Data-driven cross-domain product informatization management system

Abstract

The invention relates to the technical field of management systems, and particularly discloses a data-driven cross-domain product informatization management system, which comprises a multi-modal product data center, a cross-domain product knowledge graph construction module, an intelligent product configuration and recommendation engine, a product full life cycle tracking and feedback module and a system adaptive evolution module. According to the method, heterogeneous data in different fields are mapped to a unified semantic space through standardized processing and unified semantic coding of the multi-modal product data center, the problems that cross-field data in a traditional system are not communicated and poor in comparability are solved, and a unified data basis is provided for cross-field product management; according to the method, a personalized configuration scheme adaptive to requirements can be generated without manual intervention by means of deep mining of a cross-domain product knowledge graph on a product association relationship and a scene mapping relationship and combining with requirement analysis and multi-target optimization capability of an intelligent configuration and recommendation engine, so that the scientificity, feasibility and adaptation efficiency of the configuration scheme are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of management system technology, specifically to a data-driven cross-domain product information management system. Background Technology

[0002] In the field of product information management, with the increasing trend of cross-industry integration, enterprises have an increasingly urgent need for collaborative management of products across multiple domains. Traditional product management systems are mostly developed for single business domains, resulting in significant technical limitations: First, the system architecture is rigid, with management platforms for each domain operating independently, forming data silos and hindering effective interoperability and correlation analysis of product data across domains. Second, outdated data processing methods are difficult to adapt to multimodal heterogeneous data such as structured, semi-structured, and unstructured data, and the lack of a unified semantic coding standard leads to poor comparability of cross-domain data. Third, product configuration relies on manual experience or fixed rules, failing to intelligently generate optimized solutions based on dynamic user needs and changes in business scenarios, resulting in insufficient configuration flexibility and accuracy. Fourth, the lack of a full lifecycle performance tracking and feedback mechanism means that system decision-making strategies cannot be iteratively optimized based on actual application results, remaining in a static operating state for a long time and unable to adapt to the continuous development and changes in business. These problems lead to low management efficiency and poor adaptability of traditional systems, severely restricting the quality and efficiency of cross-domain product collaborative management and failing to meet enterprises' needs for intelligent, flexible, and evolvable product information management during digital transformation. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a data-driven cross-domain product information management system, which solves the problems mentioned in the background technology.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solution: a data-driven cross-domain product information management system, including a multimodal product data center, a cross-domain product knowledge graph construction module, an intelligent product configuration and recommendation engine, a product lifecycle tracking and feedback module, and a system adaptive evolution module; The multimodal product data center is used to connect to heterogeneous product data sources from different business domains, and generates a standardized product metadata vector set through standardized cleaning, feature extraction, and unified semantic encoding. ; The cross-domain product knowledge graph construction module is based on product metadata vector sets. It automatically mines product attribute associations, functional substitution / complementary relationships between products, and product-application scenario mapping relationships, and constructs and dynamically updates a unified cross-domain product knowledge graph. ; The intelligent product configuration and recommendation engine receives requirement descriptions from users or upper-layer applications and is based on a cross-domain product knowledge graph. Perform semantic analysis and knowledge retrieval of requirements to generate a set of multi-objective optimized product configuration solutions. And complete the feasibility assessment and benefit forecast; After the configuration scheme is implemented, the product lifecycle tracking and feedback module continuously collects product operation status data, performance indicators, and user feedback to form a performance tracking record for the product instance. ; The system's adaptive evolution module is based on performance tracking records. Analyze the discrepancy between expected and actual performance of configuration schemes to drive cross-domain product knowledge graphs. The confidence level of the association is updated to optimize the decision-making strategy of the intelligent product configuration and recommendation engine.

[0005] Preferably, the multimodal product data center includes a data access unit, an intelligent preprocessing unit, a semantic encoding unit, and a vector fusion unit; The data access unit connects to product data sources from different fields through standardized interfaces, including structured data, semi-structured data, and unstructured data, to form the original product dataset. The intelligent preprocessing unit performs AI-driven preprocessing operations on the original product dataset. The specific steps are as follows: S11. Parse the original product dataset to distinguish between numerical, textual, and image-based product data; S12. Use the Z-score standardization method to eliminate the influence of dimensions for numerical data and calculate the standardization error. ; S13. Use the TF-IDF algorithm to extract keyword features from text data and filter out low-frequency meaningless words; S14. Use an adaptive threshold segmentation method to extract contour features from image data and calculate the contour sharpness coefficient. ; S15. Substitute into the formula to calculate the data quality optimization coefficient. ,like Return to step S12 to re-perform the standardization process until... ; Among them, the data quality optimization coefficient The calculation formula is as follows: ; in, This is the standardization error, ranging from 0 to 1. The smaller the error, the better. The closer a value is to 0, the greater the consistency of the standardized numerical data. This is the outline sharpness coefficient, ranging from 0 to 1. The sharper the outline edges and the less noise, the better. The closer to 1; This is a data preprocessing correction constant, ranging from 0.03 to 0.08, used to compensate for the systematic bias of different types of data preprocessing algorithms; The semantic coding unit employs a domain-adaptive encoder to encode the first... The first field Features of each product after pretreatment Mapping to a unified semantic space yields standardized product metadata vectors. The mapping process is as follows: ; in, For the target field The trainable encoding network adopts a hybrid architecture of CNN+Transformer; These are the learnable parameters of the encoding network; A common semantic vector shared across all domains, with dimensions of 128-512, is used to ensure semantic consistency of features across domains; The vector fusion unit is used to perform weighted fusion of semantic vectors from multiple domains. The specific steps are as follows: S16. Calculate the reliability score of the product's data sources in various relevant fields. The value ranges from 0 to 1; the higher the data integrity, the better. The closer to 1; S17. Standardized product metadata vectors for various fields Assign weights ,satisfy The weight calculation formula is: ; S18. Obtain the final product metadata vector through weighted summation. .

[0006] Preferably, the cross-domain product knowledge graph construction module includes a node definition unit, an edge generation unit, a graph optimization unit, and a graph dynamic update module; The node definition unit defines three types of core nodes: product instance nodes. Attribute concept nodes Application scenario nodes Each node is assigned a unique identifier and basic attribute label. The association edge generation unit constructs three types of core edges based on product metadata vectors, and the specific steps are as follows: S21. Calculate the product metadata vector. and Using Euclidean distance, generate edges for products with similar functions. Its weight The calculation is as follows: ; in, This is a scaling parameter, with a value range of 0.5-2.0. The larger the distance, the faster the similarity decreases with distance; The square of the Euclidean distance between the two vectors; The value ranges from 0 to 1, and the closer the value is to 1, the higher the similarity of the functions of the two products. S22. Based on the co-occurrence frequency of products in historical configuration schemes or orders, generate edges for product collaboration / substitution through point mutual information measurement. Its weight The value ranges from 0 to 1, and the higher the co-occurrence frequency, the greater the weight. S23. Calculate the cosine similarity between the product metadata vector and the scene feature vector. Generate product and applicable scenario edges Set a similarity threshold (Value range: 0.7-0.85), similarity ≥ The edge should be preserved at that time. The graph optimization unit uses a graph neural network to learn the representation of the initially constructed graph, obtains node embedding vectors, completes the potential associated edges, and calculates the graph integrity coefficient. ,like Return to step S21 to re-optimize until... ; Among them, the map integrity coefficient The calculation formula is as follows: ; in: This represents the actual number of associated edges collected. This represents the number of potential associated edges completed using GNN; The theoretical maximum number of associated edges between the three types of core nodes. ; This represents the average edge weights, ranging from 0 to 1. A higher average value indicates a higher weight distribution across all associated edges. The closer to 1; The dynamic map update module is used to continuously optimize the map based on newly added product data. The specific steps are as follows: S24. Set update cycle Regularly collect metadata vectors for newly added products; S25. Calculate the associated edges between the newly added product node and the existing node, and repeat steps S21-S23 to generate new associated edges. S26. Calculate the weight of the impact of the newly added edge on the existing graph. ,like This will trigger an overall optimization of the graph; S27. Recalculate the map integrity coefficient. ,like If the update is successful, the update is complete; otherwise, return to S25 to adjust the parameters for generating associated edges. S28. Record the graph update log, including update time, number of new nodes / edges, and change in integrity coefficient.

