Knowledge graph normalization mapping method and system based on twelve-dimensional coupled signals

Abstract

The application discloses a knowledge graph normalization mapping method and system based on twelve-dimensional coupling signals, relates to the technical field of multi-source heterogeneous data processing, and specifically comprises the following steps: step S1, defining a twelve-dimensional data source set for constructing an industrial knowledge graph, wherein the twelve-dimensional data source set at least comprises a commercial behavior dynamic dimension; and step S2, performing normalization transformation of alignment with a field benchmark on original values of each-dimensional data in the twelve-dimensional data source set, and converting original values of different dimensions to a preset numerical interval. According to the application, the normalization of the twelve-dimensional coupling signals makes the entity attribute values in the knowledge graph more comparable and more insightful in business, and avoids one-sidedness of 'patent-only theory'; and through a default propagation mechanism, the relative stability of the graph structure can be maintained and uncertainty can be quantified even in the case that data is incomplete.

Description

Technical Field

[0001] This invention relates to the field of multi-source heterogeneous data processing technology, specifically to a knowledge graph normalization mapping method and system based on twelve-dimensional coupled signals. Background Technology

[0002] In technology strategy analysis of hard technology industries (such as semiconductors, new energy, and commercial aerospace), it is typically necessary to construct a knowledge graph containing multi-source data including patents, markets, supply chains, and policies for technology positioning and risk warning. However, existing technologies face the following core problems when integrating data: Single data dimension: Traditional knowledge graph ETL (Extract, Transform, Load) processes mostly only perform simple field mapping and do not normalize according to specific technical coupling signals (such as multi-dimensional signals such as technology patents, business behavior, cross-border operations, etc.), resulting in the graph failing to reflect the true market position and risk level of the technical entity.

[0003] Inconsistent dimensions: Data from different sources (such as the number of patents, the amount of financing, consumer NPS scores, and supply chain dependence) have vastly different dimensions and lack a unified dynamic normalization transformation mechanism, making it impossible to fuse and calculate data on the same scale.

[0004] Crude handling of missing data: When data in a certain dimension is missing or of low quality, existing technologies usually adopt a simple zeroing or discarding approach, ignoring the weight contribution of that dimension's data in the overall evaluation, resulting in distorted analysis results.

[0005] The data is disconnected from downstream analysis systems: the data mapping results lack effective connection with subsequent dynamic knowledge graph risk inference (such as supply interruption warning) and algorithm circuit breaker mechanisms (such as data quality triggering algorithm switching), forming data silos.

[0006] Existing knowledge graph ETL processes often involve simple field mapping or single patent data ingestion, failing to consider: fusion distortion caused by differences in data dimensions; the absence of dynamic dimensions related to business behavior during graph initialization; bias introduced by simply setting data to zero when data is missing; and the lack of standardized interfaces for downstream risk projection (calculation of difficulty weights for technology substitution, propagation of supply disruptions). To address these issues, we propose a knowledge graph normalization mapping method and system based on twelve-dimensional coupled signals.

[0007] This application, along with the applicant's other patents filed on the same day or successively, namely "Industrial Technology Supply Disruption Risk Early Warning Method and System Based on Multi-Source Signal Coupling and Dynamic Verification" and "Dynamic Knowledge Graph Circuit Breaking Optimization Method and System Based on Multi-Modal Data Fusion", constitute a patent family. This application provides a standardized mapping and dynamic normalization processing method for multi-source heterogeneous data to knowledge graph node attributes in this family, providing a reliable and unified data foundation for subsequent risk inference and algorithm circuit breaking. Summary of the Invention

[0008] To address the aforementioned technical problems, this technical solution provides a knowledge graph normalization mapping method and system based on twelve-dimensional coupled signals, thus resolving the aforementioned issues.

[0009] To achieve the above objectives, the technical solution adopted in this invention is: a knowledge graph normalization mapping method based on twelve-dimensional coupled signals, the specific steps of which are as follows: Step S1: Define a 12-dimensional data source set for constructing an industry knowledge graph, wherein the 12-dimensional data source set contains at least a dynamic dimension of business behavior; Step S2: Perform a normalization transformation on the original values ​​of each dimension of the data in the twelve-dimensional data source set, aligned with the domain benchmark, to convert the original values ​​of different dimensions to a preset numerical range. Step S3: Based on the pre-constructed anchor matrix, perform cross-language semantic alignment on the keywords of technical entities, and map the normalized data into node attributes and edge weights of the knowledge graph. Step S4: For any dimension of data in the twelve-dimensional data source set, when the data of that dimension is missing or the quality score of the data of that dimension is lower than the preset quality threshold, the data is propagated to the other effective dimensions by default according to the original weight ratio of that dimension in the preset weight vector, and the missing impact coefficient is recorded. Step S5: Output the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module as standardized inputs for graph construction and analysis.

[0010] Preferably, the twelve-dimensional data source set in step S1 includes: technology patent dimension, market dynamics dimension, consumer behavior dimension, competitive ecosystem dimension, supply chain collaboration dimension, enterprise operation dimension, capital and finance dimension, talent innovation dimension, business model paradigm dimension, cross-border operation dimension, policy resources dimension, and business behavior dynamics dimension; wherein the business behavior dynamics dimension includes at least one of public attendance record code, interview publication record code, non-public contact mark, and major contract announcement trigger mark.

[0011] Preferably, the normalization transformation in step S2 includes at least one of the following strategies: applying minimum-maximum normalization to continuous positive original values ​​and using the industry benchmark value as the upper bound of normalization; performing linear shift on the rating type original values ​​according to a preset interval; mapping Boolean original values ​​to a binary set; and dividing the level-coded original values ​​by the corresponding maximum level number to obtain the normalized value.

