A method for enterprise product compliance review and risk early warning

Abstract

The application discloses a kind of enterprise product compliance review and risk early warning method, specifically includes the following steps: S1, constructs dynamic standard compliance knowledge graph Standard-KG, S2, enterprise product specification data multimodal semantic structure, S3, double-track product compliance automatic review, S4, standard iteration atlas propagation risk early warning, S5, compliance reverse auxiliary design and supplier risk portrait generation, the present application relates to artificial intelligence knowledge engineering technical field.The enterprise product compliance review and risk early warning method pass through external standard document, enterprise internal product data two-way link, rely on big model and knowledge graph to realize whole process automatic analysis, comparison, check;Original artificial needs several days to several weeks to complete the whole product line compliance investigation, system can output complete review report in only several minutes, greatly reduce compliance post manpower investment, eliminate the problem of wrong detection, missed detection caused by artificial fatigue, knowledge blind area.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence knowledge engineering technology, specifically a method for enterprise product compliance review and risk warning. Background Technology

[0002] Product compliance management for manufacturing enterprises spans the entire lifecycle of R&D, production, and supply. Currently, the mainstream compliance review model in the industry relies heavily on manual operations, and suffers from three core pain points: data fragmentation, inability to automatically resolve complex constraints, and delayed response to risks associated with updated standards. Specific shortcomings are as follows: 1. Product data and industry standards form data silos, leading to inefficient manual review and frequent errors. While internal PLM, ERP, and BOM systems store complete product technical parameters and component specifications, national standards, industry standards, and mandatory regulations are mostly stored as unstructured documents such as PDFs, scanned paper documents, and images. These two data systems are completely isolated. In existing compliance processes, quality and compliance engineers must manually review hundreds of pages of standard texts, checking each indicator against product specifications and test reports. This model consumes a significant amount of time, and long-term manual verification is susceptible to fatigue and limitations in individual knowledge, resulting in missed indicators and mismatched standard clauses, directly creating the potential for product non-compliance and hindering market launch.

[0003] 2. Unstructured constraints with multiple coupled conditions within standards cannot be automatically parsed, resulting in extremely low system review coverage. Current national standards and industry specifications contain numerous composite constraints with preconditions, such as "under a test environment of 25℃ and 60% relative humidity, the insulation resistance of the equipment shall not be less than 5MΩ" and "for equipment used in outdoor high salt spray conditions, the outer casing must have a corrosion-resistant protective coating." Existing compliance verification systems on the market rely solely on keyword matching and simple rule engines to complete text retrieval. They cannot identify preconditions and multi-variable coupled logic within clauses, nor can they convert constraints described in natural language into machine-executable computational logic. They can only perform simple numerical comparisons, resulting in an automation rate of less than 20% for compliance reviews in complex scenarios.

[0004] 3. The lack of an automatic traceability mechanism for standard version iterations leads to a significant lag in compliance risks. National and industry standards continuously release new versions and amendments, with frequent adjustments to mandatory indicator thresholds, scope of application, and exemption clauses. Existing enterprise management systems lack the ability to automatically detect and analyze the impact of standard changes. After a standard is updated, compliance issues can only be identified through manual review of all product categories, including those under development, in mass production, and discontinued products. This often results in non-compliance issues being discovered only after mass production and regulatory inspections, which can easily lead to large-scale product recalls, administrative penalties, and commercial breaches of contract, resulting in substantial economic losses. Furthermore, supply chain risks related to component supply cannot be tracked synchronously, and compliance risks to upstream suppliers arising from standard changes cannot be predicted in advance.

