a constraint set containing hard and soft constraints using preset engineering constraint parameters and design element sets. This brings engineering constraints forward to the solution generation stage, directly eliminating solutions that do not meet hard constraints during the evolutionary algorithm search process, reducing unnecessary rendering computation and lowering decision latency. Finally, consistency checks and formal checks ensure the consistency between the generated rendering and the front-end decision parameters.

Description

Technical Field

[0001] This invention relates to the field of industrial product design technology, and more specifically, to an intelligent generation method for industrial product design decisions. Background Technology

[0002] With changes in consumption patterns, increased demand for personalization, and stronger requirements for sustainable development, product design is gradually shifting from a single functional implementation to a systematic decision-making process that comprehensively considers market trends, user needs, and engineering feasibility constraints. Design decisions must not only focus on user expression of functionality and emotional experience but also simultaneously meet engineering constraints such as material sourcing, process routes, cost boundaries, supply chain availability, and compliance requirements. Furthermore, they must formulate feasible solutions under various constraints. However, the early stages of product design typically require determining the design direction and developing a combination of feasible solutions. In existing technologies, product design decisions are generally completed in the following ways: designers, based on existing experience and market research results, combined with competitor analysis and user feedback, summarize the target audience, usage scenarios, and product selling points, and then propose several design directions and formulate solutions. User needs information is usually obtained through interviews, questionnaires, focus groups, after-sales / customer service records, or evaluation texts; market trend information is usually obtained through industry reports, policy guidance, trend analysis reports, platform popularity, and competitor activities; engineering feasibility is usually verified through material and process reviews, cost accounting, and supply chain assessments.

[0003] To improve the rationality of decision-making, some research institutions and enterprises have introduced methods such as the Analytic Hierarchy Process (AHP) and Quality Function Deployment (QFDP) into the above processes to perform weighted analysis of requirements and map them into design elements or technical indicators. Other solutions use keyword statistics, topic clustering, or traditional machine learning classification to summarize and analyze comments, Q&A, or social media texts to generate requirement keywords, pain point lists, or trend conclusions. In recent years, methods have also introduced deep learning or generative models to perform semantic parsing, summarization, intent recognition, or generate solution descriptions from text to help form design points or design drafts.

[0004] However, existing technologies still have the following technical shortcomings: (1) Multi-source heterogeneous data lacks a unified evidence processing and structured expression mechanism, making it difficult to verify and reuse conclusions. In the existing process, user expressions, trend information and engineering parameters often exist in different forms, lacking a unified data cleaning, standardization and structured representation method; at the same time, the correlation between conclusions and original data is insufficient, making it difficult to trace the evidence basis for trend judgment, demand judgment and solution selection, and making it difficult to reuse and continuously optimize across projects.

[0005] (2) There is a lack of a unified integration and balancing mechanism between market trend direction and user needs. User preferences or discussion intensity are not necessarily equivalent to market direction judgments, and trend signals may not be consistent with current explicit user preferences. Existing methods usually handle trend analysis and demand analysis separately, or rely on human experience for comprehensive judgment. There is a lack of a mechanism to map trend elements and demand elements to a unified representation and perform integrated calculations, resulting in strong subjectivity and insufficient stability in the judgment of design direction.

[0006] (3) Engineering constraints are introduced late, and there is a lack of constraint consistency verification mechanism in the solution generation stage. Constraints such as materials, processes, costs, supply chains and compliance are often centrally verified in the later stage of solution formation. This can easily lead to situations where the solution is valid at the trend or demand level but not feasible at the manufacturing and cost level, resulting in multiple rounds of rework and increasing development risks and trial and error costs.

[0007] (4) The output of the solution lacks an engineered mechanism for generating, evaluating, ranking and delivering multiple solutions. The existing process often relies on a small number of manual solutions, which makes it difficult to automatically generate multiple candidate solutions and conduct comparable evaluation and ranking under the same constraints. There is also a lack of a mechanism to standardize the output and archive of the solution results in the form of visual renderings of solutions / schemes and parameter summaries, which affects the efficiency of cross-team review and subsequent iterations. Summary of the Invention

[0008] The technical problem this invention aims to solve is how to perform evidence processing and structured modeling on large-scale unstructured data from multiple sources, and how to integrate market trends, user needs, and engineering constraints to generate and evaluate multiple candidate solutions for ranking output.

[0009] This invention provides an intelligent generation method for industrial product design decisions, comprising the following steps: S1: Collect multi-source data of the target product and preprocess it to obtain a dataset, and generate an index for each data item in the dataset; S2: Extract trend keywords from the dataset and cluster them into trend themes; count the frequency, coverage, and interaction heat of each trend theme to calculate the trend strength and trend direction; encapsulate the first evidence set formed by the index under each trend theme and its corresponding trend strength and trend direction into a trend element, and each trend element constitutes a trend set; S3: Filter relevant corpora from the dataset, identify the intent categories of the corpora, and extract the attribute items, constraints, and polarities of the corpora; cluster the corpora to obtain the demand items, and encapsulate the second evidence set formed by the index under each demand item and its corresponding intent category, attribute items, constraints, and polarities into a demand element. Each demand element constitutes a demand set. S4: The trend set and the demand set are fused through multi-dimensional scaling mapping. A constraint set is constructed by combining the preset engineering constraints and design elements. Based on the constraint set, an evolutionary algorithm is used to generate candidate solutions, which are then evaluated and ranked. The top N candidate solutions are the target solutions. S5 generates prompt information for each target solution and generates a rendering image. It verifies the consistency between the rendering image and the prompt information. If the verification fails, it modifies the parameters of the prompt information according to the verification failure type, generates a rendering image again for verification, and so on until the verification passes. Finally, it encapsulates and outputs the verified rendering image, target solution, prompt information, and associated evidence set.

[0010] Compared with existing technologies, this application has the following advantages: By collecting multi-source data and generating an index for each data record, this application enables multi-source data to have a unified representation, which is beneficial for tracing the original data records; then, by extracting trend sets and demand sets and performing multi-dimensional scaling mapping and fusion of trend sets and demand sets, market trends and user needs are uniformly mapped to the same low-dimensional space, realizing the quantitative alignment and fusion of the two, so that the design scheme can simultaneously take into account market hotspots and user pain points; next, by using preset engineering constraint parameters and design element sets, a constraint set containing hard constraints and soft constraints is constructed, and engineering constraints are brought forward to the scheme generation stage, directly eliminating schemes that do not meet hard constraints during the evolutionary algorithm search process, reducing the amount of invalid rendering computation and reducing decision delay; finally, through consistency verification and formal verification, the consistency between the generated rendering diagram and the front-end decision parameters is ensured.

[0011] In one possible implementation, in step S1, the multi-source data includes user posts and comments from various platforms, product descriptions, user reviews and Q&A from e-commerce platforms, and internal enterprise data. The preprocessing includes cleaning, deduplication, and format normalization of multi-source data to obtain a standardized dataset; The specific steps involved in generating an index for each data item in the dataset are as follows: For each piece of data in the standardized dataset, a hash algorithm is used to generate a globally unique index by combining the collection timestamp and source number, and the mapping relationship between the index and the source information, collection time, and processing rules of the multi-source data is recorded.

