Supply Chain Data Intelligence Construction and Dynamic Optimization Methods and Systems

Abstract

This invention discloses a method and system for intelligent construction and dynamic optimization of supply chain data. It constructs an ontology based on OWL and defines a semantic space. It collects text and image data of business scope, performs layout detection and OCR to extract text from images, and encodes the text into visual embeddings. It then fuses TF-IDF vectors, semantic vectors, and visual embeddings to generate a feature matrix. A rule-based expression tree with AND / OR nodes is constructed to evaluate the feature matrix, outputting candidate companies and matching degrees. The expression tree is made differentiable and gated; within a sliding window, a loss function is constructed based on target number deviation and feedback from missed detections, and gradient descent is used to update keyword weights and thresholds. A relationship graph is constructed based on candidate companies to calculate PageRank centrality, and this is fused with matching degree and time decay to output a weighted ranking. Double buffering updates are used, and the keyword system is expanded with term co-occurrence networks, Louvain, and semantic drift detection to output a traceable evidence chain.

Description

Technical Field

[0001] This invention relates to the field of big data processing technology, specifically to a method and system for intelligent construction and dynamic optimization of industrial chain data. Background Technology

[0002] Supply chain data is a crucial foundation for conducting supply chain analysis, supply chain resilience assessment, identification and profiling of enterprises within the chain, industrial policy evaluation, and investment attraction decisions. Current supply chain data construction typically relies on industry directories, manual experience rules, or single data sources (such as business scope, industry classification codes, and company directories) for enterprise aggregation and process mapping. However, the supply chain itself is characterized by complex hierarchies, cross-industry overlap, rapid technological iteration, and frequent changes in enterprise information. This makes it difficult for traditional methods to balance coverage, accuracy, and timeliness of updates, thus limiting the usability and reliability of supply chain data.

[0003] On the one hand, existing technologies mostly define supply chain links at the level of textual descriptions or static directories, lacking a computable, configurable, and updateable semantic structure expression that can evolve with the industry. Common practices rely solely on "industry classification codes + keyword searches" for link location, making it difficult to formally constrain and unify the semantic governance of the multi-level relationships of "chain segment—link—enterprise," and also hindering the integration of subsequent data collection, rule verification, and evidence tracing processes into the same semantic framework. Due to the lack of ontological modeling and the definition of a semantic feature space, supply chain definitions formed by different departments or at different times are often inconsistent, making it difficult to reproduce or audit the results of the same supply chain analysis.

[0004] On the other hand, the identification of supply chain enterprises relies heavily on textual descriptions in publicly available information, but this textual information suffers from problems such as scattered sources, non-standard expression, and serious omissions. Taking the business scope as an example, enterprises often use generic terms or marketing phrases, which can easily lead to vague keywords, synonym substitutions, and inconsistent industry terminology. In addition, a large amount of key evidence does not exist in the form of structured fields, but rather in image carriers such as webpage snapshots, scanned copies of qualification certificates, and product catalog PDF pages. Existing supply chain data construction solutions mostly focus on structured fields or plain text retrieval, lacking the ability to analyze the layout of image data, perform OCR recognition, and align text and image semantics. This results in insufficient extraction and utilization of key information such as enterprise qualifications, product range, and certificate numbers, leading to missed detections or misjudgments.

[0005] Meanwhile, existing rule-based enterprise aggregation methods typically rely on manually setting thresholds and keyword combinations, making it difficult for rules to adaptively adjust to changes in sample size and target scope. When the industry chain scope requires controlling the number of enterprises or coverage targets, traditional methods often require repeated manual trial and error, resulting in low parameter tuning efficiency and difficulty in forming a reusable optimization mechanism. More importantly, many existing solutions fail to create a closed loop between the "rule determination process" and "optimization feedback signals": for example, feedback information generated from supplementing enterprise data (correcting missed detections) is difficult to systematically update keyword weights or thresholds, leading to long-term rule rigidity and high maintenance costs.

[0006] Furthermore, assessing the importance of companies in a supply chain depends not only on whether a rule is met at a particular stage, but also on the company's position within the supply chain network, the strength of its upstream and downstream connections, and the timeliness of information. Existing company ranking methods are often based on single indicators (such as registered capital, number of patents, or simple matching scores), lacking a mechanism that integrates centrality indicators of the company relationship graph (such as PageRank) with factors like matching degree and time decay, making it difficult to output company ranking results that better reflect the characteristics of key nodes in the supply chain.

[0007] Finally, industry chain data is highly time-sensitive. Business scope, qualification certificates, product catalogs, and publicly available information change frequently over time. Relying solely on periodic full updates results in significant delays and high computational costs; while incremental updates may introduce inconsistencies, affecting the stability of online services. Simultaneously, industry terminology continuously evolves with technology, with new technologies, products, and materials causing keyword shifts and expansions. Existing industry chain data construction schemes are insufficient in recognizing terminology evolution, lack dynamic expansion mechanisms based on terminology co-occurrence networks, community discovery, and semantic drift detection, and lack the ability to write the expansion results back into the industry chain semantic space and rule system. This makes it difficult for industry chain definitions to keep pace with industry changes.

[0008] Therefore, there is an urgent need for a method and system for intelligent construction and dynamic optimization of supply chain data that can integrate text and image data within an ontology semantic framework, realize image layout parsing and image-text semantic encoding, support differentiable gating and sliding window self-optimization of rule expression trees, combine enterprise relationship graph centrality sorting, and achieve continuous dynamic optimization through double buffering updates and term evolution tracking, so as to improve the coverage, accuracy, traceability and update timeliness of on-chain enterprise identification. Summary of the Invention

[0009] To address the aforementioned technical problems, this invention discloses a method and system for intelligent construction and dynamic optimization of supply chain data. It constructs an ontology based on OWL and defines a semantic space. It collects text and image data of business scope, performs layout detection and OCR to extract image text, and encodes the text into visual embeddings. It then fuses TF-IDF vectors, semantic vectors, and visual embeddings to generate a feature matrix. A rule-based expression tree containing AND / OR nodes is constructed to evaluate the feature matrix, outputting candidate companies and matching degrees. The expression tree is made differentiable and gated; within a sliding window, a loss function is constructed based on target number deviation and feedback from missed detections, and gradient descent is used to update keyword weights and thresholds. A relationship graph is constructed based on candidate companies to calculate PageRank centrality, which is then fused with matching degree and time decay to output a weighted ranking. A double-buffered update is employed, and the keyword system is expanded using a term co-occurrence network, Louvain, and semantic drift detection to output a traceable evidence chain.