[0012] Preferably, the anchor matrix in step S3 is pre-constructed through the following steps: collecting multilingual terms and their definition paragraphs within the domain to form a structured terminology library; inputting the contextual descriptions of each term in the structured terminology library into a cross-lingual pre-trained model for encoding to obtain high-dimensional semantic vectors for each term; taking the average of the semantic vectors of the same concept in different languages ​​as the anchor vector of that concept; organizing the anchor vectors of each concept by rows to form the anchor matrix, and establishing an approximate nearest neighbor retrieval index for the anchor matrix.

[0013] Preferably, the anchor matrix supports incremental updates and version management, specifically including: when a new term is detected to appear more frequently than a preset frequency threshold within a unit time window or when a domain standard release event is detected, the anchor vector of the new concept is incrementally added to the anchor matrix; for each updated anchor matrix, a version number and timestamp are appended and historical versions are retained to support rollback.

[0014] Preferably, step S3, which involves performing cross-lingual semantic alignment on technical entity keywords based on a pre-constructed anchor matrix and mapping the normalized data to node attributes and edge weights of a knowledge graph, specifically includes: calculating a query vector using a cross-lingual pre-trained model for the technical entity keywords to be mapped; retrieving the anchor vector in the anchor matrix that has the highest cosine similarity to the query vector; when the cosine similarity is greater than a first similarity threshold, determining the technical entity keyword as a corresponding concept and assigning it a normalized attribute value; and when the cosine similarity is lower than a second similarity threshold, marking the technical entity keyword as a new term awaiting manual review.

[0015] Preferably, the step S4 of propagating the default weight of the dimension to the other valid dimensions according to the original weight ratio in the preset weight vector specifically includes: forming a set of valid dimensions with quality scores not lower than the preset quality threshold and without missing dimensions; for any valid dimension in the set of valid dimensions, setting its updated weight to the sum of the original weight of the valid dimension and the allocation increment, wherein the allocation increment is equal to the original weight of the missing dimension multiplied by the proportion of the original weight of the valid dimension to the sum of the original weights of the valid dimension set.

[0016] Preferably, the missing dimension impact coefficient is positively correlated with the original weight of the missing dimension and negatively correlated with the proportion of the number of valid dimensions in the twelve-dimensional data source set to the total number of dimensions, so that the more missing dimensions there are or the greater the original weight of the missing dimension, the greater the missing dimension impact coefficient.

[0017] Preferably, after recording the missing impact coefficient, the normalized values ​​of each effective dimension are summed according to the updated weights, and then divided by 1 and the sum of the missing impact coefficients to obtain the comprehensive confidence score. The comprehensive confidence score is then output as the confidence attribute of the corresponding knowledge graph node along with the node attributes. Step S5, which involves outputting the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module, specifically includes: outputting the node attributes, edge weights, missing impact coefficients, and comprehensive confidence scores to the supply disruption risk simulation module through a standardized data interface, as input for calculating the difficulty weight of technology substitution; and outputting the node attributes, edge weights, and quality scores to the circuit breaker optimization module through a standardized data interface, as input for determining whether data quality triggers algorithm switching.

[0018] A knowledge graph normalization mapping system based on twelve-dimensional coupled signals includes: a data source definition module, which is configured to define and manage a set of twelve-dimensional data sources; The normalization transformation module is configured to perform a normalization transformation aligned with the domain benchmark on the original values ​​of each dimension of the data in the twelve-dimensional data source set. The anchor matrix management module is configured to build, update, and retrieve anchor matrices for cross-language semantic alignment. The mapping execution module is configured to map the normalized data into node attributes and edge weights of the knowledge graph based on the anchor matrix. The default propagation module is configured to distribute weights to the remaining valid dimensions according to the original weight ratio in the preset weight vector and record the missing impact coefficient when any dimension of data is missing or its quality score is lower than the preset quality threshold. The interface connection module is configured to output the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention normalizes twelve-dimensional coupled signals, making entity attribute values ​​in the knowledge graph more comparable and providing greater business insight, thus avoiding the one-sidedness of "patent supremacy." Through a default propagation mechanism, it maintains the relative stability of the graph structure even with incomplete data and quantifies uncertainty. It establishes a standardized pipeline from raw data to the knowledge graph and then to risk inference, enabling the reuse and linkage of data assets. The anchor matrix disambiguation resolves the ambiguity of "BMS = battery management or bus manager," ensuring the accuracy of the graph. Attached Figure Description

[0020] Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0021] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0022] In some embodiments of this application, a knowledge graph normalization mapping method based on twelve-dimensional coupling signals is provided. The method includes the following steps: Step S1, defining a twelve-dimensional data source set for constructing an industry knowledge graph, wherein the twelve-dimensional data source set includes at least a dynamic dimension of business behavior; Step S2, performing a normalization transformation aligned with a domain benchmark on the original values ​​of each dimension of the data in the twelve-dimensional data source set, converting the original values ​​of different dimensions to a preset numerical range; Step S3, performing cross-language semantic alignment on technical entity keywords based on a pre-constructed anchor matrix, and mapping the normalized data to node attributes and edge weights of the knowledge graph; Step S4, for any dimension of data in the twelve-dimensional data source set, when the data of that dimension is missing or the quality score of the data of that dimension is lower than a preset quality threshold, default propagation is performed to the other effective dimensions according to the original weight ratio of that dimension in the preset weight vector, and the missing impact coefficient is recorded; Step S5, outputting the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module as standardized input for graph construction and analysis.