[0012] Compared with existing technologies, by generating a globally unique index for each piece of data in the standardized dataset using a hash algorithm, each piece of data can be accurately traced, solving the problem that conclusions cannot be verified in traditional methods and providing a reliable evidentiary basis for all subsequent analyses.

[0013] In one possible implementation, step S2 specifically includes: S201, extract text content related to the target product from the dataset, and use natural language processing technology to extract trend keywords from the text content; S202 uses a word vector model to map each trend keyword to a keyword vector and calculates the cosine similarity between keyword vectors; Based on a preset similarity threshold, trend keywords that are higher than the similarity threshold are grouped into the same trend theme to form a trend theme set. S203: For each trending topic, its frequency of occurrence, coverage, and interaction intensity are statistically analyzed within a preset time window, and the trend strength is obtained through weighted fusion calculation. : ; In the formula, , , These are the weighting coefficients. Indicates trend themes, Indicates frequency of occurrence. Indicates coverage. Indicates the level of interaction. Indicates within the time window Within the trend theme set, the trend theme with the highest frequency value is... Indicates within the time window Within the trending topic set, the interaction with the highest interaction popularity value is... It is obtained by weighting and summing user interactions such as likes, comments, reposts, and favorites, and then normalizing the sum. S204, based on the trend strength across multiple consecutive time windows, uses linear regression to fit the slope and combines it with preset rules to determine the trend direction; the preset rules include: like ,determination =Upward trend; like ,determination = Downward trend; like ,determination =Stable trend; in, Indicates the slope. For the preset threshold, Indicates the direction of the trend; S205 encapsulates each trend theme, its trend strength, trend direction, and the first set of evidence consisting of the indices under that trend theme into a trend element. : ; in, Represents a set of evidence; S206, all trend elements constitute a trend set.

[0014] Compared with existing technologies, the approach of extracting trend keywords, vectorizing them, clustering them with cosine similarity, calculating trend strength, determining trend direction, and encapsulating trend elements transforms unstructured market information into structured and computable trend elements. Through multi-dimensional quantification (frequency, coverage, popularity) and time series analysis, it objectively identifies market trend directions, solving the problems of strong subjectivity and difficulty in quantification in traditional trend judgment.

[0015] In one possible implementation, S3 specifically includes: S301, a preset demand expression mode, uses a rule matching method to select text fragments related to user demands from a standardized dataset as corpus, forming a demand corpus set, and saves the text fragments and indexes of each corpus; S302, for each piece of corpus, an intent recognition model is used to classify intents and fill slots, resulting in intent tags, attribute items, and constraint expressions in natural language form; the constraint expressions are matched using a preset polarity keyword list to determine polarity, which includes three categories: necessary, expected, and avoidable; using a preset thesaurus and numerical rule base, the numerical descriptions in the constraint expressions are mapped to numerical ranges to form standardized constraint items, and the attribute items, constraint items, and polarities are combined into demand parameter pairs; thus, each piece of corpus is transformed into a set of standardized demand parameter pairs. S303 uses a word vector model to map the text fragments of each corpus into a demand corpus vector, thus obtaining a demand corpus vector set; S304 uses the DBSCAN algorithm to cluster the demand corpus vector set to obtain several clusters. The corpus within each cluster has similar semantics; each cluster corresponds to a demand item. S305, for each cluster: The intent tags of all corpora within a cluster are counted, and the intent tag with the highest frequency of occurrence is selected as the intent category of that cluster; Calculate the Euclidean distance between each demand corpus vector within the cluster and the cluster center, and select the text fragment with the demand corpus vector with the closest Euclidean distance as the representative statement, where the cluster center is the average value of all demand corpus vectors within the cluster; Aggregate the requirement parameters of all corpora within the cluster by attribute item to obtain the requirement parameter set; The polarity of all corpora within a cluster is obtained by merging them according to the rule that must take precedence over expectation and expectation takes precedence over avoidance. The indices of all corpora within the cluster are merged to form the second evidence set for this requirement. S306 encapsulates each requirement item and its intent category, representative statement, attribute item, constraint item, polarity, and second evidence set as a requirement element; all requirement elements constitute a requirement set.

[0016] Compared with existing technologies, the technical solution of corpus screening, intent recognition and attribute constraint extraction, quantification, semantic clustering, intent category, representative statement, demand parameter aggregation, and demand element encapsulation transforms users' natural language expressions into structured and computable demand elements. It merges similar demands through clustering, determines intent categories through voting, and obtains quantitative constraints through aggregation. This solves the problem of fragmented user demands and difficulty in unified modeling, and provides accurate input for subsequent solution generation.

[0017] In one possible implementation, step S305 aggregates the demand parameter pairs of all corpora within the cluster by attribute item to obtain a demand parameter set, specifically including: For numerical attribute items, collect all constraint items associated with the attribute item within the cluster. If the numerical range of the constraint items is consistent, take the largest or smallest constraint item as the aggregation result. If the numerical range of the constraint items is conflicting, mark the conflict. For categorical attribute items, collect all constraint items associated with that attribute item within the cluster, count the frequency of occurrence of the constraint items, and use the constraint item with the highest frequency as the aggregation result; For constraints with different polarities for the same attribute, the required polarity takes precedence over the desired polarity and the avoidable polarity.

[0018] Compared with existing technologies, by setting specific rules for demand parameter aggregation—taking the tightest numerical type, taking the most frequent categorical type, and prioritizing polarity—a mechanism for resolving multi-source demand conflicts is provided. When multiple users make different requirements for the same attribute, they can be automatically integrated into a unified engineering constraint, solving the decision-making problem when demand conflicts arise and ensuring that the generated constraint set is reasonable and feasible.

[0019] In one possible implementation, S4 fuses the trend set and the demand set through a multidimensional scaling mapping, specifically including: A1, vectorizes each trend element in the trend set; For each trend element, which includes a trend theme and multiple associated trend keywords, the trend element vector for each trend element is obtained by weighted averaging based on the keyword vector of the trend keywords. A2, vectorizes each demand element in the demand set; For each demand element, its demand parameters are mapped to each attribute item in the set as word vectors, and each constraint item is encoded as a fixed-length numerical vector. The word vectors and numerical vectors are concatenated and averaged to obtain the demand element vector. ; In the formula, Represents a demand element vector. This represents a set of requirement parameter pairs. , Indicates attribute items, Indicates constraint terms. Represents word vectors; A3 merges all trend element vectors and demand element vectors to form a high-dimensional space point set. ; A4, uses a multidimensional scaling algorithm to map a high-dimensional spatial point set to a low-dimensional spatial point set. Make the distance between points in high-dimensional space Distance between the midpoint of the low-dimensional space To minimize the error, the stress function is: ; In the formula, , representing the Euclidean distance between two points in a point set in a higher-dimensional space; Let represent the Euclidean distance between two points in the low-dimensional point set; by iteratively optimizing and minimizing the stress function, the low-dimensional coordinates of each trend theme are obtained. and the low-dimensional coordinates of each demand element .

[0020] Compared with existing technologies, by vectorizing trend elements, vectorizing demand elements, constructing high-dimensional space, reducing dimensionality through multi-dimensional mapping, and optimizing stress functions, trend elements and demand elements are mapped to the same low-dimensional space, realizing the quantitative alignment and integration of the two. This solves the problem of the separation between trend analysis and demand analysis in traditional methods, enabling the design scheme to take into account both market trends and user needs.