[0007] Preferably, the intelligent product configuration and recommendation engine includes a demand analysis unit, a solution optimization unit, a solution evaluation unit, and a solution screening module; The requirement parsing unit will analyze the user's natural language requirement description. Transform into demand vectors through a pre-trained BERT model. ,calculate With knowledge graph Mid-scene nodes Cosine similarity of embedded vectors is used to pinpoint core application scenarios. The similarity calculation formula is as follows: ; The solution optimization unit models the product configuration problem as a knowledge graph. The problem of multi-objective path search on [a specific target], under constraints, generates a set of candidate configuration schemes. The specific steps are as follows: S31. Analyze the constraints in user requirements and convert them into quantitative indicator thresholds; S32. Starting from the core demand scenario node, search the knowledge graph for product node paths that meet the constraints. The path length is limited to 2-5 nodes. S33. Validate the product combination validity of the searched paths and eliminate combinations with functional conflicts or incompatible parameters; S34. Effective combinations are packaged into candidate configuration schemes, each scheme including a product list, parameter configuration, cost estimate, and delivery cycle forecast. The scheme evaluation unit evaluates each candidate configuration scheme Calculate the overall score The formula is as follows: ; in, This is a matching function between the solution and the requirements, calculated based on the similarity between the aggregated value of the solution product vector and the requirement vector, with a value range of 0-1. The robustness function of the scheme is calculated as follows: , To determine the average degree centrality of the product nodes in the solution, The value of the alternative edge richness of the solution product ranges from 0 to 1. The cost-effectiveness function of the solution is calculated as follows: After normalization, the value ranges from 0 to 1. Candidate solutions The overall performance indicators Candidate solutions Total lifecycle cost; For adjustable weighting coefficients, satisfying It can be dynamically adjusted according to business scenarios; The scheme screening module is used for the final screening of candidate schemes, and the specific steps are as follows: S35. Set the scoring threshold (Value range 0.8-0.9), filter out scores ≥ The solution; S36. If the number of solutions after screening is greater than 5, calculate the similarity between the solutions, remove redundant solutions with a similarity of ≥0.85, and retain no more than 5 differentiated solutions. S37. If the number of solutions after screening is less than 2, lower the scoring threshold by 0.05 and repeat S35-S36 until the number of solutions is greater than or equal to 2. S38. Sort the final selected solutions in descending order of comprehensive score to form a set of recommended solutions and output it.

[0008] Preferably, the product lifecycle tracking and feedback module includes a data acquisition unit, a performance calculation unit, a deviation analysis unit, and a feedback data processing module; The data acquisition unit continuously collects product operating status data, performance indicators, and user satisfaction scores through sensors, system log interfaces, user feedback questionnaires, etc. The sampling frequency is 1 time / hour to 1 time / day, which can be dynamically adjusted according to the product type. The performance calculation unit is configured for each implementation. Create tracking instances to record performance metric vectors. Over time Changes, calculating actual comprehensive effectiveness The formula is as follows: ; in, The observation period is 30-180 days, depending on the product lifecycle length. The number of performance indicators ranges from 5 to 15, covering dimensions such as operating status, performance, and user experience. As the indicator weight, satisfying This was determined using the analytic hierarchy process (AHP). For the first Each indicator in time The values ​​have been standardized to the range of 0-1; The deviation analysis unit calculates the expected overall performance based on the prediction model generated during scheme generation. And solve for the performance deviation. : ; Set deviation threshold ,when When this happens, mark the solution as one that requires priority optimization; The feedback data processing module is used to clean and classify the collected feedback data. The specific steps are as follows: S41. Perform outlier detection on numerical feedback data and use the 3σ criterion to remove outlier data. S42. Perform sentiment analysis and keyword extraction on text-based feedback data, and classify it into positive feedback, negative feedback, and neutral feedback. S43. Map negative feedback to corresponding performance indicators and generate a correlation table between feedback and indicators; S44. Calculate the percentage of negative feedback for each performance indicator. ,like Mark this indicator as an indicator that needs to be monitored closely; S45. Compare the processed feedback data with the performance deviation. The data is bound together to form a complete feedback data packet and transmitted to the system's adaptive evolution module.

[0009] Preferably, the system adaptive evolution module includes a graph update unit, a strategy optimization unit, an iterative verification unit, and a decision strategy update module; The map update unit is based on performance deviation. Dynamically adjust cross-domain product knowledge graph confidence of related edges The formula is as follows: ; in, This represents the original confidence level of the associated edge, with a value ranging from 0 to 1. The updated confidence level must satisfy 0 ≤ ≤1, if the value is outside the range, the boundary value is used; The learning rate, ranging from 0.05 to 0.15, controls the adjustment range of the confidence level. For a sign function, when hour ,when hour ,when hour ; For the edge With the plan The contribution relevance, with a value ranging from 0 to 1, is calculated by the importance of the edge in the solution generation path; The policy optimization unit models the scheme generation and selection process as a reinforcement learning problem, using a knowledge graph. The environmental state is the configuration scheme as the action, and the actual comprehensive effectiveness is... As a reward, the decision network parameters of the recommendation engine are optimized using the policy gradient method; The iterative verification unit collects performance data of newly added configuration schemes every quarter to verify the updated graph and optimized strategy. If the strategy accuracy is ≥95% and the graph integrity coefficient is [not specified], the verification is successful. If the result is positive, the current iteration is complete; otherwise, return to the graph update unit for re-optimization. The decision strategy update module is used to dynamically update the decision rules of the recommendation engine. The specific steps are as follows: S51. Calculate the fit of the current decision-making strategy with the recent configuration scheme. Value range 0-1, fit , This refers to the number of recent proposals; S52, Set the fit threshold ,like This triggers an update to the decision-making strategy; S53. Confirm the weight adjustment requirements for each performance indicator as a component of the decision-making strategy. The weight adjustment range is ≤0.1; S54. Confirm the direction of similarity threshold optimization for scheme selection, as a component of the decision-making strategy. The optimization range is ±0.05; S55. Perform constraint condition extension requirement judgment and output the judgment result; S56. When the judgment result indicates that expansion is needed, add new constraint types and thresholds as components of the decision-making strategy. ; S57. When the judgment result of S55 is "necessary", the constituent elements will be... Integrate into updated decision-making strategies; when the judgment result is that it is unnecessary, the constituent elements will be... Integrate into updated decision-making strategies; S58. Replace the current decision strategy with the updated decision strategy, and record the strategy update log, including update time, changes in adaptability, and adjustments to the constituent elements.

[0010] Preferably, the demand parsing unit of the intelligent product configuration and recommendation engine also supports multilingual demand input. It automatically identifies the input language type through a language detection model, calls the pre-trained model of the corresponding language for vector conversion, and achieves a language recognition accuracy of ≥98% and a demand matching degree loss of ≤5% after vector conversion.

[0011] Preferably, the data-driven cross-domain product information management system further includes a cross-domain product data governance dashboard; The dashboard visually displays the data source access status, data quality score, and cross-domain product knowledge graph of the multimodal product data center. The topology and edge confidence distribution, the decision accuracy trend of the intelligent recommendation engine, and the performance change curve throughout the product lifecycle; The dashboard allows administrators to manually adjust core system parameters and thresholds. The specific steps are as follows: S61. The administrator selects the type of parameter to be adjusted; S62. Input a new parameter value, and the system will automatically check whether the parameter value is within a reasonable range. S63. If the parameter value is reasonable, generate a parameter adjustment preview to show the possible impact of the adjustment on system performance. S64. After the administrator confirms the adjustment, the system updates the parameters and records the adjustment log; S65. If the parameter value is unreasonable, the system will prompt a reasonable range of values ​​and return to S62 to re-enter.

[0012] Preferably, the data-driven cross-domain product information management system is deployed using a microservice architecture, with the multimodal product data center, cross-domain product knowledge graph construction module, and intelligent product configuration and recommendation engine each serving as an independent scalable microservice, supporting elastic scaling up and down based on business load. The data-driven, cross-domain product information management system supports the configuration of adaptive encoders for different domains. Based on the domain ontology rules, new product domains can be quickly integrated without changing the core architecture. The new domain integration process is as follows: S71. Configure the data source interface parameters for the new field and connect to the data source of the new field product. S72, Training Adaptive Encoders for New Domains Generate a domain-specific coding network; S73. Import ontology rules for new domains; S74. Collect seed product data in new fields and initialize the knowledge graph; S75. Verify the accuracy of the configuration recommendation for the new product. If the accuracy is ≥90%, the integration is complete; otherwise, return to S72 to optimize the encoder. The access period for new domains is ≤15 working days. After access, cross-domain configuration recommendations with existing domain products are supported.

[0013] This invention provides a data-driven, cross-domain product information management system, which has the following beneficial effects: 1. This invention maps heterogeneous data from different fields to a unified semantic space through standardized processing and unified semantic coding of multimodal product data centers, solving the problems of cross-domain data incompatibility and poor comparability in traditional systems, and providing a unified data foundation for cross-domain product management.

[0014] 2. By leveraging the cross-domain product knowledge graph to deeply explore product relationships and scenario mapping relationships, and combining the demand analysis and multi-objective optimization capabilities of intelligent configuration and recommendation engines, personalized configuration solutions that meet specific needs can be generated without manual intervention, significantly improving the scientific nature, feasibility, and adaptation efficiency of configuration solutions.