[0010] This application provides a method for intelligent construction and dynamic optimization of supply chain data, including the following steps: S1: Construct a three-level industry chain ontology based on OWL and define the semantic feature space of the target links; S2: Collect enterprise data, which includes the business scope text field, as well as image data such as enterprise webpage snapshots, enterprise qualification certificate images, and product catalog PDF page images; S3: Perform layout area detection on the image data to obtain text area coordinates, perform OCR on the text area to obtain image text fields, and perform image-text semantic encoding on the image data and image text fields to obtain visual embedding; S4: Extract TF-IDF features from the business scope text field and the image text field respectively, and encode them semantically using a pre-trained language model. Then, fuse them with the visual embedding to generate an enterprise feature matrix. S5: Construct a rule expression tree containing AND and OR nodes. The expression tree consists of sub-condition nodes based on TF-IDF weights and semantic similarity scores. Input the enterprise feature matrix into the rule expression tree to perform sub-condition judgment, output candidate enterprises, and calculate the matching degree based on the sub-condition hit confidence. Make the expression tree differentiable and construct a loss function within a sliding window based on the deviation between the number of candidate enterprises and the target size, as well as the missed detection feedback formed by supplementary entries. Adaptively update the keyword weights and threshold parameters through gradient descent. S6: Construct an enterprise relationship graph based on candidate enterprises and calculate PageRank centrality. Combine centrality, matching degree and time decay coefficient based on enterprise data update time to obtain enterprise weights and output the ranking. S7: A double-buffered update mechanism is adopted to perform full periodic updates and / or event-driven incremental updates in the computation buffer, and the online buffer is replaced by atomic switching. A term co-occurrence network is constructed based on the business scope text field and the image text field. The keyword system is dynamically expanded through Louvain community discovery and pre-trained language model semantic drift detection, and the keyword set of the semantic feature space and rule expression tree is updated accordingly. The output includes the evidence chain containing hit sub-conditions, image region coordinates and text fragment positions.

[0011] Preferably, in step S3, after OCR, field extraction is performed on the image text field to generate a structured field set. The structured field set includes the enterprise name, unified social credit code, certificate name, certificate number, product name, and key technical keywords, and the structured field set is incorporated into the enterprise feature matrix.

[0012] Preferably, in step S4, TF-IDF vectors and semantic encoding vectors are output for the business scope text field and the image text field, respectively, and then fused with the visual embedding in a way that involves concatenation, weighted summation, or gated fusion to form an enterprise feature matrix, wherein the fusion weight is adaptively adjusted by the matching degree feedback.

[0013] Preferably, the sub-condition nodes in step S5 include: a) keyword hitting conditions based on the TF-IDF weight threshold of the business scope text field; b) semantic matching conditions based on the semantic encoding vector similarity threshold; c) regular expression matching conditions based on the certificate number and product name pattern in the image text field; and AND nodes are used to ensure that multiple conditions are met simultaneously, and OR nodes are used to ensure that one of the multiple conditions is met.

[0014] Preferably, the differentiability gating of the expression tree in step S5 includes replacing the AND node with a multiplicative gating function, replacing the OR node with an additive gating function or a LogSumExp gating function, and using the confidence level output by the sub-condition node as the gating input, so that the keyword weight, threshold parameter and gating parameter can be jointly updated through gradient descent, thereby improving the rule convergence speed and reducing manual parameter tuning.

[0015] Preferably, the missed detection feedback in step S5 is generated by an active learning strategy. The active learning strategy includes: selecting companies whose matching degree is in a preset uncertainty range as supplementary candidates from unselected companies, and prioritizing the push of samples with the highest semantic similarity to the target link keywords and the largest confidence fluctuation for manual confirmation, so as to feed the confirmation results back into the loss function as missed detection feedback.

[0016] Preferably, the evidence chain in step S7 further includes a cross-modal alignment index, which includes: file identifier and page number of image data, text region coordinates, OCR character-level or word-level position index, corresponding business scope text fragment position index, and sub-condition node identifier and threshold that trigger the selection of the candidate enterprise, so that each enterprise selection can be traced back to a specific image region and a specific text fragment.

[0017] Preferably, in step S1, the three-level industry chain ontology constructed based on OWL includes a three-layer structure of industry, chain segment, and link, and a computable semantic feature space is defined in the ontology for the target link. The semantic feature space includes: (i) The set of keywords and their initial weights for the business scope text field; (ii) The set of keywords and their initial weights for the image text field; (iii) A sub-condition template for a rule expression tree, wherein the sub-condition template includes at least a TF-IDF threshold condition and a semantic similarity threshold condition; (iv) The field mapping rules of the evidence chain are used to establish a correspondence between the keywords of the target link and the text region coordinates output in step S3 and the semantic encoding vector output in step S4; The semantic feature space serves as the configuration input for constructing the rule expression tree in step S5, and as the update object when dynamically expanding the keyword system in step S7.

[0018] Preferably, the visual embedding is obtained by performing image-text semantic encoding on the image data and the image text field, including: a) Crop the image sub-regions corresponding to the text region coordinates obtained in step S3 into several ROI image blocks, and input each ROI image block into the visual encoder to obtain the ROI visual vector; b) Segment the image text field by word or phrase and input it into a pre-trained language model to obtain text vectors; c) A cross-modal alignment network is used to calculate the similarity between the ROI visual vector and the text vector and perform attention-weighted fusion to obtain a region-level visual embedding that corresponds one-to-one with the coordinates of each text region. The region-level visual embedding is then pooled to obtain an enterprise-level visual embedding. The regional visual embedding serves as the input for the enterprise feature matrix generated in step S4, and is bound to the text region coordinates when the evidence chain is output in step S7.