[0005] This invention constructs a computable "standard compliance knowledge base," transforming unstructured standard requirements into machine-readable constraints. It overcomes the semantic mapping bottleneck between product specifications and standard clauses, achieving automated alignment and conflict detection between enterprise product parameters (BOM, specifications) and standard compliance baselines. A dynamic standard change response and risk warning model is established; when external standard libraries are updated, the system can automatically reverse-engineer to the enterprise product library, accurately locating affected components or product models, thus shifting from "passive response" to "proactive early warning" compliance risk management. Summary of the Invention

[0006] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for enterprise product compliance review and risk warning, which solves the problems of machine parsing of unstructured standard text, automatic semantic mapping of product parameters and standard clauses, and proactive risk warning throughout the entire standard iteration process.

[0007] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a method for enterprise product compliance review and risk warning, specifically including the following steps: S1. Construct a dynamic standard compliance knowledge graph Standard-KG; S2. Multimodal semantic structuring of enterprise product specification data: This involves importing BOM forms and uploading product specification documents into the enterprise PLM system. Product technical parameters are extracted using a large model, and synonym clustering is performed using an embedded model. Dimensional normalization is achieved using a built-in algorithm, generating a standardized product parameter vector library. ; S3. Dual-track automatic product compliance review: Automatic review is performed using a dual-engine parallel verification mechanism. One track is a hard numerical calculation track, used to handle numerical compliance constraints; the other track is a large model semantic reasoning track, used to handle qualitative and scenario constraints with preconditions. The two tracks are executed in parallel for verification, and the results are corrected by combining knowledge graph exemption nodes to complete the automatic product compliance review. S4, Standard Iterative Graph Propagation Risk Warning; S5. Compliance Reverse Auxiliary Design and Supplier Risk Profile Generation: Receive product functional requirements from R&D, reverse traverse the standard knowledge graph to solve the compliance parameter range, and output a compliance BOM design reference; perform a second-order drill-down graph traversal along the directed edge of the material supply for affected components, aggregate the number of supplier materials at risk, calculate the comprehensive compliance risk coefficient of the supplier using weighted averages, and generate a supplier compliance failure risk profile.

[0008] Preferably, step S1, constructing the dynamic standard compliance knowledge graph Standard-KG, specifically includes: S101. The OCR intelligent parsing system equipped with YOLOv8 and LayoutLM model performs page segmentation, multi-type content extraction and document logical DOM tree reconstruction on multi-source standard documents. S102. Build an asynchronous dual-channel constraint extraction architecture. Channel A extracts semantic entities through a LoRA fine-tuned large language model with a four-layer Prompt system. Channel B extracts numerical values ​​and unit fragments through regular expressions. After cross-validation of the two channels, standard ontology elements are output. The ontology elements include standard objects, technical attributes, test conditions, constraint operators, threshold values, and units of measurement; S103. Encapsulate the validated ontology elements into rule triples. This completes operator normalization and logical expression transformation, where... As the subject of review, As a prerequisite for testing, This is the dimensionally normalized mathematical constraint operator. S104. Store the rule triples into a graph database to generate a standard knowledge graph. Add exemption exception nodes to the graph and associate them with regular rule nodes through the directed edge exempt_from.

[0009] Preferably, the four-layer Prompt system in step S102 includes a task context definition layer, an ontology pattern constraint layer, a thought chain reasoning layer, and a few-shot example layer. The large language model is a Transformer pure decoder architecture, which uses LoRA low-rank fine-tuning to inject industry standard vertical domain knowledge.

[0010] Preferably, step S3, the dual-track product compliance automatic review, specifically includes the following steps: S301. Convert product parameters into dense semantic vectors, retrieve standard knowledge graphs using cosine similarity to match corresponding compliance rules, and set similarity thresholds to filter effective constraint baselines. S302. Build a dual parallel verification engine. Numerical constraints are sent to the hard numerical calculation engine to complete the logical comparison, and qualitative scenario constraints are sent to the large model thinking chain reasoning engine for semantic discrimination. S303. For parameter items judged as non-compliant, retrieve the exemption exception nodes of the spectrum, verify the timeliness and applicable scenarios, and correct the compliance judgment results; S304. Summarize the verification results and automatically generate a product compliance review report.