[0021] In one possible implementation, S4 constructs a constraint set by combining preset engineering constraints and design elements, specifically including: B1, a set of design elements for the pre-defined target product. And the value range of each design element, vectorizing each design element to obtain a high-dimensional vector. ; B2. Obtain preset engineering constraint parameters from the engineering side, including cost ceiling, size limit, available materials, available processes, and delivery cycle, and formalize the engineering constraint parameters into a set of engineering-side hard constraints. The engineering constraint parameters include upper limit constraints, lower limit constraints, equality constraints, or category constraints; B3, extract all requirement parameter pairs from the requirement set and classify them according to the polarity of the requirement parameter pairs: If polarity is mandatory, then this set of demand parameters constitutes a set of demand-side hard constraints. ; If polarity is to be avoided, then this requirement parameter pair is a reverse constraint; If the polarity is the desired value, then the demand parameter pair is a set of soft constraints on the demand side. B4, for each design element Integrate its set of engineering-side hard constraints and demand-side hard constraints set If both constraints are upper bound constraints, the smaller of the upper bound constraint values ​​is taken as the hard constraint expression. If both constraints are lower bound constraints, then the larger of the lower bound constraint values ​​is taken as the hard constraint expression. If two constraints contain an upper bound constraint and a lower bound constraint, then the intersection of the two constraints is taken as the hard constraint expression. If there is no intersection, the conflict handling mechanism is triggered; for equality constraints or category constraints, if they are consistent, they are retained, and if they conflict, they are handled according to priority or a conflict is triggered. Each design element obtained after integration Hard constraint expression Together with its source information, they constitute a hard constraint. , Indicates the source information of the constraint; B5, for each design element The optimization direction is determined based on the corresponding set of soft constraints on the demand side. ; B6, Calculate the soft constraint weights for each design element. : ; in, Design elements The trend weights are calculated by plotting the high-dimensional vector of the attribute items. Mapping to a lower-dimensional space yields lower-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes The Euclidean distance is used to determine the trend strength of the trend theme with the closest Euclidean distance as the design element. The trend weight is calculated using the following formula: ; Design elements The priority of demand is determined by statistically analyzing the proportion of clusters containing this attribute out of the total number of clusters. , These are the weighting coefficients; B7 outlines the optimization direction for each design element. With soft constraint weights Combining to form soft constraints ; soft constraints and hard constraints This constitutes a constraint set.

[0022] Compared with existing technologies, this method formalizes engineering hard constraints → divides demand-side hard and soft constraints → integrates hard constraints (taking tighter / intersection) → calculates soft constraint weights (trend weight + demand priority) → forms a set of soft constraints. This method unifies engineering constraints, user requirements, and user expectations into a computable set of constraints. By quantifying the importance of each design element through weight calculation, it solves the problem of the difficulty in uniformly expressing and quantifying multi-source constraints, and provides a clear search boundary and optimization target for evolutionary algorithms.

[0023] In one possible implementation, step S4 involves generating candidate solutions using an evolutionary algorithm based on the constraint set, specifically including: C1, based on the set of design elements The values ​​of each design element are encoded into a decision vector, along with the value range of the design elements. ,in, Indicates the total number of design elements. Representing design elements The values ​​of each decision vector; To be referred to as an individual; C2, randomly generated Each individual is used as the initial population; and each individual is checked to see if it satisfies the hard constraints. : If the conditions are met, the individual is retained; If the conditions are not met, a repair strategy is employed to make the individual meet the hard constraints. ; An initial population consisting of feasible individuals is obtained. ; C3, for the initial population Each individual in Utilizing the global requirement parameter set Trend Theme Collection and its low-dimensional coordinates and trend strength and soft constraints soft constraint weights in Calculate the fitness value for each individual: ; In the formula, Indicates the degree of matching of needs, by individual The values ​​of each design element and the set of global requirement parameters Corresponding constraint terms The satisfaction level is obtained by taking the weighted average of all satisfaction levels; To indicate trend consistency, by representing individual... The vector representation projected onto a low-dimensional space yields low-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes Euclidean distance, and trend strength Weighted average; Indicates the degree of conflict risk, and assesses individuals through preset rules. The risk of internal parameter conflicts is identified. These are the weighting coefficients; C4 iteratively performs evolutionary operations such as selection, crossover, mutation, and population update on the current population until the maximum number of generations is reached or fitness converges, and finally uses all individuals in the population as a set of candidate solutions that satisfy the hard constraints. .

[0024] Compared with existing technologies, the technical solution that adopts individual coding → population initialization → fitness function definition (demand matching degree + trend consistency - conflict risk degree) → evolutionary operations (selection, crossover, mutation) → candidate solution set output automatically searches and generates diverse design solutions that meet hard constraints within the constraint space. By guiding the solutions to optimize towards user needs, market trends and low risk through the fitness function, it solves the problems of low efficiency and single solution in traditional manual design, and significantly improves design efficiency and solution quality.

[0025] In one possible implementation, step S4 evaluates and ranks the candidate solutions, with the top N candidate solutions being the target solutions, specifically including: D1, for the candidate solution set Each candidate solution Calculate candidate solutions Degree of matching with the requirements set The degree of matching of requirements By candidate solutions The values ​​of each design element are compared and calculated with the corresponding constraints in the requirement set. D2, calculate each candidate solution Consistency with the trend set The trend consistency By candidate solutions Projecting onto a lower-dimensional space yields lower-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes The Euclidean distance is obtained by weighting the trend strength; D3, calculate each candidate solution Conflict risk level ; D4, according to Calculate the comprehensive score and sort the candidate solutions in descending order according to the comprehensive score. Perform diversity screening on the sorted candidate solutions, calculate the similarity between the solutions, eliminate the candidate solutions whose similarity with the selected solutions is higher than a preset threshold, and finally output the top N candidate solutions as the target solutions.

[0026] Compared with existing technologies, by objectively and quantitatively evaluating the generated candidate solutions, we can ensure that the final output solution not only meets hard constraints, but also achieves the best overall performance in terms of user needs satisfaction, market trend alignment, and internal conflict risk. At the same time, by using diversity screening, we can avoid outputting similar solutions and provide designers with a variety of high-quality options.

[0027] In one possible implementation, S5 specifically includes: S501, Generate a visual intermediate representation of the current target solution and convert it into current prompt information, the prompt information including design element vector and condition vector; S502: Input the current prompt information into the generator of the generative adversarial network to generate a rendered image of the product; S503, the discriminator of the generative adversarial network is invoked to perform consistency verification and formal verification on the current rendered image and prompt information. The consistency verification includes feature consistency verification and parameter consistency verification. The element consistency check uses a target detection network to identify whether the rendered image contains all the design element vectors specified in the intermediate representation. If so, parameter consistency check is performed; otherwise, the check fails and proceeds to step S504. The parameter consistency check estimates whether the deviation between the geometric parameters of key components in the rendered image and the parameter values ​​in the intermediate representation exceeds a preset threshold by using an image analysis network. If so, the check fails and proceeds to S504; otherwise, a formal check is performed. The formal verification is used to evaluate the generator in the specified category. The output stability under the condition that there exists an arbitrary noise vector ,satisfy This causes the discriminator to misclassify the rendered image into other categories. The probability is greater than the preset threshold If the verification fails, proceed to S504; otherwise, the verification passes and proceed to S505. S504 If the element consistency check fails, enhance the value of the missing design element vector in the intermediate representation, generate a new prompt message, and return to S502. If the parameter consistency check fails, the values ​​of the corresponding design element vectors in the intermediate representation are adjusted using gradient descent based on the deviation, a new prompt message is generated, and S502 is returned. If the formal verification fails, a random perturbation is applied to the condition vector or an adversarial resampling strategy is used to generate a new noise vector. Input the generator, generate a rendered image of the product, and return S503; S505 encapsulates and outputs the verified rendering, target solution, prompt information, and associated evidence set.