[0019] This application also provides a data intelligence construction and dynamic optimization system for the industrial chain, including: The module constructs a three-tier industry chain ontology based on OWL and defines the semantic feature space of the target links; The data collection module collects enterprise data, including the business scope text field, as well as image data such as screenshots of enterprise web pages, images of enterprise qualification certificates, and PDF pages of product catalogs. The extraction module performs layout region detection on the image data to obtain text region coordinates, performs OCR on the text region to obtain image text fields, and performs image-text semantic encoding on the image data and image text fields to obtain visual embedding; The enterprise feature matrix generation module performs TF-IDF feature extraction and pre-trained language model semantic encoding on the business scope text field and image text field respectively, and then fuses them with visual embedding to generate an enterprise feature matrix. A rule expression tree module containing AND and OR nodes is constructed. The expression tree consists of sub-condition nodes based on TF-IDF weights and semantic similarity scores. The enterprise feature matrix is ​​input into the rule expression tree for sub-condition judgment, and candidate enterprises are output. The matching degree is calculated based on the sub-condition hit confidence. The expression tree is made differentiable and gated. Within a sliding window, a loss function is constructed based on the deviation between the number of candidate enterprises and the target size, as well as the missed detection feedback formed by supplementary entries. The keyword weights and threshold parameters are adaptively updated through gradient descent. The ranking module constructs an enterprise relationship graph based on candidate enterprises and calculates PageRank centrality. It then integrates centrality, matching degree, and time decay coefficient based on enterprise data update time to obtain enterprise weights and outputs the ranking. The output module employs a double-buffered update mechanism to perform periodic full updates and / or event-driven incremental updates in the computation buffer, and replaces the online buffer with an atomic switching method. It constructs a term co-occurrence network based on the business scope text field and the image text field, dynamically expands the keyword system through Louvain community discovery and pre-trained language model semantic drift detection, and updates the keyword set of the semantic feature space and rule expression tree accordingly. The output includes a chain of evidence containing hit sub-conditions, image region coordinates, and text fragment positions.

[0020] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This application uses the OWL three-level industry chain ontology as a unified semantic framework, explicitly solidifying the keyword set, sub-condition templates, and evidence mapping rules of the target link, making the enterprise aggregation criteria configurable, reusable, and auditable. Compared with single-source matching that relies solely on business scope or industry code, this application also introduces image data such as webpage snapshots, qualification certificates, and product catalog PDFs. It locates text regions through layout area detection, extracts image text fields through OCR, and further performs image-text semantic encoding to obtain visual embedding. This is then fused with the TF-IDF features of the business scope text and the semantic vectors of the pre-trained language model to generate an enterprise feature matrix. As a result, it can reliably identify enterprises on the chain even when unstructured evidence such as certificate numbers, product names, and key technical terms are present, significantly reducing missed detections caused by missing text, synonym substitution, or non-standard expressions. At the same time, it provides an evidence chain for each hit, outputting the coordinates of the image region and the position of the text fragment, enhancing the interpretability and traceability of the results.

[0021] (2) This application constructs a rule expression tree containing AND / OR nodes, expressing sub-conditions such as TF-IDF weight threshold, semantic similarity threshold, and regular pattern matching in a combinable form, realizing collaborative judgment between "interpretable rules" and "semantic models"; further, the expression tree is made differentiable and gated, mapping the AND / OR logic to a trainable gate function, so that keyword weights, threshold parameters, and gate parameters can be jointly updated through gradient descent within a sliding window. The system uses the deviation between the number of candidate enterprises and the target size as constraints, and introduces the missed detection feedback formed by supplementary recording as a supervision signal to construct a loss function, thereby automatically finding the optimal balance between scale control and identification accuracy, avoiding the problems of long cycles and poor reusability caused by traditional rule schemes relying on repeated trial and error by experts and manual parameter tuning. Combined with the active learning supplementary recording mechanism that prioritizes the confirmation of samples in uncertain intervals, higher quality feedback can be obtained with less manual annotation, further reducing maintenance costs and improving rule convergence efficiency and stability.

[0022] (3) This application constructs an enterprise relationship graph based on the candidate enterprise set and calculates PageRank centrality. It integrates centrality, matching degree, and time decay coefficient based on update time to form enterprise weight, realizing a three-dimensional ranking of "hit degree - network position - timeliness". Compared with ranking based on a single matching score or static indicators, this scheme can highlight key node enterprises in the industrial chain, reduce the risk of "weakly related but text-hit" noisy enterprises ranking high, and improve the reliability of key enterprise discovery. At the same time, a double-buffer update mechanism is adopted to perform full-cycle updates and / or event-driven incremental updates in the calculation buffer, and the online buffer is replaced by atomic switching to ensure decoupling between online query and back-end calculation, avoiding inconsistencies or service jitter introduced during the update process. Combined with the dynamic expansion of the keyword system of term co-occurrence network, Louvain community discovery and semantic drift detection, it can continuously track the evolution of industrial technology and write back the expansion results to the semantic feature space and rule expression tree to ensure that the industrial chain caliber and data results maintain timeliness and consistency in the long term. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method for intelligent construction and dynamic optimization of supply chain data according to the present invention; Figure 2 This is a structural diagram of a data intelligence construction and dynamic optimization system for the industrial chain according to the present invention. Detailed Implementation

[0024] Those skilled in the art will understand that, in order to make the above-mentioned objects, features, and beneficial effects of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This application illustrates a method for intelligent construction and dynamic optimization of supply chain data, including the following steps: S1: Construct a three-level industry chain ontology based on OWL and define the semantic feature space of the target link. In some embodiments, the three-level industry chain ontology (OWL) is constructed using the "new energy battery industry chain - battery recycling link" as an example. This embodiment uses the OWL language to establish a three-level industry chain ontology, with the following three levels: 1) Industry layer (first level): new energy battery industry; 2) Chain segment layer (second level): upstream chain segment, midstream chain segment, downstream chain segment; 3) Link layer (third level): "battery recycling link" (target link) is set under the downstream chain segment, and "cascade utilization link", "recycled materials link", etc. can be set in parallel. The basic relationships are defined in the ontology: "link belongs to chain segment", "chain segment belongs to industry"; "link associated semantic feature space"; "link associated rule template"; "link associated evidence mapping rule". Through the above relationships, the subsequent system only needs to input the "target link identifier" to read the semantic configuration, rule structure and evidence output format corresponding to the link from the ontology.