[0011] Preferably, the formula for calculating the semantic matching cosine similarity in step S301 is: ; For product parameter semantic vectors, The semantic vector of the graph rule node is set to a preset matching threshold of 0.85.

[0012] Preferably, step S4, standard iterative graph propagation risk warning, specifically includes the following steps: S401. The comparison between the old and new standard differences identifies changes in technical indicators, and the risk triggering event nodes are instantiated in the knowledge graph. S402. Starting from the risk-triggered event node, perform a reverse BFS traversal along the directed edge based on the product design to locate all affected products and components, and perform secondary compliance verification based on the new standard rules. S403. Calculate the severity score of product risk based on the degree of violation deviation and the linear weighting of product life cycle weights, and push targeted risk warning work orders according to preset safety thresholds.

[0013] Preferably, the formula for calculating the severity of risk in step S403 is: ; Where α+β=1, The product violation deviation rate after Min-Max normalization. To differentiate the product lifecycle weights at the R&D, mass production, market launch, and delisting stages.

[0014] Preferably, the supplier compliance failure risk profile includes a list of at-risk components, associated standard codes, comprehensive risk coefficients, historical compliant shipment records, and supply chain disruption risk levels.

[0015] (III) Beneficial Effects This invention provides a method for enterprise product compliance review and risk warning. Compared with the prior art, it has the following advantages: (1) The company’s product compliance review and risk warning method is to open up the two-way link between external standard documents and internal product data, and rely on large models and knowledge graphs to realize full-process automated analysis, comparison and verification. The compliance review of the entire product line that originally required several days to several weeks to be completed manually can be completed by the system in just a few minutes, which greatly reduces the manpower input of compliance positions and eliminates the problems of false detection and missed detection caused by human fatigue and knowledge blind spots.

[0016] (2) The company’s product compliance review and risk warning method realizes compliance control before R&D, reducing product recall and administrative penalty losses from the source. This method supports reverse derivation of compliance parameter ranges in the R&D design stage, and engineers can obtain the standard-allowed index range in the early stage of design. At the same time, the machine standardized logic verification has no subjective bias and can accurately identify all non-compliant indicators, avoiding large economic losses such as rework, regulatory penalties and batch recalls due to non-compliance after product mass production and market launch.

[0017] (3) The enterprise’s product compliance review and risk warning method, by constructing an automatic response mechanism for standard changes, transforms compliance risk management from passive remediation to proactive warning. Existing technologies can only handle products after they have violated regulations. This invention realizes full-link impact analysis of standard updates through graph reverse BFS traversal, and instantly locates all affected products and components after the standard indicators are adjusted. By quantifying risk scores and pushing warnings in a graded manner, enterprises can complete product design rectification in advance and prevent the problem of delayed exposure of compliance risks.

[0018] (4) The enterprise’s product compliance review and risk warning method achieves forward control of upstream supplier compliance risks by penetrating the entire supply chain. In the standard change risk analysis, the invention drills down to trace material suppliers in the second order, automatically quantifies the compliance failure risk of each supplier, and generates a visual risk profile. The enterprise can communicate with high-risk suppliers in advance and obtain compliance information to avoid production stagnation and product violation chain risks caused by non-compliance of upstream components in advance, and improve the compliance control system of the entire industrial chain.

[0019] (5) The company’s product compliance review and risk warning method has greatly improved the accuracy of compliance verification through dual-channel extraction and exemption node dual mechanism. The asynchronous dual-channel cross-verification mechanism effectively suppresses the knowledge illusion of large language model text extraction and ensures the accuracy of standard rule extraction. The graph has built-in exemption node independent logic review, which can accurately identify the applicable scenarios of regulatory exemption clauses, avoid misjudging compliant products, take into account the rigor of review and the special requirements of industry regulations, and adapt to complex compliance scenarios in multiple industries such as lithium battery, electrical equipment, and light manufacturing. Attached Figure Description

[0020] Figure 1 This is a flowchart of the enterprise product compliance review and risk warning method of the present invention. Detailed Implementation

[0021] 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.