[0028] Compared with existing technologies, generative adversarial networks transform structured schemes into visual design drawings and introduce a triple verification mechanism of parameter consistency verification, element consistency verification, and formal verification to ensure that the generated results are completely consistent with the structured schemes, so that the generated design drawings can be directly used for engineering review. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the process of this application. Detailed Implementation

[0030] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0031] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of this application based on the specific circumstances.

[0032] In the embodiments of this application, unless otherwise expressly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0033] The present application will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0034] See Figure 1 As shown in the embodiments of this application, an intelligent generation method for industrial product design decisions is disclosed, comprising the following steps: S1: Collect multi-source data on the target product and preprocess it to obtain a dataset, then generate an index for each data item in the dataset; specifically: This application embodiment collects the following multi-source data through API interfaces and web crawler tools: A total of 15,000 backpack-related posts and comments on social media platforms from December 2025 to February 2026; The e-commerce platform had a total of 10,000 backpack product detail pages and user reviews during the same period; The company's internal after-sales records and customer service dialogue logs totaled 2,000 entries. The collected multi-source data was cleaned, deduplicated, and formatted to obtain a standardized dataset containing 25,000 valid records. The cleaning, deduplication, and formatting processes included removing HTML tags, advertising information, and invalid characters, unifying the text encoding to UTF-8, eliminating duplicate records (based on text similarity MD5 deduplication), and desensitizing sensitive information (anonymizing user nicknames). Next, the SHA-256 hash algorithm is used, combined with the collection timestamp and source number, to generate a globally unique index for each data record in the standardized dataset. For example, a comment from a social media platform, posted on February 15, 2026, "My backpack is too heavy, I hope it will be lighter," is identified by the identifier E_20260215_A_3f2a, where E represents the evidence identifier prefix, 20260215 represents the collection date, A represents the source platform code, and 3f2a is the first four digits of the hash value. All indexes are associated with their corresponding storage paths, collection times, source information, processing rules, and other metadata to form an index set, which is used for the traceability of all subsequent analysis results.

[0035] S2 extracts trend keywords from the dataset and clusters them into trend themes; calculates the frequency, coverage, and interaction intensity of each trend theme to determine trend strength and direction; and encapsulates the first evidence set formed by the indexes under each trend theme, along with its corresponding trend strength and direction, into a trend element. These trend elements constitute a trend set. Specifically, this includes: S201: Extract text content related to the target product from the dataset, and use natural language processing technology to extract a set of trend keywords from the text content. ; In this embodiment, the jieba word segmentation tool is used for word segmentation and part-of-speech tagging, and noun phrases and adjectives are selected as candidate trend keywords to obtain a keyword set. .

[0036] S202 uses a pre-trained word vector model to map each trend keyword to a keyword vector and calculates the cosine similarity between keyword vectors; Based on a preset similarity threshold, trend keywords that are higher than the similarity threshold are grouped into the same trend theme to form a trend theme set. The word vector model in this application uses the BERT model (BERT-based-Chinese) to map each trend keyword to a keyword vector. And calculate the cosine similarity between keyword vectors: ; Based on the preset similarity threshold Embodiments of this application =0.75, trend keywords with a similarity value higher than the similarity threshold are grouped into the same trend theme. To form a collection of trend themes ,in, This represents the total number of trend topics; in this embodiment, trend keywords such as "lightweight," "lightweight," and "comfortable carrying" are categorized as trend topics. Trend keywords such as "waterproof," "sealed," and "waterproof zipper" are categorized as trend themes. Trend keywords such as "quick access," "zoning," and "pocket" are categorized as trend themes. To form a collection of trend themes .

[0037] S203: For each trending topic, its frequency of occurrence, coverage, and interaction intensity are statistically analyzed within a preset time window, and the trend strength is obtained through weighted fusion calculation. : ; In the formula, , , These are the weighting coefficients. Indicates trend themes, Indicates frequency of occurrence. Indicates coverage. Indicates the level of interaction. Indicates within the time window Within the trend theme set, the trend theme with the highest frequency value is... Indicates within the time window Within the trending topic set, the interaction with the highest interaction popularity value is... It is obtained by weighting and summing user interactions such as likes, comments, reposts, and favorites, and then normalizing the sum. The embodiments of this application are as follows: The metrics for each trend theme over a 30-day period (January 15 to February 14, 2026) are shown in Table 1 below: Table 1: Frequency of occurrence within a time window Cross-border coverage and interaction popularity

[0038] ; Pick , , , Then the trend strength of each trend theme is calculated. , , .

[0039] S204, based on the trend strength across multiple consecutive time windows, uses linear regression to fit the slope and combines it with preset rules to determine the trend direction; the preset rules include: like ,determination =Upward trend; like ,determination = Downward trend; like ,determination =Stable trend; in, Indicates the slope. For the preset threshold, Indicates the direction of the trend; The embodiments of this application are based on the trend strength over three consecutive time windows, and the slope is fitted using linear regression. ,Pick =0.05, then: : =0.12 > 0.05, indicating an upward trend; : =0.08 > 0.05, indicating an upward trend; : =0.02∈[-0.05, 0.05], which is determined to be a stationary trend.

[0040] S205 encapsulates each trend theme, its trend strength, trend direction, and the first set of evidence consisting of the indices under that trend theme into a trend element. : ; in, Represents a set of evidence; S206, all trend elements constitute a trend set; The trend elements in this application embodiment include: ; ; .

[0041] S3 involves filtering demand-related corpora from the dataset, identifying the intent categories of the corpora, and extracting the attribute items, constraints, and polarities of the corpora. The corpora are clustered to obtain demand items. For each demand item, a second evidence set consisting of its index and its corresponding intent category, attribute items, constraints, and polarities are encapsulated into a demand element. These demand elements constitute a demand set. Specifically, this includes: S301, a pre-defined demand expression pattern, uses a rule-matching method to select text fragments related to user demands from a standardized dataset as corpus, forming a demand corpus. Save the text fragments and indexes for each corpus; The word set for expressing demand patterns in this application embodiment includes words such as want, hope, don't, too, suggest, and complain.

[0042] S302, for each piece of corpus, an intent recognition model is used for intent classification and slot filling to obtain intent labels, attribute items, and constraint expressions in natural language form; the constraint expressions are matched using a pre-set polarity keyword list to determine polarity. The polarity includes three categories: necessary, expected, and avoidable. Using a pre-defined thesaurus and numerical rule base, the numerical descriptions in the constraint expressions are mapped to numerical ranges, forming standardized constraint terms, and attribute terms are... Constraints Combining polarity into demand parameter pairs This transforms each corpus into a set of standardized requirement parameters. ; In this embodiment, a pre-trained intent recognition model is used to classify the intent and fill the slots for each corpus, resulting in intent tags and attribute items for each corpus. The corresponding constraints in natural language are shown in Table 2 below: Table 2: Processing Results of the Intent Recognition Model for Each Corpus .