[0015] 3. By collecting actual application data through the product lifecycle tracking and feedback module, and combining the system's adaptive evolution module to dynamically optimize the knowledge graph association confidence and engine decision-making strategy, a closed-loop mechanism of "data-configuration-feedback-optimization" is constructed, enabling the system to continuously adapt to business changes and maintain efficient decision-making and management capabilities in the long term.

[0016] 4. This invention covers the entire process of product management, from data access and configuration generation to operation tracking and feedback optimization. It not only realizes full-link informatization of product management, but also provides direct basis for system optimization and product iteration through the in-depth application of user feedback and performance data, thereby improving the integrity and closed-loop value of product management.

[0017] 5. Adopting a microservice architecture and a configurable domain adaptation mechanism, it supports rapid access to new domains and can expand management capabilities without changing the core architecture; the visual presentation and adjustable parameters of the cross-domain product data governance dashboard reduce the difficulty of system operation and maintenance and improve the transparency and convenience of management. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the process of a data-driven cross-domain product information management system according to the present invention; Figure 2 This is a schematic diagram of the principle of a data-driven cross-domain product information management system according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figures 1-2 As shown, the present invention provides a technical solution: a data-driven cross-domain product information management system, characterized in that it includes a multimodal product data center, a cross-domain product knowledge graph construction module, an intelligent product configuration and recommendation engine, a product lifecycle tracking and feedback module, and a system adaptive evolution module; The multimodal product data center is used to connect to heterogeneous product data sources from different business domains. Through standardized cleaning, feature extraction, and unified semantic encoding, it generates a standardized product metadata vector set. ; The cross-domain product knowledge graph construction module is based on product metadata vector sets. It automatically mines product attribute associations, functional substitution / complementary relationships between products, and product-application scenario mapping relationships, and constructs and dynamically updates a unified cross-domain product knowledge graph. ; The intelligent product configuration and recommendation engine receives requirement descriptions from users or upper-layer applications and is based on a cross-domain product knowledge graph. Perform semantic analysis and knowledge retrieval of requirements to generate a set of multi-objective optimized product configuration solutions. And complete the feasibility assessment and benefit forecast; After the configuration scheme is implemented, the product lifecycle tracking and feedback module continuously collects product operation status data, performance indicators, and user feedback to form a performance tracking record for the product instance. ; The system adaptive evolution module is based on performance tracking records. Analyze the discrepancy between expected and actual performance of configuration schemes to drive cross-domain product knowledge graphs. The confidence level of the association is updated to optimize the decision-making strategy of the intelligent product configuration and recommendation engine.

[0021] More specifically, the multimodal product data center includes a data access unit, an intelligent preprocessing unit, a semantic coding unit, and a vector fusion unit; The data access unit connects to product data sources from different fields such as manufacturing, finance, and healthcare through standardized interfaces, including structured data, semi-structured data, and unstructured data, to form the original product dataset. The intelligent preprocessing unit performs AI-driven preprocessing operations on the original product dataset. The specific steps are as follows: S11. Parse the original product dataset to distinguish between numerical, textual, and image-based product data; S12. Use the Z-score standardization method to eliminate the influence of dimensions for numerical data and calculate the standardization error. ; S13. Use the TF-IDF algorithm to extract keyword features from text data and filter out low-frequency meaningless words (word frequency less than 0.01% of the total word frequency). S14. For image-type data (such as product appearance images and structural diagrams), use the adaptive threshold segmentation method to extract contour features and calculate the contour sharpness coefficient. ; S15. Substitute into the formula to calculate the data quality optimization coefficient. ,like Return to step S12 to re-perform the standardization process until... ; Among them, the data quality optimization coefficient The calculation formula is as follows: ; in, This is the standardization error, ranging from 0 to 1. The smaller the error, the better. The closer a value is to 0, the greater the consistency of the standardized numerical data. This is the outline sharpness coefficient, ranging from 0 to 1. The sharper the outline edges and the less noise, the better. The closer to 1; This is a data preprocessing correction constant, ranging from 0.03 to 0.08, used to compensate for the systematic bias of different types of data preprocessing algorithms; The semantic coding unit employs a domain-adaptive encoder to convert the first... The first field Features of each product after pretreatment Mapping to a unified semantic space yields standardized product metadata vectors. The mapping process is as follows: ; in, For the target field The trainable encoding network adopts a hybrid architecture of CNN+Transformer; These are the learnable parameters of the encoding network, including convolutional kernel weights, attention mechanism parameters, etc. A common semantic vector shared across all domains, with dimensions of 128-512, is used to ensure semantic consistency of features across domains; The vector fusion unit is used to weightedly fuse semantic vectors from multiple domains. The specific steps are as follows: S16. Calculate the reliability score of the product's data sources in various relevant fields. The value ranges from 0 to 1; the higher the data integrity, the better. The closer to 1; S17. Standardized product metadata vectors for various fields Assign weights ,satisfy The weight calculation formula is: ; S18. Obtain the final product metadata vector through weighted summation. .

[0022] The multimodal product data center achieves standardized end-to-end processing of heterogeneous data (structured, semi-structured, and unstructured) from different domains through the collaborative operation of data access units, intelligent preprocessing units, semantic coding units, and vector fusion units. The data access unit breaks down barriers between domain data source interfaces, enabling unified access to multiple data types. The intelligent preprocessing unit employs AI-driven targeted processing algorithms (Z-score, TF-IDF, adaptive threshold segmentation) combined with data quality optimization coefficients. Closed-loop verification ensures data consistency and validity; the semantic encoding unit utilizes a domain-adaptive encoder based on a CNN+Transformer hybrid architecture and common semantic vectors. It maps features from various domains to a unified semantic space, solving the problem of cross-domain semantic heterogeneity; the vector fusion unit dynamically allocates weights based on the reliability of the data source to generate the most comprehensive and optimal product metadata vector.

[0023] The entire process completely solves the core pain points of traditional systems, such as data incompatibility and poor comparability across domains, and provides a standardized, highly reliable, and cross-domain compatible data foundation for subsequent modules such as cross-domain product knowledge graph construction and intelligent configuration recommendation.

[0024] In this embodiment, three cross-domain scenarios—manufacturing, finance, and healthcare—are selected for data processing verification: the manufacturing domain receives structured equipment parameter tables (including numerical data such as power and dimensions), the finance domain receives semi-structured financial product manuals (text data), and the healthcare domain receives unstructured surgical instrument diagrams (image data). The data access unit completes the unified collection of the three types of data sources through a RESTful standardized interface, forming the original product dataset. In the intelligent preprocessing stage, the Z-score standardization method is used to calculate the standardization error of the numerical equipment power data in the manufacturing domain. , the TF-IDF algorithm is used to extract keywords from the product instruction text in the financial field, and meaningless words such as "de" and "ji" with a word frequency lower than 0.01% of the total word frequency are filtered. The adaptive threshold segmentation method is used to extract contour features from the surgical instrument structure diagram in the medical field, and the contour clarity coefficient is calculated , take the data preprocessing correction constant , substitute into the calculation formula to get , meeting the data quality requirements.

[0025] The semantic encoding unit uses a domain adaptive encoder with a CNN+Transformer hybrid architecture to map the preprocessed features of the three domains to a unified semantic space of 256 dimensions (the common semantic vector dimension is 256 dimensions), obtaining the device metadata vector in the manufacturing field , the product metadata vector in the financial field , the instrument metadata vector in the medical field ; in the vector fusion stage, the reliability scores of the data sources in the manufacturing, financial, and medical fields are calculated as , , , according to the weight formula calculate to get , , and the final product metadata vector is obtained through weighted summation. This vector realizes the unified semantic expression of cross-domain product features in manufacturing, finance, and medicine, and can be directly used for subsequent cross-domain association relationship mining and intelligent configuration scheme generation.