[0025] For the "battery recycling process," a "semantic feature space" is created within the ontology, containing at least the following content, with an initial configuration example provided: A set of text keywords (for business scope / image text). Business scope keyword set A: e.g., "battery recycling, waste batteries, dismantling, cascade utilization, recycling, resource recovery, black powder, cathode material regeneration," etc.; Image text keyword set B (for certificate / catalog / OCR text): e.g., "recycling and processing qualification, hazardous waste management license, resource recovery, dismantling and processing, recycled raw materials," etc. Initial weights are assigned to the keywords (e.g., higher confidence keywords have higher weights), which will serve as initial parameters for subsequent S5 gradient descent self-optimization.

[0026] The threshold parameter set (used for rule determination) is pre-configured in the semantic feature space with three types of thresholds: TF-IDF weight threshold: to determine whether the importance of keywords in the business scope or image text meets the requirements; semantic similarity threshold: to determine whether the semantic encoding of the text is close to the semantics of the target link; regularity / pattern threshold: such as the matching of certificate number or license words, which is considered to meet the requirements. These thresholds also serve as updatable parameters for subsequent self-optimization.

[0027] Sub-condition templates and combination structures (used to generate rule expression trees) define "sub-condition templates" in the semantic feature space. For example: Sub-condition 1: The business scope TF-IDF hits keywords such as "battery recycling / dismantling" and the weight exceeds the threshold; Sub-condition 2: The semantic similarity between the business scope and "recycling and processing" exceeds the threshold; Sub-condition 3: The OCR image text contains patterns such as "hazardous waste operation license / recycling and processing qualification"; and give combination relationships (used for subsequent automatic generation of AND / OR expression trees by S5), such as: "(sub-condition 1 or sub-condition 2 or sub-condition 3) and (necessary compliance conditions)".

[0028] To ensure that S7 can output a "traceable chain of evidence" in the future, an evidence mapping is defined in the semantic feature space: if the business scope text keyword is hit: record the start and end positions of the corresponding text fragment; if the image text keyword is hit: record the page number and text area coordinates of the OCR output, as well as the corresponding word / character position index; if a sub-condition is hit: record the sub-condition number, threshold, score and the hit keyword.

[0029] Connection with S3: The "evidence mapping rules" defined in S1 specify which coordinates / positions need to be output. Therefore, S3 must retain "page numbers, text area coordinates, and character / word positions" during layout detection and OCR to meet the evidence chain requirements. Connection with S4: The keyword set and threshold defined in S1 determine the vocabulary range, weight initialization, and feature fusion strategy of S4 in TF-IDF and semantic encoding. Connection with S5: The sub-condition templates and combination structures defined in S1 directly generate the expression tree skeleton, with weights and thresholds serving as optimization variables for differentiability gating and gradient descent. Connection with S7: New terms obtained from term co-occurrence and semantic drift detection in S7 will be rewritten according to the semantic feature space structure of S1: the new terms will be added to the keyword set, weights will be initialized, and sub-condition templates will be updated, thereby achieving "automatic expansion of the link scope as the industry evolves". By mapping the keywords, thresholds, sub-condition templates, and evidence of the "target link" into the ontology in a structured definition, the subsequent identification and optimization process no longer relies on manual temporary splicing rules, but uses the ontology as a unified configuration and write-back carrier to realize the construction of configurable, traceable, self-optimizable, and evolvable industrial chain data.

[0030] S2: Collect enterprise data, including the business scope text field, as well as image data such as screenshots of enterprise web pages, images of enterprise qualification certificates, and images of product catalog PDF pages. In some embodiments (taking the candidate enterprise database for the "battery recycling link" as an example), the system uses "enterprise" as the smallest data unit and establishes a unique enterprise identifier (EnterpriseID, such as the unified social credit code or its hash value) for each enterprise. The system collects the following data for each enterprise: 1) Business scope text field: from the original text of the business scope field in the business registration information; 2) Enterprise web page screenshot: from the enterprise's official website, enterprise publicity page, enterprise product introduction page, or investment promotion platform page; 3) Enterprise qualification certificate image: from scanned copies / photos of qualification certificates publicly displayed by the enterprise (such as recycling and processing qualifications, licenses, etc.); 4) Product catalog PDF page image: from the enterprise's publicly available product catalog PDF, product manual PDF, or brochure PDF, converted into images page by page and stored. All of the above data are bound and stored with the same EnterpriseID, forming the image input required for S3 and the text input required for S4.

[0031] Acquisition and Standardization of the Business Scope Text Field: 1) The system acquires basic enterprise information from enterprise business registration data sources (such as the business registration disclosure interface, authorized enterprise information database, or internal enterprise business registration database); 2) The "Business Scope" field is extracted from the business registration information as the business scope text field and standardized: meaningless symbols and repeated spaces are removed; full-width / half-width characters and punctuation are standardized; and sentence segmentation (segmented by "; / , / .") is performed to form a list of short sentences for subsequent TF-IDF calculation; 3) The standardized business scope text and the original text are stored in the enterprise data package simultaneously, recording the data source and collection timestamp for use by S6 time decay.

[0032] The enterprise webpage snapshot screenshot collection and management system maintains a "webpage collection URL list" for each enterprise. URLs can come from: the enterprise's official website homepage, product / service pages, qualification display pages, news and updates pages, etc. During collection, a "headless browser rendering + screenshot" method is used to generate webpage snapshots: the page is fully rendered (loading main text and images); a full-page screenshot or a screenshot of a key area (e.g., containing areas such as "qualifications," "products," "service scope," etc.) is generated; the corresponding URL, screenshot time, page title, and resolution are recorded; to improve data quality, the system performs basic screening of webpage snapshot screenshots: if the text area in the screenshot is too small or appears to be a pure image advertisement, it is marked as low quality and demoted; if the screenshot contains prompts such as "qualification certificate" or "product catalog download," it is placed in a priority queue to trigger subsequent certificate / catalog collection. Webpage snapshot screenshots are directly sent as image data to S3 for layout inspection and OCR.