[0023] In practical implementation, the twelve-dimensional data source set can be formally expressed as follows: Each dimension of the data source physically corresponds to an independent data access channel, and its original value It can be a real number, integer, Boolean value, or ordered rank encoding, with different dimensions and numerical ranges. The preset numerical range is usually a closed interval [0, 1], but can also be configured as a standardized Z-score interval according to specific business needs. The anchor matrix can be obtained by encoding multilingual terms in the domain using a cross-language pre-trained model (e.g., XLM-RoBERTa-large, LaBSE, mBERT model) and taking the mean value. Each row in the matrix corresponds to an anchor vector of a concept, and the vector dimension is usually one of 512, 768, or 1024 dimensions. The anchor matrix can be indexed using FAISSindexFlatIP or IndexIVFFlat during storage to support millisecond-level nearest neighbor retrieval. The preset quality threshold... It is a real number between 0 and 1, such as 0.3, 0.4, or 0.5, and can be configured according to the sensitivity of the downstream analysis module to data integrity. The missing data impact coefficient... It is a scalar that reflects the uncertainty of the current assessment results. The larger the value, the greater the loss of confidence due to missing data. The downstream analysis modules include, but are not limited to, a supply disruption risk simulation module, a dynamic knowledge graph circuit breaker optimization module, and a technology positioning and technology substitution difficulty assessment module.

[0024] By adopting the above technical solution, a twelve-dimensional coupled data source set is defined, incorporating semi-public or non-public signals such as the dynamic dimensions of business behavior, cross-border operations, and policy resources—which are often overlooked in traditional knowledge graph ETL processes—into the graph initialization process, avoiding the one-dimensional bias caused by "patent supremacy." Through normalization transformation of data with different dimensions aligned with domain benchmarks, the data of each dimension becomes comparable, weighted, and fusionable at the same scale, solving the fusion distortion problem when amount, number of articles, rating, and NPS index coexist. Cross-language semantic alignment based on anchor matrix resolves the ambiguity of terms such as "BMS" between battery management and bus manager, ensuring the accuracy of entity recognition. Using a default propagation mechanism instead of simple zeroing, the graph structure remains relatively stable even with incomplete data, and uncertainty is quantified through the missing impact coefficient. Standardized data interfaces output to downstream supply disruption warning and algorithm circuit breaker modules, establishing a complete data pipeline from raw data to knowledge graph and risk inference, enabling the reuse and linkage of data assets.

[0025] In some preferred embodiments of this application, the twelve-dimensional data source set includes: technology patent dimension, market dynamics dimension, consumer behavior dimension, competitive ecosystem dimension, supply chain collaboration dimension, enterprise operation dimension, capital and finance dimension, talent innovation dimension, business model paradigm dimension, cross-border operation dimension, policy resources dimension, and business behavior dynamics dimension; wherein the business behavior dynamics dimension includes at least one of public attendance record code, interview publication record code, non-public contact mark, and major contract announcement trigger mark.

[0026] In the specific implementation process, the typical composition and data sources of each dimension in the above twelve-dimensional data source set can be elaborated as follows: The first dimension is the technology patent dimension d1, which can include indicators such as the number of authorized invention patents, the number of claims, the number of citations, and the number of countries covered by the same patent family. The data can come from the State Intellectual Property Office's public database, the Derwent Innovation platform, WIPOPatentScope, etc.; The second dimension is the market dynamics dimension d2, which can include indicators such as market share in a segment, revenue growth rate, and product launch frequency. The data can come from publicly available financial reports of listed companies and reports from industry research institutions such as IDC or Gartner; The third dimension is the consumer behavior dimension d3, which can include indicators such as Net Promoter Score (NPS) score, purchase history, and social listening sentiment score. The data can come from e-commerce platform reviews and social media listening tools (such as Brandwatch, Meltwater, etc.); The fourth dimension is the competitive ecosystem dimension d4, which can include indicators such as the degree of overlap of competing products' patents and the price elasticity index; The fifth dimension is the supply chain collaboration dimension d5, which can include indicators such as the concentration of first-tier suppliers, the number of alternative suppliers, and the delivery time volatility; The sixth dimension... The enterprise operation dimension d6 can include indicators such as R&D investment ratio, R&D personnel ratio, ISO certification or specialized and innovative qualifications; the seventh dimension, capital and finance dimension d7, can include indicators such as total financing amount in the past three years, investor industry background code, and "whether it is a strategic investment" Boolean value; the eighth dimension, talent and innovation dimension d8, can include indicators such as the inverse value of core technical personnel turnover rate, number of first authors of top conference papers, and the proportion of PhDs; the ninth dimension, business model paradigm dimension d9, can include indicators such as licensing model code, software and hardware integration score, and ecosystem openness index; the tenth dimension, cross-border operation dimension d10, can include export... The data source definition module includes indicators such as destination control list matching codes, number of overseas subsidiaries, and number of local compliance certifications; the eleventh dimension is the policy resource dimension d11, which may include indicators such as the number of national-level special projects, provincial-level special project level codes, and limited partner nature codes of industry funds; the twelfth dimension is the dynamic business behavior dimension d12, which may include indicators such as codes for executives' public attendance at industry forums, codes for published interviews, non-public contact markers for internal use only, and trigger markers for major contract announcements; in actual deployment, the data source definition module solidifies the above twelve-dimensional field list and its corresponding industry benchmark values ​​in a configurable manner, which facilitates subsequent dynamic expansion and operation and maintenance.