[0026] More specifically, the cross-domain product knowledge graph construction module includes a node definition unit, an associated edge generation unit, a graph optimization unit, and a graph dynamic update module; The node definition unit defines three types of core nodes: product instance nodes (store product model, parameters, price and other attributes), attribute concept nodes (store general product attribute categories), application scenario nodes (store the business scenario descriptions applicable to the product), and assign a unique identifier and a basic attribute label to each node; The associated edge generation unit constructs three types of core edges based on the product metadata vector. The specific steps are as follows: S21. Calculate the Euclidean distance between the product metadata vector and to generate an edge for similar functions between products, and its weight is calculated as: ; Among them, This is a scaling parameter, with a value range of 0.5-2.0. The larger the distance, the faster the similarity decreases with distance; The square of the Euclidean distance between the two vectors; The value ranges from 0 to 1, and the closer the value is to 1, the higher the similarity of the functions of the two products. S22. Based on the co-occurrence frequency of products in historical configuration schemes or orders, generate edges for product collaboration / substitution through point mutual information measurement. Its weight The value ranges from 0 to 1, and the higher the co-occurrence frequency, the greater the weight. S23. Calculate the cosine similarity between the product metadata vector and the scene feature vector. Generate product and applicable scenario edges Set a similarity threshold (Value range: 0.7-0.85), similarity ≥ The edge should be preserved at that time. The graph optimization unit uses a graph neural network to learn the representation of the initially constructed graph, obtains node embedding vectors, completes the potential associated edges, and calculates the graph integrity coefficient. ,like Return to step S21 to re-optimize until... ; Among them, the map integrity coefficient The calculation formula is as follows: ; in: This represents the actual number of associated edges collected. This represents the number of potential associated edges completed using GNN; The theoretical maximum number of associated edges between the three types of core nodes. ; This represents the average edge weights, ranging from 0 to 1. A higher average value indicates a higher weight distribution across all associated edges. The closer to 1; The dynamic map update module is used to continuously optimize the map based on newly added product data. The specific steps are as follows: S24. Set update cycle (Value range 7-30 days), collect metadata vectors of newly added products regularly; S25. Calculate the associated edges between the newly added product node and the existing node, and repeat steps S21-S23 to generate new associated edges. S26. Calculate the weight of the impact of the newly added edge on the existing graph. ,like (This indicates that the edge has a significant impact on the graph topology), which triggers overall graph optimization; S27. Recalculate the map integrity coefficient. ,like If the update is successful, the update is complete; otherwise, return to S25 to adjust the parameters for generating associated edges. S28. Record the graph update log, including update time, number of new nodes / edges, and change in integrity coefficient.

[0027] The cross-domain product knowledge graph construction module achieves structured modeling and dynamic evolution of cross-domain product knowledge through the collaborative linkage of node definition units, edge generation units, graph optimization units, and graph dynamic update modules. The node definition unit clarifies three core node types—product instances, attribute concepts, and application scenarios—and attribute identification rules, laying the foundation for a unified representation of cross-domain knowledge. The edge generation unit accurately constructs three core edge types—functional similarity, collaboration / substitution, and scenario adaptation—based on the Euclidean distance of product metadata vectors, mutual information of historical co-occurrence, and cosine similarity between products and scenarios, addressing the problems of single cross-domain association dimensions and insufficient accuracy in traditional graphs. The graph optimization unit leverages the representation learning capabilities of graph neural networks to complete potential edge connections and uses graph integrity coefficients to further refine the graph. The closed-loop verification ensures the integrity of the graph structure and the reliability of the associations; the graph dynamic update module ensures the timeliness and topological stability of the graph by periodically collecting new data, dynamically calculating associated edges and triggering on-demand optimization.

[0028] It addresses the core pain points of traditional cross-domain knowledge management, such as chaotic nodes, weak connections, and static solidification, providing structured, highly complete, and real-time knowledge support for subsequent functions such as intelligent configuration recommendation and cross-domain product association mining.

[0029] In this embodiment, CNC machine tools from the manufacturing field are selected. Fixed asset management in the financial sector ( Minimally invasive cutting equipment in the medical field Three types of cross-domain products, combined with the metadata vector output by the multimodal product data center ( Conduct map construction and verification: The node definition unit defines Vp1 (attributes include model CM-500, power 15kW, price 280,000 yuan). (Attributes include product code FC-2025, expected return 4.2%, term 3 years) (Attributes include model MC-30, disinfection level S, weight 2.5kg) Three types of product instance nodes are defined, along with attribute concept nodes such as technical parameters, revenue attributes, and applicable departments. Application scenario nodes such as precision machining workshop, enterprise asset allocation, and orthopedic surgery are also defined, and each node is assigned a unique UUID identifier. In the associated edge generation stage, the scale adjustment parameter β=1.2 is used to calculate... and Euclidean distance squared , and Euclidean distance squared Based on the formula for the weight of functionally similar edges Calculated , Based on historical configuration data, Co-occurrence frequency with a certain tooling fixture , , According to the collaborative edge weight formula Calculated , Co-occurrence frequency with a certain insurance product , , Calculated The scene similarity threshold τ is set to 0.75, based on the product-scene similarity formula. Calculated Vector of the "Precision Machining Workshop" Scene , Vector of "Corporate Asset Allocation" , Vector of orthopedic surgical scenarios Preserve the corresponding scene adaptation edges.

[0030] In the graph optimization stage, a GCN graph neural network is used for representation learning and completion. The potential similarity edges with high-precision measuring instruments are statistically obtained. , ,set up , , According to the theoretical formula for the maximum number of associated edges Calculated Mean edge weight Substitute into the formula for the integrity coefficient of the map Calculated Since the value was less than 0.9, we returned to S21, adjusted β to 1.5, and recalculated, finally obtaining... Dynamic update phase settings Every day, collect metadata vectors of newly added smart warehousing equipment, generate association edges between them and existing nodes, and calculate the influence weights. Without triggering overall optimization, recalculation is required. The update is completed and logged. The updated graph can accurately support cross-domain product association recommendation scenarios.

[0031] More specifically, the intelligent product configuration and recommendation engine includes a demand analysis unit, a solution optimization unit, a solution evaluation unit, and a solution screening module; The requirement parsing unit will analyze the user's natural language requirement description. Transform into demand vectors through a pre-trained BERT model. ,calculate With knowledge graph Mid-scene nodes Cosine similarity of embedded vectors is used to pinpoint core application scenarios. The similarity calculation formula is as follows: ; The solution optimization unit models the product configuration problem as a knowledge graph. This problem involves a multi-objective path search problem, under constraints (cost upper limit, performance lower limit, and delivery cycle threshold), generating a set of candidate configuration schemes. The specific steps are as follows: S31. Analyze the constraints in user requirements and convert them into quantitative indicator thresholds (such as cost ≤ 100,000 yuan, response time ≤ 500ms). S32. Starting from the core demand scenario node, search the knowledge graph for product node paths that meet the constraints. The path length is limited to 2-5 nodes. S33. Validate the product combination validity of the searched paths and eliminate combinations with functional conflicts or incompatible parameters; S34. Effective combinations are packaged into candidate configuration schemes, each scheme including a product list, parameter configuration, cost estimate, and delivery cycle forecast. The scheme evaluation unit evaluates each candidate configuration scheme Calculate the overall score The formula is as follows: ; in, This is a matching function between the solution and the requirements, calculated based on the similarity between the aggregated value of the solution product vector and the requirement vector, with a value range of 0-1. The robustness function of the scheme is calculated as follows: , To determine the average degree centrality of the product nodes in the solution, The value of the alternative edge richness of the solution product ranges from 0 to 1. The cost-effectiveness function of the solution is calculated as follows: After normalization, the value ranges from 0 to 1. Candidate solutions The overall performance indicators Candidate solutions Total lifecycle cost; For adjustable weighting coefficients, satisfying It can be dynamically adjusted according to business scenarios; The solution filtering module is used for the final filtering of candidate solutions. The specific steps are as follows: S35. Set the scoring threshold (Value range 0.8-0.9), filter out scores ≥ The solution; S36. If the number of solutions after screening is greater than 5, calculate the similarity between the solutions, remove redundant solutions with a similarity of ≥0.85, and retain no more than 5 differentiated solutions. S37. If the number of solutions after screening is less than 2, lower the scoring threshold by 0.05 and repeat S35-S36 until the number of solutions is greater than or equal to 2. S38. Sort the final selected solutions in descending order of comprehensive score to form a set of recommended solutions and output it.

[0032] The intelligent product configuration and recommendation engine achieves intelligent and precise generation and selection of cross-domain product configuration solutions through the collaborative operation of a demand analysis unit, a solution optimization unit, a solution evaluation unit, and a solution screening module. The demand analysis unit uses a pre-trained BERT model to transform natural language demands into standardized demand vectors, accurately locating core scenarios through cosine similarity, thus solving the problems of vague demand understanding and large scenario matching deviations in traditional methods. The solution optimization unit, based on a knowledge graph, models the configuration problem as a multi-objective path search, combining constraint analysis and combination validity verification to generate candidate solutions containing complete information such as product lists and parameter configurations, breaking down functional and parameter compatibility barriers in cross-domain product combinations. The solution evaluation unit constructs a comprehensive scoring system from three core dimensions: demand matching degree, solution robustness, and cost-effectiveness. Adjustable weight coefficients adapt to different business scenario requirements, ensuring the comprehensiveness and flexibility of solution evaluation. The solution screening module ensures high-quality, differentiated, and quantitatively reasonable output solutions through a closed-loop process of scoring threshold control, redundant solution elimination, and dynamic threshold adjustment.