[0022] Please see Figure 1 This invention provides two technical solutions: a method for enterprise product compliance review and risk warning, specifically including the following embodiments: Example 1: A method for enterprise product compliance review and risk warning, specifically including the following steps: S1. Construct a dynamic standard compliance knowledge graph Standard-KG; S2. Multimodal semantic structuring of enterprise product specification data: This involves importing BOM forms and uploading product specification documents into the enterprise PLM system. Product technical parameters are extracted using a large model, and synonym clustering is performed using an embedded model. Dimensional normalization is achieved using a built-in algorithm, generating a standardized product parameter vector library. ; Step S2 specifically includes the following steps: Multi-source product data fusion and access: The system bidirectionally connects to enterprise PLM and ERP systems, directly exporting structured BOM data tables; it also supports engineers to upload unstructured documents such as PDF and Word format product manuals, component test reports, and R&D drawings, which are uniformly included in the dataset to be processed.

[0023] Large model-driven product technical parameter extraction: Reuse the fine-tuned large language model and information extraction paradigm of step S1, replace the dedicated Prompt ontology adapted to the product document, perform NER entity recognition and relation extraction on BOM and specification document, extract product component names, various technical indicators, measured values ​​of indicators, product application scenarios, supporting supplier information, and generate the original product parameter set.

[0024] Semantic and dimensional dual-layer normalization processing: (1) Synonym semantic clustering: Call the BGE / BERT text embedding model to convert product parameter names into dense semantic vectors, and complete the synonym parameter mapping through cosine similarity calculation to achieve the unification of internal enterprise terminology and national standard terminology, such as the product document "tensile strength" automatically matching the standard text "tensile strength"; (2) Physical dimension conversion normalization: Built-in international unit conversion algorithm to unify parameter measurement units and complete the conversion of units such as mA and A, MΩ and Ω; After normalization, a standardized product parameter vector library is generated: .

[0025] S3, Dual-track system for automatic product compliance review; S4, Standard Iterative Graph Propagation Risk Warning; S5, Compliance Reverse Auxiliary Design and Supplier Risk Profile Generation.

[0026] In this embodiment of the invention, step S1, constructing the dynamic standard compliance knowledge graph Standard-KG, specifically includes: S101. The OCR intelligent parsing system equipped with YOLOv8 and LayoutLM model performs page segmentation, multi-type content extraction and document logical DOM tree reconstruction on multi-source standard documents. Page layout segmentation: The document page is identified using a deep learning object detection algorithm, and independent bounding boxes are divided for text paragraphs, hierarchical headings, data tables, formulas, and charts. Extraction of multiple content types: Text areas use OCR (Optical Character Recognition) to output a complete character stream; Table areas perform table structure recognition, restore row and column coordinates, and convert to Markdown structured table data; Formula areas extract mathematical symbols and threshold values. Document logic reconstruction: Reconstruct the full-text reading order based on document title hierarchy rules, and generate a standardized DOM logic tree as the underlying input data for subsequent constraint extraction.

[0027] S102. Build an asynchronous dual-channel constraint extraction architecture. Channel A extracts semantic entities through a LoRA fine-tuned large language model with a four-layer Prompt system. Channel B extracts numerical values ​​and unit fragments through regular expressions. After cross-validation of the two channels, standard ontology elements are output. Ontology elements include standard objects, technical attributes, test conditions, constraint operators, threshold values, and units of measurement; Channel A: Deep Semantic Extraction Channel. It employs a large language model with a Transformer pure decoder architecture fine-tuned using LoRA low-rank, coupled with a four-layer structured knowledge extraction Prompt system to complete entity extraction. The four Prompt systems include a task context definition layer, an ontology pattern constraint layer, a thought chain reasoning layer, and a few-shot example layer: The task context definition layer limits the model to a standardized compliance audit expert, with input document DOM trees and table matrices, and the task objective is lossless extraction of compliance rules; the ontology pattern constraint layer forces the model to output six types of entities according to a preset ontology, with the output format limited to program-parsable JSON; the thought chain reasoning layer guides the model to decompose the preconditions within the clauses and distinguish between quantitative indicators and qualitative descriptions; the few-shot example layer includes 3-5 sets of GB and ISO standard extraction examples, improving the accuracy of industrial terminology extraction through context learning.