[0033] The enterprise qualification certificate image acquisition system locates relevant "qualification / certificate" entry points (such as "Qualification & Honor," "Certificate Download," and "License" sections) from web snapshots or enterprise websites. It then collects certificate resources, including: scanned copies of certificates (JPG / PNG); screenshots of certificate pages; and PDF certificates (if PDF, each page is converted to an image and saved). After acquisition, metadata is added to each certificate image: EnterpriseID, certificate source URL, acquisition timestamp, certificate category (which can be roughly divided into "License / Qualification / Certification / Honor"), file format, and page number information (if from PDF). This allows S3 to locate the certificate number area, seal area, and certificate name area through layout detection, providing reliable evidence for the regular expression pattern sub-conditions in S5.

[0034] The product catalog PDF page image acquisition system obtains product catalog PDFs / product manual PDFs from the company's official website, product pages, or the "Downloads" section. It then performs page conversion on the PDF files: rendering each page as an image at a fixed resolution (e.g., 300dpi), resulting in "product catalog PDF page images." For each page image, it records the PDF_ID + page_no and binds it to the company's EnterpriseID. To reduce invalid pages, the system performs simple filtering: only pages containing keywords such as product name, model, parameter table, and applicable scenarios are retained. Pages consisting solely of covers or images are marked, and their inclusion in the S3 OCR quality assessment is determined accordingly. After the product catalog page images enter S3, product names, models, and key technical terms can be extracted, thus compensating for omissions caused by insufficient business scope text.

[0035] After data collection is complete, the system generates an "Enterprise Data Package" for each enterprise (closely integrated with S3 / S4), which includes at least: EnterpriseID; a business scope text field (original text + standardized version + sentence list); an image data list: a collection of webpage snapshot screenshots, a collection of qualification certificate images, and a collection of product catalog PDF page images; and metadata for each image: source, timestamp, page number / resolution / URL, etc. This enterprise data package directly serves as: S3 input: image data list; S4 input: business scope text field (and the image text field and visual embeddings output by S3).

[0036] By uniformly collecting and binding image data such as "business scope text field" and "webpage / certificate / directory" by enterprise, when the enterprise's business scope description is not standardized or missing, the image text in the certificate and directory can be used to strengthen the identification evidence, improve the coverage and interpretability of the on-chain enterprise collection, and provide traceable data sources and timestamps for subsequent evidence chain output.

[0037] S3: Perform layout area detection on the image data to obtain text region coordinates, perform OCR on the text region to obtain image text fields, and perform image-text semantic encoding on the image data and image text fields to obtain visual embedding; In some embodiments, the specific embodiment is as follows: layout detection, OCR and image-text semantic encoding of image data (S3), input data and identifier. This embodiment processes the image data of a single enterprise. The image data includes at least: webpage snapshot screenshots (e.g., screenshots of the enterprise's "qualifications / products / services scope" page); qualification certificate images (scanned copies or photos); product catalog PDF page images (page images obtained by rendering the PDF page by page). The system assigns a unique identifier ImageID to each image. If the source is a PDF page, the PDF_ID and PageNo are additionally recorded for subsequent evidence chain reference of page numbers.

[0038] Preprocessing for layout area detection (obtaining text area coordinates) Perform unified preprocessing on the input image: rotation correction (de-skewing), denoising, contrast enhancement, and resolution normalization; for certificate photos, perspective correction can be performed to reduce OCR errors caused by tilting. Layout area detection Execute a system call to the layout area detection model to perform layout analysis on the image, and output multiple region boxes and their categories. The region categories at least include: title / large text area, body paragraph area, table area, list area; for certificate images, it also includes: certificate name area, certificate number area, issuing authority / seal area, date area. Coordinate output For each detected text-related region, output its coordinates in the image (such as the upper-left and lower-right coordinates), forming a TextRegion list: TextRegion = {RegionID, ImageID, PageNo(optional), RegionType, coordinates, confidence}, note (aligned with S7): The above "coordinates" are the "image area coordinates" that need to be output in the subsequent evidence chain, so the system must retain RegionID and coordinates at this step. OCR recognition (obtaining image text fields) OCR input and strategy Crop the region image corresponding to each TextRegion to obtain an ROI image block, and perform OCR recognition on the ROI respectively: for paragraph-type ROIs, perform line text recognition; for table ROIs, first perform table structure splitting (row / column / cell positioning) and then OCR; for areas such as certificate numbers / dates, adopt a "number + symbol" enhanced recognition strategy (such as higher resolution, character set constraint).

[0039] OCR output structure The OCR output not only includes the recognized text, but also includes character-level / word-level positions and confidence, forming an ImageTextField: ImageTextField = {RegionID, recognized text, token sequence, token position box, confidence, text normalization result}, where the "text normalization result" includes: removing extra spaces, unifying full-width and half-width, correcting common misspelled words (such as "回 收→回收"), unifying units and symbols, etc. Note (aligned with S4, S7): The recognized text and token sequence are used as the input of the "image text field" in subsequent S4; the token position box and RegionID are used to output the "word / character position index" in the subsequent evidence chain.

[0040] This embodiment uses a "region-level image-text alignment → enterprise-level aggregation" approach to generate visual embeddings, enabling them to be used for both feature fusion and evidence chain interpretation. The region-level visual vector is input into a visual encoder for each ROI image patch (i.e., the cropped region corresponding to each RegionID) to obtain a region-level visual vector V_region. This vector characterizes visual semantic information such as certificate format, seals / logos, table layout, and combinations of product images and text. The region-level text vector is input into a pre-trained language model using the OCR-recognized text corresponding to the RegionID to obtain a region-level text vector T_region, used to represent semantics such as "recycling processing qualifications, licenses, product models, and technical keywords." The cross-modal alignment fusion system uses a cross-modal alignment network to calculate the semantic consistency between V_region and T_region, and obtains the region-level visual embedding E_region through attention-weighted fusion, while outputting an alignment score (which can be used as one of the sources of confidence for subsequent S5 sub-conditions). The enterprise-level visual embedding aggregation pools all E_regions under the same enterprise according to confidence weights (or groups and pools them according to region type before fusion) to obtain the enterprise-level visual embedding E_enterprise. For example, higher weights are given to the certificate number and certificate name regions; lower weights are given to low-quality screenshot regions. Output: E_region (bound to RegionID, used for evidence chain location and interpretation); E_enterprise (used for S4 fusion to generate enterprise feature matrix).