[0033] In this embodiment, cross-domain scenarios of manufacturing, logistics, and finance are selected to verify the actual effect of intelligent product configuration and recommendation engine: user natural language demand description The goal is to provide small and medium-sized precision parts processing plants with an integrated solution for automated production, intelligent warehousing, and equipment financing, with a total cost of ≤800,000 yuan, a core equipment response time of ≤300ms, and a delivery cycle of ≤90 days. The requirement parsing unit uses a pre-trained BERT model to... Convert to demand vector ,calculate With knowledge graph Nodes in medium-precision machining automation scenarios The cosine similarity of the embedded vectors is calculated using the formula... have to The node is identified as the core demand scenario; the constraints of the solution optimization unit are quantifiable indicator thresholds (total cost ≤ 800,000 yuan, response time ≤ 300ms, delivery cycle ≤ 90 days). Starting from the knowledge graph, a product node path with a path length of 3 is searched, resulting in 6 candidate paths. After validity verification, 2 paths with incompatible parameters are removed (e.g., the load-bearing capacity of a high-speed processing equipment does not match that of a regular warehouse rack). The remaining 4 valid paths are encapsulated into candidate configuration schemes. , , , Each plan includes a product list (such as...) Includes CNC precision machine tools, intelligent automated shelving, and equipment financing and leasing services; parameter configuration, cost estimation, and delivery cycle forecasting; adjustable weighting coefficients are set for the solution evaluation unit. , , Calculate the overall score for each option: by For example, (Solution product vector aggregation value and) (similarity) ( For the average degree centrality of product nodes, (to replace edge richness) ( For comprehensive performance indicators, The total cost of the entire life cycle is 10,000 yuan (normalized). Substituting this into the scoring formula yields... .

[0034] The solution selection module sets a scoring threshold. Filter out (0.901) (0.887) (0.862) (0.855) Four qualified solutions, with the number of solutions ranging from 2 to 5, do not require redundancy removal or threshold adjustment. Finally, the solutions are sorted in descending order of comprehensive score to form a recommended solution set and output. This solution set can directly meet the user's cross-domain integrated configuration needs.

[0035] More specifically, the product lifecycle tracking and feedback module includes a data acquisition unit, a performance calculation unit, a deviation analysis unit, and a feedback data processing module; The data acquisition unit continuously collects product operation status data (failure rate, response speed), performance indicators (throughput, accuracy) and user satisfaction scores through sensors, system log interfaces, user feedback questionnaires, etc. The sampling frequency is 1 time / hour to 1 time / day, which can be dynamically adjusted according to product type. The performance computing unit is configured for each implementation. Create tracking instances to record performance metric vectors. Over time Changes, calculating actual comprehensive effectiveness The formula is as follows: ; in, The observation period (unit: days) ranges from 30 to 180 days, depending on the product lifecycle length. The number of performance indicators ranges from 5 to 15, covering dimensions such as operating status, performance, and user experience. As the indicator weight, satisfying This was determined using the analytic hierarchy process (AHP). For the first Each indicator in time The values ​​have been standardized to the range of 0-1; The deviation analysis unit calculates the expected overall effectiveness based on the prediction model generated during scheme generation. And solve for the performance deviation. : ; Set deviation threshold (Value range -0.15~0.15), when When this happens, mark the solution as one that requires priority optimization; The feedback data processing module is used to clean and classify the collected feedback data. The specific steps are as follows: S41. Perform outlier detection on numerical feedback data (such as failure rate and response time) and use the 3σ criterion to remove outlier data. S42. Perform sentiment analysis and keyword extraction on text-based feedback data (such as user reviews), and classify it into positive feedback, negative feedback, and neutral feedback; S43. Map negative feedback to corresponding performance indicators and generate a correlation table between feedback and indicators; S44. Calculate the percentage of negative feedback for each performance indicator. ,like Mark this indicator as an indicator that needs to be monitored closely; S45. Compare the processed feedback data with the performance deviation. The data is bound together to form a complete feedback data packet and transmitted to the system's adaptive evolution module.

[0036] The product lifecycle tracking and feedback module, through the coordinated operation of the data acquisition unit, performance calculation unit, deviation analysis unit, and feedback data processing module, achieves full-process dynamic tracking, performance evaluation, and accurate feedback after the implementation of cross-domain configuration solutions. The data acquisition unit relies on multi-channel collection methods (sensors, system logs, user questionnaires) and a dynamic sampling frequency mechanism to comprehensively cover multi-dimensional data such as product operation, performance, and user experience, solving the problems of incomplete and untimely traditional tracking data. The performance calculation unit constructs performance indicator vectors of tracking instances and time dimensions, and calculates the actual comprehensive performance by combining the weights determined by the analytic hierarchy process, realizing a quantitative evaluation of solution performance. The deviation analysis unit solves the deviation by comparing the actual and expected comprehensive performance, and accurately locates the solutions that need to be optimized by combining the deviation threshold, filling the gap of traditional configuration lacking performance verification. The feedback data processing module uses outlier detection, sentiment analysis, indicator association mapping and other processing processes to explore the correlation between negative feedback and performance indicators, mark key monitoring indicators and form a complete feedback data package. A closed-loop link of configuration implementation, tracking evaluation, deviation identification and feedback optimization is constructed, which not only provides data support for the iterative optimization of existing solutions, but also delivers core feedback data to the system adaptive evolution module, ensuring the continuous optimization capability of the entire management system.

[0037] In this embodiment, the recommendation scheme is based on the intelligent product configuration and the output of the recommendation engine. (CNC precision machine tools, intelligent automated shelving, and equipment financing and leasing services) are the tracking targets, and full lifecycle tracking and feedback verification are carried out: The data acquisition unit collects failure rate and response speed through the machine tool's built-in sensors, throughput and inventory accuracy through the warehouse system log interface, and user satisfaction ratings through online questionnaires, with a sampling frequency set at 1 time / 6 hours (adapted to the operating characteristics of industrial equipment); the performance calculation unit is... Create a tracking instance and set the observation period. On that day, five performance indicators were selected: failure rate, response speed, throughput, accuracy, and user satisfaction. The weights of the indicators are determined using the analytic hierarchy process. (Failure rate) (Response speed) (throughput) (Accuracy) (User satisfaction), record the standardized values ​​of each indicator over time t. Substituting into the actual comprehensive efficiency formula: Calculated ; The expected overall performance of the deviation analysis unit when generating the scheme Solving for performance bias Set deviation threshold , Marked as a standard monitoring solution; during the feedback data processing phase, the following is adopted: The criteria removed 12 sets of abnormal failure rate data. Sentiment analysis was performed on the 80 collected user reviews, extracting positive keywords such as fast response and accurate inventory, and neutral keywords such as cumbersome questionnaire operation. There was no negative feedback. The percentage of negative feedback for each performance indicator was also analyzed. (All < 0.2), no need to mark key monitoring indicators, and finally the processed feedback data will be compared with... The data is bound together and a feedback data packet is transmitted to the system's adaptive evolution module.

[0038] More specifically, the system adaptive evolution module includes a graph update unit, a strategy optimization unit, an iterative verification unit, and a decision strategy update module; The map update unit is based on performance deviation. Dynamically adjust cross-domain product knowledge graph confidence of related edges The formula is as follows: ; in, This represents the original confidence level of the associated edge, with a value ranging from 0 to 1. The updated confidence level must satisfy 0 ≤ ≤1, if the value is outside the range, the boundary value is used; The learning rate, ranging from 0.05 to 0.15, controls the adjustment range of the confidence level. For a sign function, when hour ,when hour ,when hour ; For the edge With the plan The contribution relevance, with a value ranging from 0 to 1, is calculated by the importance of the edge in the solution generation path; The strategy optimization unit models the solution generation and selection process as a reinforcement learning problem, using knowledge graphs. The environmental state is the configuration scheme as the action, and the actual comprehensive effectiveness is... As a reward, the decision network parameters of the recommendation engine are optimized using the policy gradient method; The iterative verification unit collects performance data of newly added configuration schemes every quarter to verify the updated graph and optimized strategy. If the strategy accuracy is ≥95% and the graph integrity coefficient is [not specified], the verification is successful. If the result is positive, the current iteration is complete; otherwise, return to the graph update unit for re-optimization. The decision strategy update module is used to dynamically update the decision rules of the recommendation engine. The specific steps are as follows: S51. Calculate the fit of the current decision-making strategy with the recent configuration scheme. Value range 0-1, fit , This refers to the number of recent proposals; S52, Set the fit threshold (Value range 0.85-0.9), if This triggers an update to the decision-making strategy; S53. Confirm the weight adjustment requirements for each performance indicator as a component of the decision-making strategy. The weight adjustment range is ≤0.1; S54. Confirm the direction of similarity threshold optimization for scheme selection, as a component of the decision-making strategy. The optimization range is ±0.05; S55. Perform constraint condition extension requirement judgment and output the judgment result; S56. When the judgment result indicates that expansion is needed, add new constraint types and thresholds as components of the decision-making strategy. ; S57. When the judgment result of S55 is "necessary", the constituent elements will be... Integrate into updated decision-making strategies; when the judgment result is that it is unnecessary, the constituent elements will be... Integrate into updated decision-making strategies; S58. Replace the current decision strategy with the updated decision strategy, and record the strategy update log, including update time, changes in adaptability, and adjustments to the constituent elements.