[0043] S303 uses a word vector model to map each text segment of the corpus into a vector of the required corpus. The required corpus vector set is obtained. ; S304, using the DBSCAN algorithm to process the required corpus vector set Clustering is performed to obtain several clusters. The corpora within each cluster have similar semantics; each cluster corresponds to a single requirement. S305, for each cluster : The intent tags of all corpora within a cluster are statistically analyzed, and the intent tag with the highest frequency of occurrence is selected as the intent category of that cluster. ; Calculate the Euclidean distance between each demand corpus vector within the cluster and the cluster center, and select the text fragment with the demand corpus vector closest in Euclidean distance as the representative statement. Among them, the cluster center is the demand corpus. The average value of all demand corpus vectors; The requirement parameters of all corpora within the cluster are sorted by attribute item. Clustering is performed to obtain a set of demand parameters. ; The polarity of all corpora within a cluster is obtained by merging them according to the rule that must take precedence over expectation and expectation takes precedence over avoidance. The indices of all corpora within the cluster are merged to form the evidence set for this requirement. ; In this embodiment of the application, the clusters in this embodiment of the application The corpus contains 84 utterances. Statistical analysis shows that the intent tag "comfort" appeared 68 times within the cluster. Therefore, this cluster... Intent category =Comfort, representative statement =The shoulder straps need to be widened, otherwise they'll be too tight; Demand parameters by attribute item Perform aggregation: Weight: Collect ≤900, ≤850, ≤950, etc., take the lower limit ≤850; Shoulder strap width: Collect ≥70, ≥75, etc., take the upper limit ≥75; Chest buckle: Collect must-have, preferably have, etc., polarity "must-have" takes priority, take = have; Aggregation results ={(weight≤850),(shoulder strap width≥75),(chest buckle=present)}; ={(weight≤850),(shoulder strap width≥75),(chest buckle=present)}; By merging the evidence identifiers of all corpora within a cluster, we obtain the evidence set. .

[0044] S306, classify each requirement item and its intent category. Representative statements Attributes Constraints ,polarity Second set of evidence Encapsulate them as requirement elements; all requirement elements constitute a requirement set. : ; All demand elements constitute the demand set. .

[0045] S4 merges the trend set and demand set through multi-dimensional scaling mapping, constructs a constraint set by combining preset engineering constraints and design elements, generates candidate solutions using an evolutionary algorithm based on the constraint set, evaluates and ranks the candidate solutions, and outputs the top N target solutions; where: The trend set and demand set are merged through a multi-dimensional scaling mapping, specifically including: A1, vectorizes each trend element in the trend set; For each trend element, which includes a trend theme and multiple associated trend keywords, the trend element vector for each trend element is obtained by weighted averaging based on the keyword vector of the trend keywords. ; In the formula, Indicating trend elements, Indicates trend themes, Keywords indicating trends Represents a vector of trend elements. Indicates the weighting coefficient; In this application embodiment, trend elements The trend theme is lightweight carrying, and the associated trend keywords are {"lightweight", "reduced burden", "carrying comfort"}. The weighted average of the keyword vectors is taken: .

[0046] A2, vectorizes each demand element in the demand set; For each demand element, its demand parameters are mapped to each attribute item in the set as word vectors, and each constraint item is encoded as a fixed-length numerical vector. The word vectors and numerical vectors are concatenated and averaged to obtain the demand element vector. ; In the formula, Represents a demand element vector. This represents a set of requirement parameter pairs. , Indicates attribute items, Indicates constraint terms. Represents word vectors; In this embodiment of the application, for the demand elements Its requirement parameters for the set : Weight attribute vector Concatenate with constraint code [0.85]; Shoulder strap width attribute vector Concatenate with constraint code [0.875]; Chest buckle attribute vector Concatenate with constraint coding [1]; Take the average to get .

[0047] A3 merges all trend element vectors and demand element vectors to form a high-dimensional space point set. ; A4, uses a multidimensional scaling algorithm to map a high-dimensional spatial point set to a low-dimensional spatial point set. Make the distance between points in high-dimensional space Distance between the midpoint of the low-dimensional space To minimize the error, the stress function is: ; In the formula, , representing the Euclidean distance between two points in a point set in a higher-dimensional space; Let represent the Euclidean distance between two points in the low-dimensional point set; by iteratively optimizing and minimizing the stress function, the low-dimensional coordinates of each trend theme are obtained. and the low-dimensional coordinates of each demand element .

[0048] A constraint set is constructed by combining pre-defined engineering constraints and design elements, specifically including: B1, a set of design elements for the pre-defined target product. And the value range of each design element, vectorizing each design element to obtain a high-dimensional vector. ; In this embodiment, a set of pre-defined design elements is used based on the characteristics of backpack products. The value ranges for each design element are shown in Table 3 below: Table 3: Set of Design Elements and Their Value Ranges .

[0049] B2. Obtain preset engineering constraint parameters from the engineering side, including cost ceiling, size limit, available materials, available processes, and delivery cycle, and formalize the engineering constraint parameters into a set of engineering-side hard constraints. The engineering constraint parameters include upper limit constraints, lower limit constraints, equality constraints, or category constraints; The engineering constraint parameters obtained from the engineering side in this embodiment are specifically as follows: Cost ceiling: cost ≤ 260 yuan; Size restrictions: length ≤ 48cm, width ≤ 32cm, height ≤ 18cm; Available materials: material∈{Nylon composite, TPU coating}; Available processes: process⊆{sewing, partial sealing, heat sealing} Delivery time: lead_time ≤ 30 days.

[0050] B3, extract all requirement parameter pairs from the requirement set and classify them according to the polarity of the requirement parameter pairs: If polarity is mandatory, then this set of demand parameters constitutes a set of demand-side hard constraints. ; If polarity is to be avoided, then this requirement parameter pair is a reverse constraint; If the polarity is the desired value, then the demand parameter pair is a set of soft constraints on the demand side. The embodiments in this application are classified according to polarity: Polarity "must" should be considered as a candidate for hard constraints on the demand side, such as "shoulder strap width ≥ 75" and "chest buckle = present"; Polarity "expectation" can be used as a candidate for soft constraints, such as "weight ≤ 850". Polarity "avoidance" is transformed into reverse constraint, such as "fastener material ∉ plastic".

[0051] B4, for each design element Integrate its set of engineering-side hard constraints and demand-side hard constraints set If both constraints are upper bound constraints, the smaller of the upper bound constraint values ​​is taken as the hard constraint expression. If both constraints are lower bound constraints, then the larger of the lower bound constraint values ​​is taken as the hard constraint expression. If two constraints contain an upper bound constraint and a lower bound constraint, then the intersection of the two constraints is taken as the hard constraint expression. If there is no intersection, the conflict handling mechanism is triggered; for equality constraints or category constraints, if they are consistent, they are retained, and if they conflict, they are handled according to priority or a conflict is triggered. Each design element obtained after integration Hard constraint expression Together with its source information, they constitute a hard constraint. , Indicates the source information of the constraint; The hard constraints integrated in the embodiments of this application ...