[0027] By employing the aforementioned technical solution, and refining the composition of each dimension of the twelve-dimensional data source set, the system covers all key driving dimensions in the hard technology industry, including technology, market, consumers, competition, supply chain, operations, finance, talent, business models, cross-border transactions, policies, and behaviors. This effectively reflects the true market position and risk level of a technology entity. By incorporating semi-public or non-public signals such as executive public attendance records, interview publication records, non-public contact markers, and major contract announcement trigger markers from the dynamic dimension of business behavior into the initialization of the graph, the system compensates for the signal lag blind spots in the traditional ETL process, which is mainly based on patents and financial reports. This allows the graph to capture potential major industry events such as technological cooperation, supply chain restructuring, and M&A intentions earlier.

[0028] In some preferred embodiments of this application, the normalization transformation includes at least one of the following strategies: applying minimum-maximum normalization to continuous positive original values ​​and using an industry benchmark value as the upper bound of normalization; performing linear shift on the rating-type original values ​​according to a preset interval; mapping Boolean original values ​​to a binary set; and dividing the level-coded original values ​​by the corresponding maximum level number to obtain the normalized value.

[0029] In practice, the normalization transformation module can automatically route to different normalization strategies based on the type of data in each dimension: for continuous positive original values (For example, the amount of funding, the number of patents, the number of R&D personnel, etc.), using the minimum-maximum normalization formula, that is:

[0030] in and By referencing the historical statistical benchmarks of the top 100 companies in the industry, thus... fall into The range; for raw rating values ​​(e.g., NPS scores), the raw value range is [...]. [100, 100]) can be represented by a linear translation formula:

[0031] Translate it to Ranges; for Boolean values ​​or state flag primitives (e.g., "whether it is a strategic investment", "whether there is any non-public contact"), they can be directly mapped to sets. Where 1 represents true and 0 represents false; for the original values ​​of ordered level codes (e.g., policy-specific level codes, whose values ​​range from 1 to 5), the normalized value can be obtained by dividing by the maximum level number, i.e.:

[0032] Industry benchmarks can be fixed in the form of configuration files and can be updated quarterly or annually to adapt to the dynamic evolution of the overall industry level. As a specific example, if a company has 50 authorized invention patents, and the lower limit for the number of patents among the top 100 companies in the industry is 10 and the upper limit is 60, then the normalized attribute value for this dimension is approximately 0.8. If a company's financing amount in the past year is 200 million RMB, and the lower limit for financing amount among the top 100 companies in the industry is 50 million RMB and the upper limit is 300 million RMB, then the normalized attribute value for this dimension is approximately 0.6. If a company's NPS score is 60, then the normalized attribute value is 0.8. If a company obtains a national-level major project with a level code of 4, then the normalized attribute value is 0.8.

[0033] By adopting the above technical solution and designing corresponding normalization strategies for four typical data types, it is possible to unify heterogeneous data such as amount, number of articles, rating, Boolean value and grade code into the same numerical range while preserving the original data distribution characteristics. This lays a unified measurement foundation for subsequent multi-dimensional fusion based on vector operations. By introducing industry benchmark values ​​as the upper bound of normalization, the problem of being easily disturbed by outliers when normalizing based solely on the extreme values ​​of the sample is avoided. This ensures that the normalization results not only reflect the company's own level, but also allow for horizontal comparison with other companies in the same industry.

[0034] In some preferred embodiments of this application, the anchor matrix is ​​pre-constructed through the following steps: collecting multilingual terms and their definition paragraphs within the domain to form a structured terminology library; inputting the contextual descriptions of each term in the structured terminology library into a cross-lingual pre-trained model for encoding to obtain high-dimensional semantic vectors for each term; taking the average of the semantic vectors of the same concept in different languages ​​as the anchor vector of that concept; organizing the anchor vectors of each concept by rows to form the anchor matrix, and establishing an approximate nearest neighbor retrieval index for the anchor matrix.

[0035] In the specific implementation process, the pre-construction process of the anchor point matrix can be refined into the following five stages: The first stage is the construction of a domain terminology database. The input is a set of core terms in the hard technology field (such as "solid-state electrolyte", "SiCMOSFET", "Lidar", "perovskite battery", etc.). The sources can include the definition of International Patent Classification (IPC) classification numbers, terminology sections of industry standards such as IEC or SAE, authoritative technical dictionaries, and Wikipedia's classification directory, etc. The output format is a structured triple (term, language, domain tag), for example (Solid-state electrolyte, language, domain tag, etc.). The second stage is multilingual semantic vectorization, which uses cross-lingual pre-trained models (such as XLM-RoBERTa-large, LaBSE, mBERT, etc.) to encode each term along with its contextual description (such as Wikipedia abstracts or standard original text paragraphs). The output of the last layer of the model, CLStoken, is taken as the semantic vector v of the term, and the vector dimension is usually 768 dimensions. The third stage is concept alignment and merging, which is the vector set of the same concept in different languages. Calculate its average vector as the anchor point for this concept, that is...

[0036] As an anomaly handling mechanism, if the mean cosine similarity between a language vector and other language vectors is less than 0.70, the language expression is removed and marked as a "low-confidence translation" for manual review; the fourth stage is anchor matrix generation, which involves concatenating the anchor vectors of all concepts as rows to form... matrix ,in The total number of core concepts; to accelerate retrieval, [the following is omitted]. The first stage involves creating an FAISSindexFlatIP or IndexIVFFlat index to achieve millisecond-level similarity queries; the fifth stage is version persistence. The initial version is marked as The data is then written to a distributed object storage or a dedicated vector database to provide online retrieval services for subsequent mapping execution modules.

[0037] By adopting the above technical solution, through multilingual semantic vectorization and mean anchor alignment, the expression of the same concept in different languages ​​can correspond to the same coordinate position in the semantic space, fundamentally eliminating cross-language and cross-source terminological ambiguity. By introducing a low confidence threshold of 0.70, noise pollution of anchor vectors introduced by machine translation or non-authoritative terms is avoided. Through the FAISS approximate nearest neighbor index, millisecond-level retrieval latency can still be maintained even with a scale of millions of concepts, meeting the real-time requirements of online mapping.