[0039] The system's adaptive evolution module, through the coordinated operation of the graph update unit, strategy optimization unit, iterative verification unit, and decision strategy update module, constructs a closed-loop evolution mechanism for system feedback, optimization, and verification: The graph update unit dynamically adjusts the confidence of knowledge graph association edges based on performance deviations, ensuring accurate matching between graph associations and actual application effects; the strategy optimization unit, with reinforcement learning at its core, models the configuration scheme generation process as a sequential decision problem, continuously optimizing recommendation engine decision parameters through actual performance feedback; the iterative verification unit ensures the reliability of optimization results through quarterly performance data collection and dual-indicator verification (strategy accuracy and graph integrity coefficient); and the decision strategy update module dynamically triggers strategy updates based on recent scheme adaptability, integrating three elements—weight adjustment, threshold optimization, and constraint expansion—to achieve flexible iteration of decision rules. This addresses the pain points of traditional systems' static and fixed decision strategies, which cannot adapt to business changes, enabling the system to continuously absorb the value of actual application data, achieve adaptive optimization of the knowledge graph and decision strategies, and maintain efficient cross-domain product management capabilities in the long term.

[0040] In this embodiment, a smart product configuration scheme is used. Based on feedback data from (CNC precision machine tools, intelligent automated shelving, and equipment financing and leasing services), the actual effectiveness of the system's adaptive evolution module is verified; (Known) Performance deviation Select the associated edge between CNC precision machine tool and intelligent automated storage and retrieval system in the path generated by this scheme. Its original confidence level Set the learning rate The importance of the edge in the path is calculated. ,because Sign function Substitute into the confidence update formula Calculated (If the value is in the 0-1 range, retain it directly); the strategy optimization unit will use the knowledge graph As an environmental state, the configuration scheme is the action. Actual comprehensive effectiveness As a reward, the decision network parameters of the recommendation engine are optimized through the policy gradient method to improve the accuracy of solution generation in similar scenarios; The iterative verification unit collects performance data for 100 newly added configuration schemes every quarter, verifying a strategy accuracy of 96.2% ≥ 95% and a graph integrity coefficient. The iteration completion conditions are met; in the decision strategy update phase, the 20 most recent configuration schemes are selected. ), calculate the fit Set the fit threshold ,because Trigger update; The need to adjust the weight of performance indicators (increasing the weight of the failure rate indicator from 0.25 to 0.3, with an adjustment range of 0.05 ≤ 0.1) has been confirmed as a component element. The similarity threshold for scheme selection was optimized from 0.85 to 0.8 (optimization range -0.05) as a component. After assessment, it was determined that no further constraint extension was required. and The system integrates and updates decision-making strategies, replaces the current strategy, and records logs. After the update, the system's adaptation accuracy for similar cross-domain configuration requirements is improved.

[0041] More specifically, the demand parsing unit of the intelligent product configuration and recommendation engine also supports multilingual demand input. It automatically identifies the input language type (Chinese, English, Japanese, etc.) through a language detection model, calls the pre-trained model of the corresponding language for vector conversion, and achieves a language recognition accuracy of ≥98% and a demand matching degree loss of ≤5% after vector conversion.

[0042] The intelligent product configuration and recommendation engine's requirement parsing unit integrates a language detection model and a multilingual pre-trained model to achieve accurate identification and efficient conversion of multilingual requirement inputs. It ensures accurate judgment of the input language type with a language recognition accuracy of ≥98% and guarantees the consistency of cross-language semantic transmission with a requirement matching loss of ≤5%. This breaks through the limitations of traditional single-language input systems and solves the pain points of large deviations in cross-language requirement understanding and serious semantic loss. It enables the system to adapt to the cross-domain product configuration needs of users in different regions around the world, significantly improving the system's universality, international adaptability, and user experience.

[0043] More specifically, a data-driven cross-domain product information management system also includes a cross-domain product data governance dashboard; The dashboard visually displays the data source access status (number of connections, online rate) and data quality score (based on) of the multimodal product data center. Calculation, real-time updates), cross-domain product knowledge graph The topology and edge confidence distribution, the decision accuracy trend of the intelligent recommendation engine, and the performance change curve throughout the product lifecycle; The dashboard allows administrators to manually adjust core system parameters and thresholds. The specific steps are as follows: S61. The administrator selects the type of parameter to be adjusted (data preprocessing parameters, map construction parameters, recommendation engine parameters). S62. Input a new parameter value, and the system will automatically check whether the parameter value is within a reasonable range. S63. If the parameter value is reasonable, generate a parameter adjustment preview to show the possible impact of the adjustment on system performance. S64. After the administrator confirms the adjustment, the system updates the parameters and records the adjustment log; S65. If the parameter value is unreasonable, the system will prompt a reasonable range of values ​​and return to S62 to re-enter.

[0044] The cross-domain product data governance dashboard achieves transparent management of system operation status and efficient implementation of operation and maintenance operations through the dual functions of visualization and closed-loop parameter adjustment. The visualization function comprehensively covers core dimensions such as data source access, data quality, knowledge graph structure, recommendation engine performance, and product effectiveness, presenting key data (such as real-time data quality score and edge confidence distribution of Qd calculation) in an intuitive form, solving the pain points of opaque system operation status and difficulty in locating problems in traditional systems. The parameter adjustment function allows administrators to accurately select parameter types. Through a closed-loop process of automatic system verification, adjustment preview, and log recording, it ensures the rationality and traceability of parameter modifications. It can adapt to changes in business needs without modifying core code, reducing the operation and maintenance threshold and operational risks, while improving the system's flexibility in responding to business adjustments.

[0045] In this embodiment, a cross-domain management scenario involving manufacturing and finance is selected to verify the application effect of the cross-domain product data governance dashboard. The dashboard uses a visual interface to display the multimodal product data center's access to three types of data sources in the manufacturing domain (equipment parameter tables, production logs, and product drawings) and two types of data sources in the financial domain (financial product brochures and financing service agreements). The online rate of all data sources is 99.8%. Based on the data quality optimization coefficient... Real-time calculation and display of data quality scores for numerical equipment parameters in the manufacturing field. Financial field text-based instruction manuals Cross-domain product knowledge graph The topology is presented using a force-directed graph, with the edge confidence of core product nodes (such as CNC machine tools and fixed asset investment) and scenario nodes ranging from 0.75 to 0.93. The intelligent recommendation engine's decision accuracy trend curve over the past three months shows a steady increase in accuracy from 92% to 96.5%. Configuration schemes... The life-cycle efficiency change curve of (CNC machine tools and equipment financial leasing) shows that the actual comprehensive efficiency is stable in the range of 0.87-0.89 within a 90-day observation period.

[0046] To further improve data quality, the administrator performs the following steps: S61 Select data preprocessing parameters as the parameter type; S62 Input new... The threshold is 0.88, and the system automatically verifies that this value is within a reasonable range (0.8-0.95); S63 generates a preview prompt. After raising the threshold to 0.88, the data preprocessing pass rate is expected to decrease by 3%, but the average data quality will improve by 5%. The S64 administrator confirms the adjustment, the system updates the parameters and records the log (including update time, parameter type, original value 0.85, new value 0.88, and estimated impact). The dashboard updates in real time after the adjustment, supporting numerical data in the manufacturing sector. Improved to 0.94, for text-based data in the financial sector The accuracy has been improved to 0.91, significantly enhancing the reliability of system data.

[0047] More specifically, the data-driven cross-domain product information management system adopts a microservice architecture deployment. The multimodal product data center, the cross-domain product knowledge graph construction module, and the intelligent product configuration and recommendation engine are each deployed as independent scalable microservices, supporting elastic scaling up and down based on business load. The data-driven, cross-domain product information management system supports the configuration of adaptive encoders for different domains. Based on the domain ontology rules, new product domains (such as new energy, artificial intelligence, and the Internet of Things) can be quickly integrated without changing the core architecture. The process for integrating new domains is as follows: S71. Configure the data source interface parameters for the new field and connect to the data source of the new field product. S72, Training Adaptive Encoders for New Domains Generate a domain-specific coding network; S73. Import ontology rules for new domains (product attribute classification, scenario definition specifications); S74. Collect seed product data (no less than 500 records) in new fields and initialize the knowledge graph; S75. Verify the accuracy of the configuration recommendation for the new product. If the accuracy is ≥90%, the integration is complete; otherwise, return to S72 to optimize the encoder. The access period for new domains is ≤15 working days. After access, cross-domain configuration recommendations with existing domain products are supported.