[0028] Channel B: Hard Regular Expression Channel. It has a pre-built library of industrial unit of measurement regular expressions that independently extracts pure numbers, physical units, and logical operator fragments from the original standard text, without relying on semantic judgments from a large model.

[0029] Dual-channel cross-validation: Construct a high-dimensional vector alignment matrix, calculate the intersection-union ratio (IoU) of the page coordinates of entity text in channel A and numerical fragments in channel B, and complete the semantic vector space alignment; retain the extraction results only when the extracted elements of the two channels are completely matched; if the matching fails, a reflection prompt word is triggered, and the extracted text is verified a second time by the large model.

[0030] S103. Encapsulate the validated ontology elements into rule triples. This completes operator normalization and logical expression transformation, where... As the subject of review, As a prerequisite for testing, This is the dimensionally normalized mathematical constraint operator. Semantic slot filling: Slot binding standard object technical attributes; Fill the slots with test conditions; leave them empty if there are no prior scenario constraints. Slot integration constraint operators, thresholds, and units of measurement; Operator normalization: Establish a mapping relationship between natural language descriptions and mathematical operators, unifying "not less than" and "at least" into ≥, "not exceeding" and "upper limit" into ≤, and "within the interval" into ∈[min,max]. Formal expression generation: converts encapsulated content into directly computable logical expressions. Example: The standard clause "The discharge capacity of the lithium battery at 25℃ is not less than 3000mAh" extracts the following entities as standard objects: lithium battery, technical attributes: discharge capacity, test conditions: ambient temperature = 25℃, operator: not less than, threshold: 3000, unit: mAh; packaged as a ternary array. Lithium-ion battery. Discharge capacity. Ambient temperature == 25°C : value>=3000.

[0031] S104. All compliance rule triples are stored in a graph database to construct a standard knowledge graph: Graph nodes include standard clause nodes, component entity nodes, technical indicator nodes, and exemption exception nodes; semantic relationships between nodes are defined through directed edges, including four basic association edges: "constrained by", "test preconditions", and "belongs to"; for exemption clauses within regulations, a separate exception node is created, and the node metadata stores the exempted substance, applicable scenario, and exemption effective / expiration time. It is associated with regular rule nodes through the exempt_from directed edge for subsequent compliance verification and exemption determination.

[0032] In step S102, the four-layer Prompt system includes a task context definition layer, an ontology pattern constraint layer, a thought chain reasoning layer, and a few-shot example layer. The large language model is a Transformer pure decoder architecture, which uses LoRA low-rank fine-tuning to inject industry-standard vertical domain knowledge.

[0033] In this embodiment of the invention, step S3, the dual-track product compliance automatic review, specifically includes the following steps: S301. Convert product parameters into dense semantic vectors, retrieve standard knowledge graphs using cosine similarity to match corresponding compliance rules, and set similarity thresholds to filter effective constraint baselines. S302. Build a dual parallel verification engine. Numerical constraints are sent to the hard numerical calculation engine to complete the logical comparison, and qualitative scenario constraints are sent to the large model thinking chain reasoning engine for semantic discrimination. S303. For parameter items judged as non-compliant, retrieve the exemption exception nodes of the spectrum, verify the timeliness and applicable scenarios, and correct the compliance judgment results; S304. Summarize the verification results and automatically generate a product compliance review report.