[0038] In some preferred embodiments of this application, the anchor matrix supports incremental updates and version management, specifically including: when a new term is detected to appear more frequently than a preset frequency threshold within a unit time window or when a domain standard release event is detected, the anchor vector of the new concept is incrementally added to the anchor matrix; for each updated anchor matrix, a version number and timestamp are appended and historical versions are retained to support rollback.

[0039] In practical implementation, the anchor matrix management module can incorporate two types of update triggers: one is a frequency trigger, such as triggering the generation and storage of anchor points for a new concept when a new term appears more than 100 times in the data inflow over the past 7 days; the other is an event trigger, such as subscribing to new standard release events from standards organizations like IEC, IEEE, SAE, and GB / T, and actively scanning the terminology section and triggering incremental updates when a new standard involving hard technology emerges. The incremental update process only requires performing cross-language vectorization once for newly added terms and appending their anchor vectors to the matrix. At the end of the update, an add operation is performed on the FAISS index, eliminating the need to rebuild the entire matrix and significantly reducing update costs. After each update, the system generates a semantic version number for the new anchor matrix (e.g., ...). The version number, along with an update timestamp, a summary of the update scope (number of new concepts, domain tags involved), and metadata such as the reason for the update; historical versions , , , All data is persistently stored in object storage. When an anomaly is found in the new version or it is incompatible with downstream modules, the management module can roll back to any historical version with one click.

[0040] By employing the above technical solution and utilizing a dual-channel triggering mechanism of frequency triggers and event triggers, it can capture emerging and trending terms that appear during the natural evolution of data, while also timely synchronizing with standardized terms released by authoritative standards organizations, ensuring that the anchor matrix is ​​synchronized with the latest industry terminology database. Furthermore, by replacing full reconstruction with incremental updates, the update cost is reduced from... Reduce to This significantly shortens the update window; the version management mechanism, which combines version number, timestamp, and historical retention, provides deterministic anchor references for downstream modules, avoiding the problem of unreproducible historical evaluation results due to dynamic changes in the anchor matrix.

[0041] In some preferred embodiments of this application, step S3, which involves performing cross-lingual semantic alignment on technical entity keywords based on a pre-constructed anchor matrix and mapping the normalized data to node attributes and edge weights of a knowledge graph, specifically includes: calculating a query vector using a cross-lingual pre-trained model for the technical entity keywords to be mapped; retrieving the anchor vector in the anchor matrix that has the highest cosine similarity to the query vector; when the cosine similarity is greater than a first similarity threshold, determining the technical entity keyword as a corresponding concept and assigning it a normalized attribute value; and when the cosine similarity is lower than a second similarity threshold, marking the technical entity keyword as a new term awaiting manual review.

[0042] In the specific implementation process, the specific processing flow of the mapping execution module can be broken down into the following steps: First, the technical entity keyword k to be mapped (e.g., "sulfide solid electrolyte", "SiCMOSFET", "perovskite component", etc.) along with its context paragraphs in the data source are input into the cross-language pre-trained model, and the 768-dimensional vector corresponding to CLStoken is taken as the query vector. Then, using the anchor point matrix The pre-built FAISS index, for Perform top- Nearest neighbor retrieval yields the anchor vector with the highest similarity score. and their corresponding concept tags Next, cosine similarity is used as the metric to calculate...

[0043] Then the judgment phase begins: if... Greater than the first similarity threshold (For example, 0.82 or 0.85), then the keyword is determined. Hit concept Map k to the corresponding node of the concept in the knowledge graph, with its normalized attribute value taken from the output of step S2, and the initial edge weight w automatically assigned by the mapping rules; if Located between the first similarity threshold and the second similarity threshold (For example, between 0.60) then Marked as "candidate for review", it is temporarily filled with the node attribute values ​​of the candidate concept, while the downstream fusion operation is suspended and pushed to the manual review queue; if Below the second similarity threshold Then Marked as "new terminology", it will enter the manual review process and will not participate in this round of graph fusion calculation to avoid polluting other nodes.

[0044] By adopting the above technical solution, a dual-threshold judgment mechanism is used to divide the semantic space into three intervals: "high-confidence mapping", "to be verified" and "new terms". This achieves high-precision alignment of known concepts and controllable discovery of unknown concepts. By adopting the strategy of not mapping new terms after marking them, the chain pollution of erroneous mapping to the fusion operation of other parts of the graph is avoided. By introducing the cosine similarity of CLStoken vectors as a metric, both computational efficiency and semantic representation ability are taken into account, and similarity retrieval of a million-level concept database can be completed in milliseconds.

[0045] In some preferred embodiments of this application, the step S4 of propagating the default weight of the dimension to the remaining valid dimensions according to the original weight ratio in the preset weight vector specifically includes: forming a set of valid dimensions with quality scores not lower than the preset quality threshold and without missing dimensions; for any valid dimension in the set of valid dimensions, setting its updated weight to the sum of the original weight of the valid dimension and the allocation increment, wherein the allocation increment is equal to the original weight of the missing dimension multiplied by the proportion of the original weight of the valid dimension to the sum of the original weights of the set of valid dimensions.