[0048] The system adopts a microservice architecture that breaks down the core module into independent scalable units, enabling elastic scaling up and down based on business load. This effectively solves the pain points of low resource utilization and easy lag in high-load scenarios of traditional monolithic architectures. Through the flexible extension mechanism of the configuration domain adaptive encoder and ontology rules, new domains can be quickly accessed without changing the core architecture. Combined with standardized access processes and accuracy verification mechanisms, the system ensures the efficiency (cycle ≤ 15 working days) and reliability (configuration recommendation accuracy ≥ 90%) of new domain access. At the same time, it supports cross-domain configuration recommendations for new and old domains, completely breaking the limitations of traditional systems that are "domain-fixed and difficult to extend", and significantly improving the system's scalability, resource adaptability and business coverage.

[0049] In this embodiment, the new energy field is selected as the new access field to verify the system's scalability and the elastic adaptation effect of the microservice architecture. At the microservice deployment level, during peak business periods of cross-domain scenarios in manufacturing and finance (such as a surge in configuration demand at the end of the quarter), the system detects that the load rate of the intelligent product configuration and recommendation engine reaches 85%, automatically triggering elastic expansion to increase the number of microservice instances from 3 to 6. After expansion, the response latency is reduced from 200ms to 80ms. During business off-peak periods, when the load rate drops to 30%, the system automatically shrinks to 2 instances, improving resource utilization by 40%. In the new field access phase, the following process is executed: S71 Configure the data source interface parameters for the new energy field, connecting to three types of data sources, including battery production parameter tables and photovoltaic module performance reports; S72 Train an adaptive encoder based on 600 product data points from the new energy field. The system generates a domain-specific CNN+Transformer encoding network; S73 imports ontology rules from the new energy domain (including product attribute classifications such as battery capacity and photovoltaic conversion efficiency, and definition specifications for distributed generation scenarios and energy storage scenarios); S74 collects 600 new energy seed product data (meeting the requirement of ≥500 data points) to complete the initialization of the new energy branch of the cross-domain product knowledge graph; S75 verifies the accuracy of cross-domain configuration recommendations between new energy products and existing manufacturing equipment, achieving an accuracy rate of ≥90% (93%), thus completing the integration. The entire integration cycle is 12 working days (≤15 working days). After integration, the system can accurately generate cross-domain configuration solutions for new energy storage equipment, manufacturing workshop power supply schemes, and equipment financing services, meeting users' integrated needs.

[0050] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data-driven, cross-domain product information management system, characterized in that, It includes a multimodal product data center, a cross-domain product knowledge graph construction module, an intelligent product configuration and recommendation engine, a product lifecycle tracking and feedback module, and a system adaptive evolution module; The multimodal product data center is used to connect to heterogeneous product data sources from different business domains, and generates a standardized product metadata vector set through standardized cleaning, feature extraction, and unified semantic encoding. ; The cross-domain product knowledge graph construction module is based on product metadata vector sets. It automatically mines product attribute associations, functional substitution / complementary relationships between products, and product-application scenario mapping relationships, and constructs and dynamically updates a unified cross-domain product knowledge graph. ; The intelligent product configuration and recommendation engine receives requirement descriptions from users or upper-layer applications and is based on a cross-domain product knowledge graph. Perform semantic analysis and knowledge retrieval of requirements to generate a set of multi-objective optimized product configuration solutions. And complete the feasibility assessment and benefit forecast; After the configuration scheme is implemented, the product lifecycle tracking and feedback module continuously collects product operation status data, performance indicators, and user feedback to form a performance tracking record for the product instance. ; The system's adaptive evolution module is based on performance tracking records. Analyze the discrepancy between expected and actual performance of configuration schemes to drive cross-domain product knowledge graphs. The confidence level of the association is updated to optimize the decision-making strategy of the intelligent product configuration and recommendation engine.

2. The data-driven cross-domain product information management system according to claim 1, characterized in that, The multimodal product data center includes a data access unit, an intelligent preprocessing unit, a semantic encoding unit, and a vector fusion unit; The data access unit connects to product data sources from different fields through standardized interfaces, including structured data, semi-structured data, and unstructured data, to form the original product dataset. The intelligent preprocessing unit performs AI-driven preprocessing operations on the original product dataset. The specific steps are as follows: S11. Parse the original product dataset to distinguish between numerical, textual, and image-based product data; S12. Use the Z-score standardization method to eliminate the influence of dimensions for numerical data and calculate the standardization error. ; S13. Use the TF-IDF algorithm to extract keyword features from text data and filter out low-frequency meaningless words; S14. Use an adaptive threshold segmentation method to extract contour features from image data and calculate the contour sharpness coefficient. ; S15. Substitute into the formula to calculate the data quality optimization coefficient. ,like Return to step S12 to re-perform the standardization process until... ; Among them, the data quality optimization coefficient The calculation formula is as follows: ； in, This is the standardization error, ranging from 0 to 1. The smaller the error, the better. The closer a value is to 0, the greater the consistency of the standardized numerical data. This is the outline sharpness coefficient, ranging from 0 to 1. The sharper the outline edges and the less noise, the better. The closer to 1; This is a data preprocessing correction constant, ranging from 0.03 to 0.08, used to compensate for the systematic bias of different types of data preprocessing algorithms; The semantic coding unit employs a domain-adaptive encoder to encode the first... The first field Features of each product after pretreatment Mapping to a unified semantic space yields standardized product metadata vectors. The mapping process is as follows: ； in, For the target field The trainable encoding network adopts a hybrid architecture of CNN+Transformer; These are the learnable parameters of the encoding network; A common semantic vector shared across all domains, with dimensions of 128-512, is used to ensure semantic consistency of features across domains; The vector fusion unit is used to perform weighted fusion of semantic vectors from multiple domains. The specific steps are as follows: S16. Calculate the reliability score of the product's data sources in various relevant fields. The value ranges from 0 to 1; the higher the data integrity, the better. The closer to 1; S17. Standardized product metadata vectors for various fields Assign weights ,satisfy The weight calculation formula is: ; S18. Obtain the final product metadata vector through weighted summation. .

3. The data-driven cross-domain product information management system according to claim 2, characterized in that, The cross-domain product knowledge graph construction module includes a node definition unit, an edge generation unit, a graph optimization unit, and a graph dynamic update module. The node definition unit defines three types of core nodes: product instance nodes. Attribute concept nodes Application scenario nodes Each node is assigned a unique identifier and basic attribute label. The association edge generation unit constructs three types of core edges based on product metadata vectors, and the specific steps are as follows: S21. Calculate the product metadata vector. and Using Euclidean distance, generate edges for products with similar functions. Its weight The calculation is as follows: ； in, This is a scaling parameter, with a value range of 0.5-2.

0. The larger the distance, the faster the similarity decreases with distance; The square of the Euclidean distance between the two vectors; The value ranges from 0 to 1, and the closer the value is to 1, the higher the similarity of the functions of the two products. S22. Based on the co-occurrence frequency of products in historical configuration schemes or orders, generate edges for product collaboration / substitution through point mutual information measurement. Its weight The value ranges from 0 to 1, and the higher the co-occurrence frequency, the greater the weight. S23. Calculate the cosine similarity between the product metadata vector and the scene feature vector. Generate product and applicable scenario edges Set a similarity threshold (Value range: 0.7-0.85), similarity ≥ The edge should be preserved at that time. The graph optimization unit uses a graph neural network to learn the representation of the initially constructed graph, obtains node embedding vectors, completes the potential associated edges, and calculates the graph integrity coefficient. ,like Return to step S21 to re-optimize until... ; Among them, the map integrity coefficient The calculation formula is as follows: ； in: This represents the actual number of associated edges collected. This represents the number of potential associated edges completed using GNN; The theoretical maximum number of associated edges between the three types of core nodes. ; This represents the average edge weights, ranging from 0 to 1. A higher average value indicates a higher weight distribution across all associated edges. The closer to 1; The dynamic map update module is used to continuously optimize the map based on newly added product data. The specific steps are as follows: S24. Set update cycle Regularly collect metadata vectors for newly added products; S25. Calculate the associated edges between the newly added product node and the existing node, and repeat steps S21-S23 to generate new associated edges. S26. Calculate the weight of the impact of the newly added edge on the existing graph. ,like This will trigger an overall optimization of the graph; S27. Recalculate the map integrity coefficient. ,like If the update is successful, the update is complete; otherwise, return to S25 to adjust the parameters for generating associated edges. S28. Record the graph update log, including update time, number of new nodes / edges, and change in integrity coefficient.