[0034] The formula for calculating the semantic matching cosine similarity in step S301 is as follows: ; For product parameter semantic vectors, The semantic vector of the graph rule node is set to a preset matching threshold of 0.85.

[0035] In this embodiment of the invention, step S4, standard iterative graph propagation risk warning, specifically includes the following steps: S401. The comparison between the old and new standard differences identifies changes in technical indicators, and the risk triggering event nodes are instantiated in the knowledge graph. S402. Starting from the risk-triggered event node, perform a reverse BFS traversal along the directed edge based on the product design to locate all affected products and components, and perform secondary compliance verification based on the new standard rules. S403. Calculate the severity score of product risk based on the degree of violation deviation and the linear weighting of product life cycle weights, and push targeted risk warning work orders according to preset safety thresholds.

[0036] The formula for calculating the severity of risk in step S403 is as follows: ; Where α+β=1, The product violation deviation rate after Min-Max normalization. To differentiate the product lifecycle weights at the R&D, mass production, market launch, and delisting stages.

[0037] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that: Step S5, compliance reverse auxiliary design and supplier risk profile generation, specifically includes the following steps: S501: Receive the functional requirements of the R&D product, reverse traverse the standard knowledge graph to solve the compliance parameter range, and output a compliant BOM design reference. S502. Perform a second-order drill-down traversal of the material supply path along the directed edge of the affected components, aggregate the number of materials at risk from suppliers, calculate the comprehensive compliance risk coefficient of suppliers using a weighted average, and generate a profile of supplier compliance failure risk.

[0038] The supplier compliance failure risk profile in step S502 includes a list of at-risk components, associated standard codes, comprehensive risk coefficients, historical compliant shipment records, and supply chain disruption risk levels.

[0039] After first-order graph traversal to identify components affected by standard changes, continue second-order drill-down along the directed edge of material supply [provided_by] in the graph to complete the supply chain risk analysis: (1) Supplier material aggregation statistics: All affected parts are classified according to upstream supplier entities, and the total number of materials associated with each supplier affected is counted; (2) Calculation of supplier's comprehensive risk coefficient: sum up the product risk scores corresponding to all risky components of the supplier. The weighted average yields the comprehensive coefficient of supplier compliance failure. (3) Generate standardized supplier risk profiles: The profiles include a list of risky components, the corresponding mandatory standard numbers, comprehensive risk coefficients, historical compliant shipment records, and potential supply disruption risk levels; the compliance department requests new versions of compliance test reports from high-risk suppliers in advance based on the profiles to proactively resolve supply chain compliance risks.

[0040] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for enterprise product compliance review and risk warning, characterized in that: Specifically, the following steps are included: S1. Construct a dynamic standard compliance knowledge graph Standard-KG; S2. Multimodal semantic structuring of enterprise product specification data: This involves importing BOM forms and uploading product specification documents into the enterprise PLM system. Product technical parameters are extracted using a large model, and synonym clustering is performed using an embedded model. Dimensional normalization is achieved using a built-in algorithm, generating a standardized product parameter vector library. ; S3. Dual-track automatic product compliance review: Automatic review is performed using a dual-engine parallel verification mechanism. One track is a hard numerical calculation track, used to handle numerical compliance constraints; the other track is a large model semantic reasoning track, used to handle qualitative and scenario constraints with preconditions. The two tracks are executed in parallel for verification, and the results are corrected by combining knowledge graph exemption nodes to complete the automatic product compliance review. S4, Standard Iterative Graph Propagation Risk Warning; S5. Compliance Reverse Auxiliary Design and Supplier Risk Profile Generation: Receive product functional requirements from R&D, reverse traverse the standard knowledge graph to solve the compliance parameter range, and output a compliance BOM design reference; perform a second-order drill-down graph traversal along the directed edge of the material supply for affected components, aggregate the number of supplier materials at risk, calculate the comprehensive compliance risk coefficient of the supplier using weighted averages, and generate a supplier compliance failure risk profile.