[0046] In the specific implementation process, the processing flow of the default propagation module can be described in detail as follows: Let the original weight vector be:

[0047] and:

[0048] For the technical entity to be evaluated, the quality score of its data in each dimension is denoted as follows: Given a preset quality threshold (e.g., 0.3), effective dimension set:

[0049] For any valid dimension The updated weights are calculated as follows:

[0050] in The original weights of the currently missing dimension. This represents the incremental share allocated to that effective dimension; if multiple dimensions are missing simultaneously, their original weights are summed. As the total allocable budget, it is distributed proportionally to each valid dimension in S_valid. As a concrete numerical example, assume the original twelve-dimensional weights are uniformly distributed, i.e., each dimension... If the third dimension (supply chain collaboration dimension) is missing, and the remaining eleven dimensions are valid and their quality scores are all above the threshold, then the sum of the original weights of the valid dimensions is considered. The allocation increment obtained for each effective dimension is approximately Therefore, the new weight for each effective dimension is approximately , exactly equal to Consistent with intuition, this verifies the additivity and conservation of weight redistribution.

[0051] By employing the above technical solution, missing weights are allocated according to the original weight ratio of the effective dimensions, preserving the relative importance ranking of each dimension in the overall evaluation and avoiding the dilution of key dimension weights caused by simple average allocation. Threshold gating based on quality scores ensures that low-quality data is identified as "fact missing" and enters the default propagation branch, avoiding the distortion risk of low-quality data directly participating in the fusion operation. Furthermore, by uniformly allocating budgets for scenarios where multiple dimensions are simultaneously missing, the additivity and conservation of weight allocation are guaranteed. .

[0052] In some preferred embodiments of this application, the missing dimension impact coefficient is positively correlated with the original weight of the missing dimension and negatively correlated with the proportion of the number of valid dimensions in the twelve-dimensional data source set to the total number of dimensions, so that the more missing dimensions there are or the greater the original weight of the missing dimension, the greater the missing dimension impact coefficient.

[0053] In the specific implementation process, the missing influence coefficient It can be implemented in the following form:

[0054] in The original weights of the missing dimensions (if multiple dimensions are missing, this can be generalized to the sum of the original weights of the missing dimensions). ), This represents the number of effective dimensions. It can be intuitively seen from the above expression that: when... When the factor increases, Monotonically increasing, thus Increase; when When the number of missing dimensions decreases (i.e., the number of missing dimensions increases), the factor Increase, thus Increase; when all dimensions are valid, , This indicates that the evaluation process is complete and the confidence level is not discounted. As a numerical example, if the twelve dimensions are evenly distributed and only the third dimension is missing, then... The impact coefficient of missing data remains at a low level; if all six dimensions of data are missing, then... The missing impact coefficient increases significantly, resulting in a lower confidence level for the current assessment results when used to warn downstream users.

[0055] By employing the above technical solution, and coupling the missing dimension impact coefficient with both the original weight of the missing dimension and the proportion of the effective dimension, a two-dimensional uncertainty indicator is formed that reflects both "what is missing" and "how much is missing"; through exist The continuous values ​​between these values ​​provide a fine-grained basis for confidence level classification for the downstream circuit breaker optimization module, which can be used to trigger a multi-level algorithm switching strategy from Bayesian fusion to interval estimation and then to manual intervention.

[0056] In some preferred embodiments of this application, after recording the missing impact coefficient, the normalized values ​​of each effective dimension are summed according to the updated weights, and then divided by 1 and the sum of the missing impact coefficients to obtain a comprehensive confidence score. The comprehensive confidence score is then output as the confidence attribute of the corresponding knowledge graph node along with the node attribute. Step S5, which outputs the node attribute and edge weight as the initial value of the time-series industry knowledge graph to the downstream analysis module, specifically includes: outputting the node attribute, edge weight, missing impact coefficient, and comprehensive confidence score to the supply disruption risk simulation module through a standardized data interface as the calculation input for the difficulty weight of technology substitution; and outputting the node attribute, edge weight, and quality score to the circuit breaker optimization module through a standardized data interface as the judgment input for data quality triggering algorithm switching.

[0057] In the specific implementation process, comprehensive confidence scoring The calculation formula can be expressed as:

[0058] in For the first The normalized value (already located in) (interval) The new weights after default propagation. This represents the aforementioned missing dimension impact coefficient. As a numerical example, following the assumption that "only the third dimension is missing, and the normalized scores of the remaining eleven dimensions are all 0.7," the numerator... denominator ,final Compared to the score obtained by simply setting the missing dimension to zero (approximately...), This is significantly closer to the true level, and also through the denominator A slight penalty was applied to the confidence level. Regarding the connection with downstream modules, the interface connection module uses a standardized data interface (e.g., CSV or Parquet attribute files following Neo4j or JanusGraph batch import specifications, Kafka message topics, or gRPC services) to output node attribute vectors, initial edge weights, missing impact coefficients, comprehensive confidence scores, and anchor matrix version numbers to downstream modules: It outputs (node ​​attributes, edge weights, missing impact coefficients, and comprehensive confidence scores) to the supply disruption risk simulation module as one of the inputs for calculating the "technology substitution difficulty weight" in that module; and it outputs (node ​​attributes, edge weights, and quality scores for each dimension) to the dynamic knowledge graph circuit breaker optimization module as the basis for that module to determine whether to trigger a switch from a deep fusion model to a linear fusion model or a manual rule model.

[0059] Using the above technical solution, by summing weighted normalized values ​​in the numerator of the comprehensive confidence score and introducing weighted normalized values ​​into the denominator... The factor enables the decoupling of the "assessment result itself" and the "confidence level of the assessment result", allowing downstream modules to obtain both the point estimate and the uncertainty description of that point estimate. Through standardized data interfaces, it outputs to the supply disruption risk simulation module and the circuit breaker optimization module, realizing a complete data pipeline closed loop from data collection, semantic alignment, multi-dimensional fusion to risk simulation and algorithm circuit breaker, forming a three-in-one linkage mechanism.