4. The data-driven cross-domain product information management system according to claim 3, characterized in that, The intelligent product configuration and recommendation engine includes a demand analysis unit, a solution optimization unit, a solution evaluation unit, and a solution screening module; The requirement parsing unit will analyze the user's natural language requirement description. Transform into demand vectors through a pre-trained BERT model. ,calculate With knowledge graph Mid-scene nodes Cosine similarity of embedded vectors is used to pinpoint core application scenarios. The similarity calculation formula is as follows: ; The solution optimization unit models the product configuration problem as a knowledge graph. The problem of multi-objective path search on [a specific target], under constraints, generates a set of candidate configuration schemes. The specific steps are as follows: S31. Analyze the constraints in user requirements and convert them into quantitative indicator thresholds; S32. Starting from the core demand scenario node, search the knowledge graph for product node paths that meet the constraints. The path length is limited to 2-5 nodes. S33. Validate the product combination validity of the searched paths and eliminate combinations with functional conflicts or incompatible parameters; S34. Effective combinations are packaged into candidate configuration schemes, each scheme including a product list, parameter configuration, cost estimate, and delivery cycle forecast. The scheme evaluation unit evaluates each candidate configuration scheme Calculate the overall score The formula is as follows: ； in, This is a matching function between the solution and the requirements, calculated based on the similarity between the aggregated value of the solution product vector and the requirement vector, with a value range of 0-1. The robustness function of the scheme is calculated as follows: , To determine the average degree centrality of the product nodes in the solution, The value of the alternative edge richness of the solution product ranges from 0 to 1. The cost-effectiveness function of the solution is calculated as follows: After normalization, the value ranges from 0 to 1. Candidate solutions The overall performance indicators Candidate solutions Total lifecycle cost; For adjustable weighting coefficients, satisfying It can be dynamically adjusted according to business scenarios; The scheme screening module is used for the final screening of candidate schemes, and the specific steps are as follows: S35. Set the scoring threshold (Value range 0.8-0.9), filter out scores ≥ The solution; S36. If the number of solutions after screening is greater than 5, calculate the similarity between the solutions, remove redundant solutions with a similarity of ≥0.85, and retain no more than 5 differentiated solutions. S37. If the number of solutions after screening is less than 2, lower the scoring threshold by 0.05 and repeat S35-S36 until the number of solutions is greater than or equal to 2. S38. Sort the final selected solutions in descending order of comprehensive score to form a set of recommended solutions and output it.

5. A data-driven cross-domain product information management system according to claim 4, characterized in that, The product lifecycle tracking and feedback module includes a data acquisition unit, a performance calculation unit, a deviation analysis unit, and a feedback data processing module. The data acquisition unit continuously collects product operating status data, performance indicators, and user satisfaction scores through sensors, system log interfaces, user feedback questionnaires, etc. The sampling frequency is 1 time / hour to 1 time / day, which can be dynamically adjusted according to the product type. The performance calculation unit is configured for each implementation. Create tracking instances to record performance metric vectors. Over time Changes, calculating actual comprehensive effectiveness The formula is as follows: ； in, The observation period is 30-180 days, depending on the product lifecycle length. The number of performance indicators ranges from 5 to 15, covering dimensions such as operating status, performance, and user experience. As the indicator weight, satisfying This was determined using the analytic hierarchy process (AHP). For the first Each indicator in time The values ​​have been standardized to the range of 0-1; The deviation analysis unit calculates the expected overall performance based on the prediction model generated during scheme generation. And solve for the performance deviation. : ； Set deviation threshold ,when When this happens, mark the solution as one that requires priority optimization; The feedback data processing module is used to clean and classify the collected feedback data. The specific steps are as follows: S41. Perform outlier detection on numerical feedback data and use the 3σ criterion to remove outlier data. S42. Perform sentiment analysis and keyword extraction on text-based feedback data, and classify it into positive feedback, negative feedback, and neutral feedback. S43. Map negative feedback to corresponding performance indicators and generate a correlation table between feedback and indicators; S44. Calculate the percentage of negative feedback for each performance indicator. ,like Mark this indicator as an indicator that needs to be monitored closely; S45. Compare the processed feedback data with the performance deviation. The data is bound together to form a complete feedback data packet and transmitted to the system's adaptive evolution module.

6. The data-driven cross-domain product information management system according to claim 5, characterized in that, The system adaptive evolution module includes a graph update unit, a strategy optimization unit, an iterative verification unit, and a decision strategy update module; The map update unit is based on performance deviation. Dynamically adjust cross-domain product knowledge graph confidence of related edges The formula is as follows: ； in, This represents the original confidence level of the associated edge, with a value ranging from 0 to 1. The updated confidence level must satisfy 0 ≤ ≤1, if the value is outside the range, the boundary value is used; The learning rate, ranging from 0.05 to 0.15, controls the adjustment range of the confidence level. For a sign function, when hour ,when hour ,when hour ; For the edge With the plan The contribution relevance, with a value ranging from 0 to 1, is calculated by the importance of the edge in the solution generation path; The policy optimization unit models the scheme generation and selection process as a reinforcement learning problem, using a knowledge graph. The environmental state is the configuration scheme as the action, and the actual comprehensive effectiveness is... As a reward, the decision network parameters of the recommendation engine are optimized using the policy gradient method; The iterative verification unit collects performance data of newly added configuration schemes every quarter to verify the updated graph and optimized strategy. If the strategy accuracy is ≥95% and the graph integrity coefficient is [not specified], the verification is successful. If the result is positive, the current iteration is complete; otherwise, return to the graph update unit for re-optimization. The decision strategy update module is used to dynamically update the decision rules of the recommendation engine. The specific steps are as follows: S51. Calculate the fit of the current decision-making strategy with the recent configuration scheme. Value range 0-1, fit , This refers to the number of recent proposals; S52, Set the fit threshold ,like This triggers an update to the decision-making strategy; S53. Confirm the weight adjustment requirements for each performance indicator as a component of the decision-making strategy. The weight adjustment range is ≤0.1; S54. Confirm the direction of similarity threshold optimization for scheme selection, as a component of the decision-making strategy. The optimization range is ±0.05; S55. Perform constraint condition extension requirement judgment and output the judgment result; S56. When the judgment result indicates that expansion is needed, add new constraint types and thresholds as components of the decision-making strategy. ; S57. When the judgment result of S55 is "necessary", the constituent elements will be... Integrate into updated decision-making strategies; when the judgment result is that it is unnecessary, the constituent elements will be... Integrate into updated decision-making strategies; S58. Replace the current decision strategy with the updated decision strategy, and record the strategy update log, including update time, changes in adaptability, and adjustments to the constituent elements.

7. A data-driven cross-domain product information management system according to claim 6, characterized in that, The intelligent product configuration and recommendation engine's requirement parsing unit also supports multilingual requirement input. It automatically identifies the input language type through a language detection model, calls the corresponding pre-trained model for vector conversion, and achieves a language recognition accuracy of ≥98% and a requirement matching degree loss of ≤5% after vector conversion.

8. A data-driven cross-domain product information management system according to claim 7, characterized in that, The data-driven cross-domain product information management system also includes a cross-domain product data governance dashboard; The dashboard visually displays the data source access status, data quality score, and cross-domain product knowledge graph of the multimodal product data center. The topology and edge confidence distribution, the decision accuracy trend of the intelligent recommendation engine, and the performance change curve throughout the product lifecycle; The dashboard allows administrators to manually adjust core system parameters and thresholds. The specific steps are as follows: S61. The administrator selects the type of parameter to be adjusted; S62. Input a new parameter value, and the system will automatically check whether the parameter value is within a reasonable range. S63. If the parameter value is reasonable, generate a parameter adjustment preview to show the possible impact of the adjustment on system performance. S64. After the administrator confirms the adjustment, the system updates the parameters and records the adjustment log; S65. If the parameter value is unreasonable, the system will prompt a reasonable range of values ​​and return to S62 to re-enter.

9. A data-driven cross-domain product information management system according to claim 8, characterized in that, The data-driven cross-domain product information management system is deployed using a microservice architecture. The multimodal product data center, the cross-domain product knowledge graph construction module, and the intelligent product configuration and recommendation engine are each deployed as independent scalable microservices, supporting elastic scaling up and down based on business load. The data-driven, cross-domain product information management system supports the configuration of adaptive encoders for different domains. Based on the domain ontology rules, new product domains can be quickly integrated without changing the core architecture. The new domain integration process is as follows: S71. Configure the data source interface parameters for the new field and connect to the data source of the new field product. S72, Training Adaptive Encoders for New Domains Generate a domain-specific coding network; S73. Import ontology rules for new domains; S74. Collect seed product data in new fields and initialize the knowledge graph; S75. Verify the accuracy of the configuration recommendation for the new product. If the accuracy is ≥90%, the integration is complete; otherwise, return to S72 to optimize the encoder. The access period for new domains is ≤15 working days. After access, cross-domain configuration recommendations with existing domain products are supported.

Images