2. The method for enterprise product compliance review and risk warning according to claim 1, characterized in that: Step S1, constructing the dynamic standard compliance knowledge graph Standard-KG, specifically includes: S101. The OCR intelligent parsing system equipped with YOLOv8 and LayoutLM model performs page segmentation, multi-type content extraction and document logical DOM tree reconstruction on multi-source standard documents. S102. Build an asynchronous dual-channel constraint extraction architecture. Channel A extracts semantic entities through a LoRA fine-tuned large language model with a four-layer Prompt system. Channel B extracts numerical values ​​and unit fragments through regular expressions. After cross-validation of the two channels, standard ontology elements are output. The ontology elements include standard objects, technical attributes, test conditions, constraint operators, threshold values, and units of measurement; S103. Encapsulate the validated ontology elements into rule triples. This completes operator normalization and logical expression transformation, where... As the subject of review, As a prerequisite for testing, This is the dimensionally normalized mathematical constraint operator. S104. Store the rule triples into a graph database to generate a standard knowledge graph. Add exemption exception nodes to the graph and associate them with regular rule nodes through the directed edge exempt_from.

3. The method for enterprise product compliance review and risk warning according to claim 2, characterized in that: The four-layer Prompt system in step S102 includes a task context definition layer, an ontology pattern constraint layer, a thought chain reasoning layer, and a few-shot example layer. The large language model is a Transformer pure decoder architecture, which uses LoRA low-rank fine-tuning to inject industry standard vertical domain knowledge.

4. The method for enterprise product compliance review and risk warning according to claim 1, characterized in that: The S3 dual-track product compliance automatic review specifically includes the following steps: S301. Convert product parameters into dense semantic vectors, retrieve standard knowledge graphs using cosine similarity to match corresponding compliance rules, and set similarity thresholds to filter effective constraint baselines. S302. Build a dual parallel verification engine. Numerical constraints are sent to the hard numerical calculation engine to complete the logical comparison, and qualitative scenario constraints are sent to the large model thinking chain reasoning engine for semantic discrimination. S303. For parameter items judged as non-compliant, retrieve the exemption exception nodes of the spectrum, verify the timeliness and applicable scenarios, and correct the compliance judgment results; S304. Summarize the verification results and automatically generate a product compliance review report.

5. The method for enterprise product compliance review and risk warning according to claim 4, characterized in that: The formula for calculating the semantic matching cosine similarity in step S301 is as follows: ; For product parameter semantic vectors, The semantic vector of the graph rule node is set to a preset matching threshold of 0.

85.

6. The method for enterprise product compliance review and risk warning according to claim 1, characterized in that: The step S4, standard iterative graph propagation risk warning, specifically includes the following steps: S401. The comparison between the old and new standard differences identifies changes in technical indicators, and the risk triggering event nodes are instantiated in the knowledge graph. S402. Starting from the risk-triggered event node, perform a reverse BFS traversal along the directed edge based on the product design to locate all affected products and components, and perform secondary compliance verification based on the new standard rules. S403. Calculate the severity score of product risk based on the degree of violation deviation and the linear weighting of product life cycle weights, and push targeted risk warning work orders according to preset safety thresholds.

7. The method for enterprise product compliance review and risk warning according to claim 6, characterized in that: The formula for calculating the severity of risk in step S403 is as follows: ； Where α+β=1, The product violation deviation rate after Min-Max normalization. To differentiate the product lifecycle weights at the R&D, mass production, market launch, and delisting stages.

8. The method for enterprise product compliance review and risk warning according to claim 1, characterized in that: The supplier compliance failure risk profile includes a list of at-risk components, associated standard codes, comprehensive risk coefficients, historical compliant shipment records, and supply chain disruption risk levels.

Images