[0052] B5, for each design element The optimization direction is determined based on the corresponding set of soft constraints on the demand side. ; B6, Calculate the soft constraint weights for each design element. : ; in, Design elements The trend weights are calculated by plotting the high-dimensional vector of the attribute items. Mapping to a lower-dimensional space yields lower-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes The Euclidean distance is used to determine the trend strength of the trend theme with the closest Euclidean distance as the design element. The trend weight is calculated using the following formula: ; Design elements The priority of demand is determined by statistically analyzing the proportion of clusters containing this attribute out of the total number of clusters. , These are the weighting coefficients; Trend weight in the embodiments of this application Taking weight as an example, its low-dimensional coordinates Calculate the Euclidean distance from the low-dimensional coordinates of the trend theme: and The Euclidean distance is 0.15, and the trend strength is taken as 0.925; and The Euclidean distance is 0.45, and the trend strength is taken as 0.703; and The Euclidean distance is 0.35, and the trend strength is taken as 0.638; Demand Priority Weight appears in 3 demand clusters Total number of clusters: 5, priority: 0.6.

[0053] B7 outlines the optimization direction for each design element. With soft constraint weights Combining to form soft constraints ; The soft constraint weights in this application's embodiments The soft constraints formed .

[0054] soft constraints and hard constraints This constitutes a constraint set.

[0055] Based on the constraint set, an evolutionary algorithm is used to generate candidate solutions, specifically including: C1, based on the set of design elements The values ​​of each design element are encoded into a decision vector, along with the value range of the design elements. ,in, Indicates the total number of design elements. Representing design elements The values ​​of each decision vector; To be referred to as an individual; C2, randomly generated Each individual is used as the initial population; and each individual is checked to see if it satisfies the hard constraints. : If the conditions are met, the individual is retained; If the conditions are not met, a repair strategy is employed to make the individual meet the hard constraints. ; An initial population consisting of feasible individuals is obtained. ; C3, for the initial population Each individual in Utilizing the global requirement parameter set Trend Theme Collection and its low-dimensional coordinates and trend strength and soft constraints soft constraint weights in Calculate the fitness value for each individual: ; In the formula, Indicates the degree of matching of needs, by individual The values ​​of each design element and the set of global requirement parameters Corresponding constraint terms The satisfaction level is obtained by taking the weighted average of all satisfaction levels; To indicate trend consistency, by representing individual... The vector representation projected onto a low-dimensional space yields low-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes Euclidean distance, and trend strength Weighted average; Indicates the degree of conflict risk, and assesses individuals through preset rules. The risk of internal parameter conflicts is identified; These are the weighting coefficients; C4 iteratively performs evolutionary operations such as selection, crossover, mutation, and population update on the current population until the maximum number of generations is reached or fitness converges, and finally uses all individuals in the population as a set of candidate solutions that satisfy the hard constraints. .

[0056] This application employs the NSGA-II algorithm, iterating for 100 generations. Selection uses tournament selection (k=2); crossover uses simulated binary crossover with a probability of 0.9; mutation uses polynomial mutation with a probability of 1 / 6; and population update uses non-dominated sorting combined with crowding distance. After evolution, a set of 200 candidate solutions satisfying the hard constraints is obtained. .

[0057] The candidate solutions are evaluated and ranked, and the top N candidate solutions are the target solutions, specifically including: D1, for the candidate solution set Each candidate solution Calculate candidate solutions Degree of matching with the requirements set The degree of matching of requirements By candidate solutions The values ​​of each design element are compared and calculated with the corresponding constraints in the requirement set. D2, calculate each candidate solution Consistency with the trend set The trend consistency By candidate solutions Projecting onto a lower-dimensional space yields lower-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes The Euclidean distance is obtained by weighting the trend strength; D3, calculate each candidate solution Conflict risk level ; D4, according to Calculate the comprehensive score and sort the candidate solutions in descending order according to the comprehensive score. Perform diversity screening on the sorted candidate solutions, calculate the similarity between the solutions, eliminate the candidate solutions whose similarity with the selected solutions is higher than a preset threshold, and finally output the top N candidate solutions as the target solutions.

[0058] In this embodiment of the application, a comprehensive score is calculated for each solution in the candidate solution set. Sort the results in descending order to obtain the top 3 target solutions, as shown in Table 4 below: Table 4: Target Plan S-01 820 75 3 TPU coating 1 245 0.92 0.88 0.08 0.708 1 S-02 880 70 4 Nylon composite 1 235 0.88 0.85 0.12 0.671 2 S-03 790 80 2 TPU coating 1 258 0.85 0.90 0.15 0.665 3

[0059] S5 generates prompts for each target solution and generates a rendered image. It verifies the consistency between the rendered image and the prompts. If the verification fails, it modifies the parameters of the prompts according to the failure type, generates another rendered image, and performs verification again until the verification passes. Finally, it encapsulates and outputs the verified rendered image, target solution, prompts, and associated evidence set, specifically including: S501, Generate a visual intermediate representation of the current target solution and convert it into current prompt information, the prompt information including design element vector and condition vector; S502: Input the current prompt information into the generator of the generative adversarial network to generate a rendered image of the product; S503, the discriminator of the generative adversarial network is invoked to perform consistency verification and formal verification on the current rendered image and prompt information. The consistency verification includes feature consistency verification and parameter consistency verification. The element consistency check uses a target detection network to identify whether the rendered image contains all the design element vectors specified in the intermediate representation. If so, parameter consistency check is performed; otherwise, the check fails and proceeds to step S504. The parameter consistency check estimates whether the deviation between the geometric parameters of key components in the rendered image and the parameter values ​​in the intermediate representation exceeds a preset threshold by using an image analysis network. If so, the check fails and proceeds to S504; otherwise, a formal check is performed. The formal verification is used to evaluate the generator in the specified category. The output stability under the condition that there exists an arbitrary noise vector ,satisfy This causes the discriminator to misclassify the rendered image into other categories. The probability is greater than the preset threshold If the verification fails, proceed to S504; otherwise, the verification passes and proceed to S505. S504 If the element consistency check fails, enhance the value of the missing design element vector in the intermediate representation, generate a new prompt message, and return to S502. If the parameter consistency check fails, the values ​​of the corresponding design element vectors in the intermediate representation are adjusted using gradient descent based on the deviation, a new prompt message is generated, and S502 is returned. If the formal verification fails, a random perturbation is applied to the condition vector or an adversarial resampling strategy is used to generate a new noise vector. Input the generator, generate a rendered image of the product, and return S503; S505 encapsulates and outputs the verified rendering, target solution, prompt information, and associated evidence set.

[0060] The visual intermediate representation of JSON format generated by the target scheme S-01 in this application embodiment; The intermediate representation is converted into a text prompt, which is then input into the pre-trained StyleGAN3 generator to generate a 512×512 pixel knapsack rendering.

[0061] Verification is performed using a ResNet-50-based discriminator: Element consistency: The target detection and recognition image contains elements such as a chest buckle and three pockets, passing the test. Parameter consistency: Image analysis estimated the shoulder strap width ratio, which matches 75mm, passing the test. Formal verification: Verify the noise vector for any input. The adversarial network generator will not misclassify the knapsack as another category. Solving the optimization problem. The maximum probability of violation is 0.02 < 0.05, so it passes.

[0062] In the description of the embodiments of this application, it should be noted that the terms "inner" and "outer" and other terms indicating direction or positional relationship are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and does not indicate or imply that the device or component must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this application.