[0060] In some embodiments of this application, a knowledge graph normalization mapping system based on twelve-dimensional coupled signals is also provided. The system includes: a data source definition module configured to define and manage a set of twelve-dimensional data sources; a normalization transformation module configured to perform a normalization transformation aligned with a domain benchmark on the original values ​​of each dimension of the data in the set of twelve-dimensional data sources; an anchor matrix management module configured to construct, update, and retrieve anchor matrices for cross-language semantic alignment; a mapping execution module configured to map the normalized data to node attributes and edge weights of the knowledge graph based on the anchor matrix; a default propagation module configured to allocate weights to the remaining valid dimensions according to the original weight ratio in a preset weight vector and record the missing impact coefficient when any dimension of data is missing or its quality score is lower than a preset quality threshold; and an interface connection module configured to output the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module.

[0061] In practical implementation, the modules of the above system can be deployed and integrated using a microservice architecture: the data source definition module uses a configuration center (such as Apollo, Nacos, etc.) as a carrier to solidify the field list, industry benchmark values, unit conversion rules, and metadata of the twelve-dimensional data source; the normalization transformation module uses a stateless computing service (such as based on Flink, Spark Streaming, or an independent containerized service) as a carrier to realize the normalization routing and execution of four types of data: continuous positive values, rating types, booleans, and grade codes; the anchor matrix management module uses a vector database (such as Milvus, Weaviate, FAISSonRedis, etc.) as the core to expose four types of interfaces: construction, incremental update, version rollback, and similarity retrieval; the mapping execution module uses an online inference service (such as based on TorchServe, TensorRT, or ONNXRuntime) as a carrier to load cross-language pre-trained models and call the anchor matrix interface to complete real-time mapping; the default propagation module uses a combination of a rule engine and computing services to realize weight redistribution and The computation and interface connection module is jointly deployed with message queues (such as Kafka, Pulsar, etc.) and API gateways (such as Kong, APISIX, etc.) to push standardized data to the supply disruption risk simulation module and the circuit breaker optimization module. The above modules can communicate with each other using gRPC or RESTful HTTP protocols, and cross-module calls can be monitored through distributed tracing (such as OpenTelemetry, Jaeger, etc.).

[0062] By adopting the above technical solution, through modular functional decomposition and microservice deployment, each module can be developed, tested, and scaled up or down independently, significantly improving the maintainability and elasticity of the system. Through a unified interface connection module and standardized data interface, the system can seamlessly connect to downstream knowledge graphs and risk inference systems from different vendors and with different technology stacks. Through the version rollback capability of the anchor matrix management module and the confidence recording capability of the default propagation module, the entire system is provided with complete observability and traceability assurance.

[0063] In some embodiments of this application, this application also provides an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the knowledge graph normalization mapping method based on the twelve-dimensional coupled signal described above.

[0064] In specific implementations, mobile electronic devices can be, for example, mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), televisions (TVs), ATMs, or self-service machines, etc. This application embodiment does not impose specific limitations. The electronic device used for knowledge graph normalization mapping based on twelve-dimensional coupled signals in this application embodiment can be an electronic device with an operating system. This operating system can be Android, iOS, or other possible operating systems, which this application embodiment does not specifically limit.

[0065] By adopting the above technical solution, the knowledge graph normalization mapping method based on twelve-dimensional coupled signals is solidified into a program or instruction that can run on electronic devices. This enables the method of this application to be deployed and run on various hardware forms such as servers, workstations or cloud computing nodes. It facilitates integration with existing enterprise data warehouses, graph platforms and risk warning platforms, and significantly expands the product application scenarios of the method.

[0066] In some embodiments of this application, this application also provides a readable non-transitory storage medium storing a program or instructions that, when executed by a processor, implement the steps of the knowledge graph normalization mapping method based on twelve-dimensional coupled signals described above.

[0067] In specific implementation, the readable non-transitory storage medium includes computer-readable non-transitory storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0068] By adopting the above technical solution, the program or instructions of the method of this application are pre-fixed on a readable non-transitory storage medium, which facilitates external distribution, cross-system migration and disaster recovery. At the same time, it can be used as a software product or software module to enter the functional matrix of related software products such as knowledge graphs, risk warning, and supply chain analysis, thus expanding the application channels of the method of this application.

[0069] Based on the above implementation schemes, this application defines a twelve-dimensional coupled data source set including the dynamic dimension of business behavior, unifies the dimensions of heterogeneous data with four types of normalization strategies, achieves cross-language and cross-source semantic alignment with a multilingual anchor matrix, handles missing data and quantifies uncertainty with a default propagation mechanism based on quality thresholds and original weight ratios, and finally outputs node attributes, edge weights, missing impact coefficients, and comprehensive confidence scores to the supply disruption risk inference module and the dynamic knowledge graph circuit breaker optimization module through a standardized data interface. This forms a complete closed loop from raw data to knowledge graph to risk inference and algorithm circuit breaker, systematically solving the core problems of existing knowledge graph ETL processes from four aspects: data dimension, dimensions, missing data handling, and downstream connection. It can effectively support high-value application scenarios such as technology positioning, risk warning, and strategic decision-making in the hard technology industry.