[0041] After executing S3, the system outputs three types of results for each enterprise and writes them into the enterprise data package: 1) Text Region Coordinate List: TextRegion list (used for evidence chain coordinates); 2) Image Text Field: ImageTextField list (used for S4 text feature and rule matching); 3) Visual Embedding: E_region and E_enterprise (used for S4 fusion and S7 interpretation). For S4: The "image text field" output from S3 is processed using TF-IDF and semantic encoding; E_enterprise is used as a visual embedding and fused with text vectors to generate the enterprise feature matrix. For S5: Region alignment scores and OCR confidence can be used for sub-condition node confidence calculation, improving the stability of the rule expression tree judgment. For S7: RegionID + coordinates + token position directly form the "image region coordinates and text fragment positions" in the evidence chain, enabling traceable interpretation.

[0042] S4: The business scope text field and image text field are subjected to TF-IDF feature extraction and pre-trained language model semantic encoding, and fused with visual embedding to generate an enterprise feature matrix. In some embodiments, the enterprise feature matrix generation and fusion (S4) takes input data (from S2 and S3) for a certain enterprise (EnterpriseID=E001). The system obtains: Business scope text field: original and standardized text of the business scope registered with the industrial and commercial administration (from S2); Image text field: text obtained by OCR of webpage screenshots / certificates / directory pages and its regional identifier (from S3); Visual embedding: enterprise-level visual embedding E_enterprise and optional regional-level visual embedding E_region (from S3 image-text semantic encoding). Text preprocessing and segmentation of business scope text: The business scope text is segmented into multiple short sentences according to delimiters, and word segmentation, stop word removal and synonym unification are performed (e.g., "recycling processing ≈ recycling disposal ≈ resource utilization processing").

[0043] Image text processing groups OCR text by RegionID: certificate name area, certificate number area, product name area, table cell text, etc.; it filters out obviously noisy text (such as watermarks, headers and footers). Table text can be concatenated into structured short sentences by "row / column" (e.g., "product name + model + purpose"). After processing, two types of computable text sets are obtained: BS_TextSet: a set of short sentences on business scope; IMG_TextSet: a set of short sentences on image text (the RegionID can be retained for interpretation).

[0044] The TF-IDF feature extraction (sparse keyword evidence) vocabulary construction system uses a set of keywords from the target process (e.g., "battery recycling") and historical corpora to jointly construct a vocabulary, ensuring that industry terms can be stably extracted. TF-IDF calculation is performed on BS_TextSet and IMG_TextSet respectively: obtaining the TF-IDF vector V_tfidf_bs for the business scope; obtaining the TF-IDF vector V_tfidf_img for the image text; and calculating the "keyword hit strength" index: for example, whether the TF-IDF of words such as "battery recycling / dismantling / hazardous waste permit" exceeds a threshold, which is used in the "keyword hit condition" sub-node of S5.

[0045] The semantic encoding (semantic matching evidence) objects of the pre-trained language model are input into the BS_TextSet and IMG_TextSet respectively to obtain semantic vectors: business scope semantic vector V_sem_bs (which can be obtained by weighted pooling of short sentence vectors to obtain enterprise-level vectors); image text semantic vector V_sem_img (also pooled to obtain enterprise-level vectors). The semantic matching score (serving S5) calculates the similarity between the enterprise semantic vector and the target process semantic vector (generated from the target process description / keyword center vector configured in S1), obtaining: score_sem_bs: business scope semantic similarity score; score_sem_img: image text semantic similarity score; these two scores can be directly used as the input confidence scores for the S5 "semantic matching condition" sub-node.

[0046] The visual embedding fusion (multimodal joint feature) system incorporates the enterprise-level visual embedding E_enterprise output by S3 into the fusion process because visual cues such as certificate format, seal features, and directory page structure can enhance credibility and can be mutually verified with OCR text. In this embodiment, the fusion method employs "gated fusion": first, dimensional alignment (linear projection / normalization) is performed on V_tfidf_bs, V_tfidf_img, V_sem_bs, V_sem_img, and E_enterprise respectively; then, the gating weights g_bs, g_img, and g_vis are calculated, where the gating weights can be determined based on quality indicators such as OCR confidence, image-text alignment score, and text length; finally, the fusion vector F = g_bs·[V_tfidf_bs, V_sem_bs] + g_img·[V_tfidf_img, V_sem_img] + g_vis·E_enterprise is obtained.

[0047] To facilitate subsequent rule expression tree determination, the enterprise feature matrix is ​​constructed and output in a "matrix / vectorized structure" in this embodiment, including at least the following fields: V_tfidf_bs, V_tfidf_img (keyword sparse evidence); score_sem_bs, score_sem_img (semantic matching evidence); E_enterprise (visual evidence); quality features such as quality_ocr (OCR average confidence) and quality_page (page quality score); and timestamp (data update time, used for S6 time decay). This set constitutes the enterprise feature matrix M_enterprise (which can be viewed as a matrix or feature vector expanded according to "feature dimensions"), and is bound to EnterpriseID for output.

[0048] In the next step, S5, the rule expression tree directly reads the key components in M_enterprise as sub-condition inputs. For example, the "keyword hit condition" node reads whether the TF-IDF value of the target keyword in V_tfidf_bs or V_tfidf_img exceeds the threshold; the "semantic matching condition" node reads whether score_sem_bs or score_sem_img exceeds the threshold; the "certificate pattern condition" node reads the regular expression hit flag from IMG_TextSet and can combine it with quality_ocr to improve the confidence. After differentiability gating, the gating parameters and threshold parameters are updated within the sliding window based on the missed detection feedback, thereby synchronously optimizing the weight contribution of the above features in the judgment.

[0049] By extracting "keyword evidence (TF-IDF)" and "semantic evidence (semantic encoding)" from the business scope text and image OCR text respectively, and then fusing them with visual embedding to generate an enterprise feature matrix, this solution can still stably represent the correlation strength between the enterprise and the target link even when the text expression is not standardized, there are synonym substitutions, or the key evidence only exists in the certificate / catalog image. This provides a unified input basis for subsequent rule tree judgment, self-optimization, and evidence chain interpretation.