[0063] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligently generating industrial product design decisions, characterized in that, Including the following steps: S1: Collect multi-source data of the target product and preprocess it to obtain a dataset, and generate an index for each data item in the dataset; S2: Extract trend keywords from the dataset and cluster them into trend themes; count the frequency, coverage, and interaction heat of each trend theme to calculate the trend strength and trend direction; encapsulate the first evidence set formed by the index under each trend theme and its corresponding trend strength and trend direction into a trend element, and each trend element constitutes a trend set; S3: Filter relevant corpora from the dataset, identify the intent categories of the corpora, and extract the attribute items, constraints, and polarities of the corpora; Clustering the corpus yields demand items. The second evidence set formed by the index under each demand item, along with its corresponding intent category, attribute item, constraint item, and polarity, is encapsulated into a demand element. Each demand element constitutes a demand set. S4: The trend set and the demand set are fused through multi-dimensional scaling mapping. A constraint set is constructed by combining the preset engineering constraints and design elements. Based on the constraint set, an evolutionary algorithm is used to generate candidate solutions, which are then evaluated and ranked. The top N candidate solutions are the target solutions. S5 generates prompt information for each target solution and generates a rendering image. It verifies the consistency between the rendering image and the prompt information. If the verification fails, it modifies the parameters of the prompt information according to the verification failure type, generates a rendering image again for verification, and so on until the verification passes. Finally, it encapsulates and outputs the verified rendering image, target solution, prompt information, and associated evidence set.

2. The intelligent generation method for industrial product design decisions according to claim 1, characterized in that, In step S1, the multi-source data includes user posts and comments from various platforms, product descriptions, user reviews and Q&A from e-commerce platforms, and internal enterprise data. The preprocessing includes cleaning, deduplication, and format normalization of multi-source data to obtain a standardized dataset; The specific steps involved in generating an index for each data item in the dataset are as follows: For each piece of data in the standardized dataset, a hash algorithm is used to generate a globally unique index by combining the collection timestamp and source number, and the mapping relationship between the index and the source information, collection time, and processing rules of the multi-source data is recorded.

3. The intelligent generation method for industrial product design decisions according to claim 1, characterized in that, Step S2 specifically includes: S201, extract text content related to the target product from the dataset, and use natural language processing technology to extract trend keywords from the text content; S202 uses a word vector model to map each trend keyword to a keyword vector and calculates the cosine similarity between keyword vectors; Based on a preset similarity threshold, trend keywords that are higher than the similarity threshold are grouped into the same trend theme to form a trend theme set. S203: For each trending topic, its frequency of occurrence, coverage, and interaction intensity are statistically analyzed within a preset time window, and the trend strength is obtained through weighted fusion calculation. : ； In the formula, , , These are the weighting coefficients. Indicates trend themes, Indicates frequency of occurrence. Indicates coverage. Indicates the level of interaction. Indicates within the time window Within the trend theme set, the trend theme with the highest frequency value is... Indicates within the time window Within the trending topic set, the interaction with the highest interaction popularity value is... It is obtained by weighting and summing user interactions such as likes, comments, reposts, and favorites, and then normalizing the sum. S204, based on the trend strength across multiple consecutive time windows, uses linear regression to fit the slope and combines it with preset rules to determine the trend direction; the preset rules include: like ,determination =Upward trend; like ,determination = Downward trend; like ,determination =Stable trend; in, Indicates the slope. For the preset threshold, Indicates the direction of the trend; S205 encapsulates each trend theme, its trend strength, trend direction, and the first set of evidence consisting of the indices under that trend theme into a trend element. : ； in, Represents a set of evidence; S206, all trend elements constitute a trend set.

4. The intelligent generation method for industrial product design decisions according to claim 3, characterized in that, S3 specifically includes: S301, a preset demand expression mode, uses a rule matching method to select text fragments related to user demands from a standardized dataset as corpus, forming a demand corpus set, and saves the text fragments and indexes of each corpus; S302, for each piece of corpus, an intent recognition model is used to classify intents and fill slots, resulting in intent tags, attribute items, and constraint expressions in natural language form; the constraint expressions are matched using a preset polarity keyword list to determine polarity, which includes three categories: necessary, expected, and avoidable; using a preset thesaurus and numerical rule base, the numerical descriptions in the constraint expressions are mapped to numerical ranges to form standardized constraint items, and the attribute items, constraint items, and polarities are combined into demand parameter pairs; thus, each piece of corpus is transformed into a set of standardized demand parameter pairs. S303 uses a word vector model to map the text fragments of each corpus into a demand corpus vector, thus obtaining a demand corpus vector set; S304 uses the DBSCAN algorithm to cluster the demand corpus vector set to obtain several clusters. The corpus within each cluster has similar semantics; each cluster corresponds to a demand item. S305, for each cluster: The intent tags of all corpora within a cluster are counted, and the intent tag with the highest frequency of occurrence is selected as the intent category of that cluster; Calculate the Euclidean distance between each demand corpus vector within the cluster and the cluster center, and select the text fragment with the demand corpus vector with the closest Euclidean distance as the representative statement, where the cluster center is the average value of all demand corpus vectors within the cluster; Aggregate the requirement parameters of all corpora within the cluster by attribute item to obtain the requirement parameter set; The polarity of all corpora within a cluster is obtained by merging the polarities of all corpora according to the rule that must take precedence over expectation and expectation takes precedence over avoidance. The indices of all corpora within the cluster are merged to form the second evidence set for this requirement. S306 encapsulates each requirement item and its intent category, representative statement, attribute item, constraint item, polarity, and second evidence set as a requirement element; all requirement elements constitute a requirement set.

5. The intelligent generation method for industrial product design decisions according to claim 4, characterized in that, In S305, the requirement parameters of all corpora within a cluster are aggregated by attribute item to obtain a set of requirement parameters, specifically including: For numerical attribute items, collect all constraint items associated with the attribute item within the cluster. If the numerical range of the constraint items is consistent, take the largest or smallest constraint item as the aggregation result. If the numerical range of the constraint items is conflicting, mark the conflict. For categorical attribute items, collect all constraint items associated with that attribute item within the cluster, count the frequency of occurrence of the constraint items, and use the constraint item with the highest frequency as the aggregation result; For constraints with different polarities for the same attribute, the required polarity takes precedence over the desired polarity and the avoidable polarity.

6. The intelligent generation method for industrial product design decisions according to claim 4, characterized in that, S4 merges the trend set and demand set through a multidimensional scaling mapping, specifically including: A1, vectorizes each trend element in the trend set; For each trend element, which includes a trend theme and multiple associated trend keywords, the trend element vector for each trend element is obtained by weighted averaging based on the keyword vector of the trend keywords. A2, vectorizes each demand element in the demand set; For each demand element, its demand parameters are mapped to each attribute item in the set as word vectors, and each constraint item is encoded as a fixed-length numerical vector. The word vectors and numerical vectors are concatenated and averaged to obtain the demand element vector. ； In the formula, Represents a demand element vector. This represents a set of requirement parameter pairs. , Indicates attribute items, Indicates constraint terms. Represents word vectors; A3 merges all trend element vectors and demand element vectors to form a high-dimensional space point set. ; A4, uses a multidimensional scaling algorithm to map a high-dimensional spatial point set to a low-dimensional spatial point set. Make the distance between points in high-dimensional space Distance between the midpoint of the low-dimensional space To minimize the error, the stress function is: ； In the formula, , representing the Euclidean distance between two points in a point set in a higher-dimensional space; Let represent the Euclidean distance between two points in the low-dimensional point set; by iteratively optimizing and minimizing the stress function, the low-dimensional coordinates of each trend theme are obtained. and the low-dimensional coordinates of each demand element .