[0070] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A knowledge graph normalization mapping method based on twelve-dimensional coupled signals, characterized in that, The specific steps are as follows: Step S1: Define a 12-dimensional data source set for constructing an industry knowledge graph, wherein the 12-dimensional data source set contains at least a dynamic dimension of business behavior; Step S2: Perform a normalization transformation on the original values ​​of each dimension of the data in the twelve-dimensional data source set, aligned with the domain benchmark, to convert the original values ​​of different dimensions to a preset numerical range. Step S3: Based on the pre-constructed anchor matrix, perform cross-language semantic alignment on the keywords of technical entities, and map the normalized data into node attributes and edge weights of the knowledge graph. Step S4: For any dimension of data in the twelve-dimensional data source set, when the data of that dimension is missing or the quality score of the data of that dimension is lower than the preset quality threshold, the data is propagated to the other effective dimensions by default according to the original weight ratio of that dimension in the preset weight vector, and the missing impact coefficient is recorded. Step S5: Output the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module as standardized inputs for graph construction and analysis.

2. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that, The twelve-dimensional data source set mentioned in step S1 includes: technology patent dimension, market dynamics dimension, consumer behavior dimension, competitive ecosystem dimension, supply chain collaboration dimension, enterprise operation dimension, capital and finance dimension, talent innovation dimension, business model paradigm dimension, cross-border operation dimension, policy resources dimension, and business behavior dynamics dimension; wherein the business behavior dynamics dimension includes at least one of the following: public attendance record code, interview publication record code, non-public contact mark, and major contract announcement trigger mark.

3. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that, The normalization transformation in step S2 includes at least one of the following strategies: applying minimum-maximum normalization to continuous positive original values ​​and using the industry benchmark value as the upper bound of normalization; performing linear shift on the rating type original values ​​according to a preset interval; mapping Boolean original values ​​to a binary set; and dividing the level-coded original values ​​by the corresponding maximum level number to obtain the normalized value.

4. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that, The anchor matrix mentioned in step S3 is pre-constructed through the following steps: collecting multilingual terms and their definition paragraphs in the domain to form a structured terminology library; inputting the context description of each term in the structured terminology library into a cross-language pre-trained model for encoding to obtain a high-dimensional semantic vector for each term; taking the average of the semantic vectors of the same concept in different languages ​​as the anchor vector of the concept; organizing the anchor vectors of each concept by rows to form the anchor matrix, and establishing an approximate nearest neighbor retrieval index for the anchor matrix.

5. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 4, characterized in that: The anchor matrix supports incremental updates and version management, specifically including: when a new term is detected to appear more frequently than a preset frequency threshold within a unit time window or when a domain standard release event is detected, the anchor vector of the new concept is incrementally added to the anchor matrix; for each updated anchor matrix, a version number and timestamp are appended and historical versions are retained to support rollback.

6. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that: Step S3, which involves performing cross-lingual semantic alignment on technical entity keywords based on a pre-constructed anchor matrix and mapping the normalized data to node attributes and edge weights of a knowledge graph, specifically includes: calculating a query vector using a cross-lingual pre-trained model for the technical entity keywords to be mapped; retrieving the anchor vector with the highest cosine similarity to the query vector from the anchor matrix; determining the technical entity keyword as a corresponding concept and assigning it a normalized attribute value when the cosine similarity is greater than a first similarity threshold; and marking the technical entity keyword as a new term awaiting manual review when the cosine similarity is lower than a second similarity threshold.

7. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that, The step S4, which involves propagating the default weight of the dimension to the remaining valid dimensions according to the original weight ratio in the preset weight vector, specifically includes: forming a set of valid dimensions with quality scores not lower than the preset quality threshold and without missing dimensions; for any valid dimension in the set of valid dimensions, updating its weight to the sum of the original weight of the valid dimension and the allocation increment, wherein the allocation increment is equal to the original weight of the missing dimension multiplied by the proportion of the original weight of the valid dimension to the sum of the original weights of the set of valid dimensions.

8. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that: The missing dimension impact coefficient is positively correlated with the original weight of the missing dimension and negatively correlated with the proportion of the number of valid dimensions in the twelve-dimensional data source set to the total number of dimensions. This means that the more missing dimensions there are or the greater the original weight of the missing dimension, the greater the missing dimension impact coefficient.

9. The knowledge graph normalization mapping method based on twelve-dimensional coupled signals according to claim 1, characterized in that: After recording the missing impact coefficient, the normalized values ​​of each effective dimension are summed according to the updated weights, and then divided by 1 and the sum of the missing impact coefficients to obtain the comprehensive confidence score. The comprehensive confidence score is then output as the confidence attribute of the corresponding knowledge graph node along with the node attributes. Step S5, which involves outputting the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module, specifically includes: outputting the node attributes, edge weights, missing impact coefficients, and comprehensive confidence scores to the supply disruption risk simulation module through a standardized data interface, as input for calculating the difficulty weight of technology substitution; and outputting the node attributes, edge weights, and quality scores to the circuit breaker optimization module through a standardized data interface, as input for determining whether data quality triggers algorithm switching.

10. A knowledge graph normalization mapping system based on twelve-dimensional coupled signals, applied to the knowledge graph normalization mapping method based on twelve-dimensional coupled signals as described in any one of claims 1 to 9, characterized in that, include: The data source definition module is configured to define and manage a twelve-dimensional data source set; The normalization transformation module is configured to perform a normalization transformation aligned with the domain benchmark on the original values ​​of each dimension of the data in the twelve-dimensional data source set. The anchor matrix management module is configured to build, update, and retrieve anchor matrices for cross-language semantic alignment. The mapping execution module is configured to map the normalized data into node attributes and edge weights of the knowledge graph based on the anchor matrix. The default propagation module is configured to distribute weights to the remaining valid dimensions according to the original weight ratio in the preset weight vector and record the missing impact coefficient when any dimension of data is missing or its quality score is lower than the preset quality threshold. The interface connection module is configured to output the node attributes and edge weights as initial values ​​of the time-series industry knowledge graph to the downstream analysis module.

Images