[0050] S5: Construct a rule expression tree containing AND and OR nodes. The expression tree consists of sub-condition nodes based on TF-IDF weights and semantic similarity scores. Input the enterprise feature matrix into the rule expression tree to perform sub-condition judgment, output candidate enterprises, and calculate the matching degree based on the sub-condition hit confidence. Make the expression tree differentiable and construct a loss function within a sliding window based on the deviation between the number of candidate enterprises and the target size, as well as the missed detection feedback formed by supplementary entries. Adaptively update the keyword weights and threshold parameters through gradient descent. In some embodiments, the rule expression tree judgment and sliding window self-optimization (taking "battery recycling" as an example) input and initialization for the target stage "battery recycling" involves the system reading from the S1 ontology semantic feature space: a set of business scope keywords and initial weights (e.g., "battery recycling, dismantling, cascade utilization, regeneration, resource utilization," etc.); a set of image and text keywords / patterns (e.g., "hazardous waste operation license, recycling and processing qualification, license number pattern," etc.); and initial thresholds: TF-IDF threshold θ_tfidf, semantic similarity threshold θ_sem, and regularization hit threshold. For each enterprise, S4 outputs the enterprise feature matrix M_enterprise, which includes at least: a business scope TF-IDF vector, an image and text TF-IDF vector; a business scope semantic similarity score score_sem_bs, an image and text semantic similarity score score_sem_img; OCR quality score, and image-text alignment score (which can be used as a confidence weighting factor), etc.

[0051] The system constructs a rule expression tree (AND / OR structure). In this embodiment, the system constructs the following expression tree (example structure) for the "battery recycling process": Root node: AND; Subtree A: OR ("Business relevance" must satisfy one of the criteria); Node A1: TF-IDF matching condition for business scope keywords; Decision logic: If the weighted sum of keywords such as "battery recycling / dismantling / recycling" in the business scope TF-IDF vector exceeds θ_tfidf, then it is a match. Node A2: Semantic matching condition for business scope; Decision logic: If score_sem_bs exceeds θ_sem, then it is a match. Node A3: Semantic matching condition for image and text; Decision logic: If score_sem_img exceeds θ_sem, then it is a match. Subtree B: OR ("Evidence strengthening" must satisfy one of the criteria); Node B1: Certificate / License pattern matching condition; Decision logic: If the image and text field regular expression matches "Hazardous waste operation license / recycling and processing qualification / license number pattern", then it is a match. Node B2: Key Product Catalog Item Hit Decision Logic: The image text contains keywords such as "recycling / dismantling line / black powder treatment / recycled materials" for products or processes, and the TF-IDF exceeds the threshold. Explanation: This structure reflects "relevance (A) × evidence reinforcement (B)," avoiding false inclusions due to vague terms in the business scope alone, while using certificate / catalog image evidence to supplement missed detections.

[0052] For each enterprise, its M_enterprise is input into an expression tree, and the hit confidence score is output node by node (not a simple 0 / 1, but a confidence level between 0 and 1). For example: A1 confidence score: obtained by mapping the difference between the "weighted sum of keyword TF-IDF and the threshold"; the larger the difference, the higher the confidence score. A2 / A3 confidence scores: obtained by mapping the semantic similarity score to the threshold difference. B1 confidence score: high if the regular expression hits, and can be reduced by the OCR quality score (lower confidence score when OCR quality is low). B2 confidence score: determined by the hit strength (TF-IDF) of the directory entry and the OCR quality. Then, AND / OR nodes are aggregated according to rules: OR node: "any strong evidence in the sub-condition" will result in a higher confidence score for OR output. AND node: both subtrees need to be relatively credible for the root node to output a high confidence score. The confidence level output by the final root node is denoted as the matching degree M of the enterprise to the target stage, and the candidate admission rule is set: if M ≥ θ_accept (e.g. 0.70), the output is a candidate enterprise; at the same time, "which sub-conditions are hit, the score of each sub-condition, the threshold, and the corresponding evidence position" are recorded to serve the subsequent S7 evidence chain.

[0053] To enable automatic updates to keyword weights and thresholds, this embodiment changes each sub-condition node from "hard decision" to "differentiable gating": the sub-condition output is obtained by smoothing continuous quantities such as "TF-IDF weighted sum / semantic similarity score" (so that gradient signals can be generated near the threshold); AND / OR nodes no longer use hard logic short-circuiting, but use continuous gating aggregation, making the entire tree numerically differentiable. In this way, keyword weights and threshold parameters (including θ_tfidf, θ_sem, θ_accept) become trainable parameters.

[0054] The sliding window and loss function are constructed (scale bias + missed detection feedback). A sliding window is set, for example, "the 1000 companies processed in the most recent round" or "data batches from the last 7 days". Within each sliding window, the following statistics are calculated: the current number of candidate companies N_actual (the number of companies with a matching degree ≥ the admission threshold); the business target scale N_target (e.g., maintaining the number of candidate companies in the battery recycling segment at around 500); and the manually added (missed detection) set S_add: a list of companies confirmed by business personnel as "should have been selected but were not". In this embodiment, the loss function consists of two parts: scale bias loss: constraining N_actual to move closer to N_target, avoiding too many or too few candidate companies; and missed detection penalty loss: imposing a penalty on the added companies S_add, prompting these companies to improve their matching degree above the admission threshold. Intuitively, the system aims for both "controllable scale" and "minimizing the omission of added companies," with both factors driving the threshold and weights towards a more reasonable direction.

[0055] Gradient descent update and write-back (keyword weights, threshold parameters): At the end of each sliding window, the system performs one or more iterative updates to the parameters based on the above loss: Update content: keyword weights (business scope keyword weights, image text keyword weights), semantic threshold θ_sem, TF-IDF threshold θ_tfidf, admission threshold θ_accept, and gating parameters; Update constraints: The threshold is kept within a preset reasonable range (e.g., 0.5~0.95), the weights are kept non-negative and normalizable to prevent parameter drift from causing the rules to go out of control; Update result write-back: The new weights / thresholds are written back to the "version configuration" of the semantic feature space of the target stage, and a version number is generated for reuse and rollback in subsequent batches.