7. The intelligent generation method for industrial product design decisions according to claim 4, characterized in that, S4 constructs a constraint set by combining preset engineering constraints and design elements, specifically including: B1, a set of design elements for the target product. And the value range of each design element, vectorizing each design element to obtain a high-dimensional vector. ; B2. Obtain preset engineering constraint parameters from the engineering side, including cost ceiling, size limit, available materials, available processes, and delivery cycle, and formalize the engineering constraint parameters into a set of engineering-side hard constraints. The engineering constraint parameters include upper limit constraints, lower limit constraints, equality constraints, or category constraints; B3, extract all requirement parameter pairs from the requirement set and classify them according to the polarity of the requirement parameter pairs: If polarity is mandatory, then this set of demand parameters constitutes a set of demand-side hard constraints. ; If polarity is to be avoided, then this requirement parameter pair is a reverse constraint; If the polarity is the desired value, then the demand parameter pair is a set of soft constraints on the demand side. B4, for each design element Integrate its set of engineering-side hard constraints and demand-side hard constraints set If both constraints are upper bound constraints, the smaller of the upper bound constraint values ​​is taken as the hard constraint expression. If both constraints are lower bound constraints, then the larger of the lower bound constraint values ​​is taken as the hard constraint expression. If two constraints contain an upper bound constraint and a lower bound constraint, then the intersection of the two constraints is taken as the hard constraint expression. If there is no intersection, the conflict handling mechanism is triggered; for equality constraints or category constraints, if they are consistent, they are retained, and if they conflict, they are handled according to priority or a conflict is triggered. Each design element obtained after integration Hard constraint expression Together with its source information, they constitute a hard constraint. , Indicates the source information of the constraint; B5, for each design element The optimization direction is determined based on the corresponding set of soft constraints on the demand side. ; B6, Calculate the soft constraint weights for each design element. : ； in, Design elements The trend weights are calculated by plotting the high-dimensional vector of the attribute items. Mapping to a lower-dimensional space yields lower-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes The Euclidean distance is used to determine the trend strength of the trend theme with the closest Euclidean distance as the design element. The trend weight is calculated using the following formula: ; Design elements The priority of demand is determined by statistically analyzing the proportion of clusters containing this attribute out of the total number of clusters. , These are the weighting coefficients; B7 outlines the optimization direction for each design element. With soft constraint weights Combining to form soft constraints ; soft constraints and hard constraints This constitutes a constraint set.

8. The intelligent generation method for industrial product design decisions according to claim 7, characterized in that, In step S4, an evolutionary algorithm is used to generate candidate solutions based on the constraint set, specifically including: C1, based on the set of design elements The values ​​of each design element are encoded into a decision vector, along with the value range of the design elements. ,in, Indicates the total number of design elements. Representing design elements The values ​​of each decision vector; To be referred to as an individual; C2, randomly generated Each individual is used as the initial population; and each individual is checked to see if it satisfies the hard constraints. : If the conditions are met, the individual is retained; If the conditions are not met, a repair strategy is employed to make the individual meet the hard constraints. ; An initial population consisting of feasible individuals is obtained. ; C3, for the initial population Each individual in Utilizing the global requirement parameter set Trend Theme Collection and its low-dimensional coordinates and trend strength and soft constraints soft constraint weights in Calculate the fitness value for each individual: ； In the formula, Indicates the degree of matching of needs, by individual The values ​​of each design element and the set of global requirement parameters Corresponding constraint terms The satisfaction level is obtained by taking the weighted average of all satisfaction levels; To indicate trend consistency, by representing individual... The vector representation projected onto a low-dimensional space yields low-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes Euclidean distance, and trend strength Weighted average; Indicates the degree of conflict risk, and assesses individuals through preset rules. The risk of internal parameter conflicts is identified; These are the weighting coefficients; C4 iteratively performs evolutionary operations such as selection, crossover, mutation, and population update on the current population until the maximum number of generations is reached or fitness converges, and finally uses all individuals in the population as a set of candidate solutions that satisfy the hard constraints. .

9. The intelligent generation method for industrial product design decisions according to claim 8, characterized in that, Step S4 evaluates and ranks the candidate solutions, with the top N candidate solutions being the target solutions. Specifically, this includes: D1, for the candidate solution set Each candidate solution Calculate candidate solutions Degree of matching with the requirements set The degree of matching of requirements By candidate solutions The values ​​of each design element are compared and calculated with the corresponding constraints in the requirement set. D2, calculate each candidate solution Consistency with the trend set The trend consistency By candidate solutions Projecting onto a lower-dimensional space yields lower-dimensional coordinates. Then calculate the low-dimensional coordinates. Low-dimensional coordinates of various trend themes The Euclidean distance is obtained by weighting the trend strength; D3, calculate each candidate solution Conflict risk level ; D4, according to Calculate the comprehensive score and sort the candidate solutions in descending order according to the comprehensive score. Perform diversity screening on the sorted candidate solutions, calculate the similarity between the solutions, eliminate the candidate solutions whose similarity with the selected solutions is higher than a preset threshold, and finally output the top N candidate solutions as the target solutions.

10. The intelligent generation method for industrial product design decisions according to claim 9, characterized in that, S5 specifically includes: S501, Generate a visual intermediate representation of the current target solution and convert it into current prompt information, the prompt information including design element vector and condition vector; S502: Input the current prompt information into the generator of the generative adversarial network to generate a rendered image of the product; S503, the discriminator of the generative adversarial network is invoked to perform consistency verification and formal verification on the current rendered image and prompt information. The consistency verification includes feature consistency verification and parameter consistency verification. The element consistency check uses a target detection network to identify whether the rendered image contains all the design element vectors specified in the intermediate representation. If so, parameter consistency check is performed; otherwise, the check fails and proceeds to step S504. The parameter consistency check estimates whether the deviation between the geometric parameters of key components in the rendered image and the parameter values ​​in the intermediate representation exceeds a preset threshold by using an image analysis network. If so, the check fails and proceeds to S504; otherwise, a formal check is performed. The formal verification is used to evaluate the generator in the specified category. Output stability under the condition that there exists an arbitrary noise vector ,satisfy This causes the discriminator to misclassify the rendered image into other categories. The probability is greater than the preset threshold If the verification fails, proceed to S504; otherwise, if the verification passes, proceed to S505. S504 If the element consistency check fails, enhance the value of the missing design element vector in the intermediate representation, generate a new prompt message, and return to S502. If the parameter consistency check fails, the values ​​of the corresponding design element vectors in the intermediate representation are adjusted using gradient descent based on the deviation, a new prompt message is generated, and S502 is returned. If the formal verification fails, a random perturbation is applied to the condition vector or an adversarial resampling strategy is used to generate a new noise vector. Input the generator, generate a rendered image of the product, and return S503; S505 encapsulates and outputs the verified rendering, target solution, prompt information, and associated evidence set.

Images