[0056] In one embodiment, during a certain sliding window: N_target=500, the system output N_actual=420 (too low); business personnel added 20 companies, many of which have broad business scopes, but their qualification certificate images clearly show the words "recycling and processing qualification / license," and their directory pages contain "disassembly and processing technology." The system's update strategy might be as follows: appropriately lower the threshold or increase the weight of the sub-condition related to "certificate pattern hit" (because it is more crucial for identifying missed companies); not significantly increase the weight of "generalized terms for business scope" (to avoid introducing a large number of false inclusions); moderately lower the admission threshold θ_accept or adjust the AND / OR gating parameters so that "strong evidence (certificates / directories)" contributes more to the final matching score. After the update, similar companies are more likely to be included in the next sliding window, thus reducing duplicate entries.

[0057] Through a closed-loop mechanism of "rule expression tree + differentiable gating + sliding window size constraint + supplementary recording and feedback of missed detections", this embodiment achieves: the rule determination process is interpretable and traceable (each candidate enterprise has hit sub-conditions and evidence positions); parameters can be self-optimized, reducing repeated manual parameter tuning; and missed detections can be minimized under the premise of "controllable scale", thereby improving the quality and stability of enterprise collection in the industrial chain.

[0058] S6: Construct an enterprise relationship graph based on candidate enterprises and calculate PageRank centrality. Combine centrality, matching degree and time decay coefficient based on enterprise data update time to obtain enterprise weights and output the ranking. In some embodiments, the enterprise relationship graph construction is integrated with PageRank ranking. Input data (timestamps from S5 and S2 / S3) is used for the target stage (e.g., "battery recycling stage"). In step S5, the system has already obtained: a candidate enterprise set C = {E1, E2, …, En}; and the matching degree M(Ei) of each candidate enterprise (obtained from the confidence score of the root node of the rule expression tree). Simultaneously, the system reads the data update time t_update(Ei) of each enterprise from the enterprise data packet. This timestamp can be the latest value of the business scope collection time, the webpage / certificate / directory image collection time, or a weighted summary value, used for subsequent time decay.

[0059] The enterprise relationship graph (Graph) is constructed by using each enterprise in the candidate enterprise set as a node, forming a node set V=C. The edges and edge weights of the graph are defined (this embodiment uses "multi-source relationship fusion edge construction") to better align with the semantics of the industry chain. This embodiment establishes three types of relationship edges between candidate enterprises and assigns weight factors to each type: Text-Sim Edge: Edge construction is based on the semantic vector similarity between the enterprise's business scope text field and the image text field; generation method: if the semantic similarity between two enterprises is higher than a threshold (e.g., 0.75), an edge is established between the two enterprises; edge weight: the higher the similarity, the greater the edge weight. Keyword Co-occurrence Edge: Edge construction is based on the overlap of the high-weight keyword hit sets of the two enterprises in the "target link keyword set"; generation method: calculate the overlap of the keyword hit sets of the two enterprises, and if it exceeds a threshold, an edge is constructed; edge weight: the greater the overlap, the greater the edge weight.

[0060] Evidence type consistency is used to construct edges based on similar image evidence types, such as certificates / catalogs, between the two companies (e.g., both matching the "license / qualification" pattern or both matching the "disassembly process / equipment" entry). Generation method: Edges are constructed if the evidence type consistency exceeds a threshold. Edge weight: Higher consistency results in higher edge weights, which can be reduced using OCR quality adjustments. Finally, these edges are merged into a weighted graph G=(V,E,W), where each edge weight is the weighted sum of weights from multiple sources, ensuring the graph structure reflects both "semantic proximity" and "evidence similarity." Explanation: This graph construction method reduces noisy connections caused by relying solely on a single indicator, allowing PageRank to more stably reflect the status of "key node companies" within the industry chain.

[0061] To calculate PageRank centrality, the system performs PageRank calculations on the constructed weighted firm relationship graph, obtaining a centrality score C(Ei) for each candidate firm. A firm receives a higher PageRank score if it is connected to more "high-quality evidence firms" and its connection edge weights are higher; firms with lower edge weights or that are isolated receive relatively lower scores. The system can normalize the PageRank scores (e.g., map them to the range of 0-1) to facilitate integration with matching degree and time decay.

[0062] To reflect "data freshness," this embodiment calculates the time decay coefficient T(Ei) based on the enterprise's data update time: Let the current batch calculation time be T_now; let Δt = T_now - t_update(Ei) represent the degree of data staleness; the larger Δt is, the smaller the time decay coefficient; the smaller Δt is (the newer the data), the closer the coefficient is to 1. The update time t_update(Ei) can adopt one of the following strategies: take the enterprise's most recent collection time (the latest time from any source such as business scope / webpage screenshot / certificate / directory); or take the weighted update time (for example, certificate / directory update has a higher weight than webpage screenshot).

[0063] The system merges three indicators to obtain enterprise weights and outputs a ranking. The enterprise weights W(Ei) are: Matching degree M(Ei): reflecting the "rule hit strength" between the enterprise and the target stage; Centrality C(Ei): reflecting the "structural importance" of the enterprise in the candidate enterprise network; and Time decay T(Ei): reflecting the "timeliness reliability" of the data. The fusion strategy can employ weighted fusion or gated fusion. In this embodiment, weighted fusion with normalization is used: given fusion weights (e.g., matching degree is more important, centrality is secondary, and timeliness is used as a reduction factor), the comprehensive enterprise weight is calculated; the candidate enterprises are sorted from high to low according to their comprehensive weights, and a Top-K enterprise list is output. The output results must include at least: EnterpriseID, comprehensive weight, matching degree, centrality, and update time / decay coefficient, for easy subsequent interpretation and auditing.

[0064] The connection with preceding and following steps (ensuring logical closure) and with S5: the matching degree of S5 output is directly incorporated into the fusion; at the same time, the evidence types matched by S5 can be used for the "evidence type consistency edge" in the graph construction. Connection with S7: the ranking output can be used as an online display or a key analysis object; the evidence chain can be attached to the output to explain "why the company ranks so high" (high matching degree, high centrality, fresh data, certificate hit, etc.).

[0065] Within the same batch, company A has a slightly higher match with the target link, but its evidence is outdated and its connections with other companies are weak. Company B has a slightly lower match, but it holds key license evidence, has close connections with multiple companies with high-evidence data, and its data is more recent. Sorting solely by match might place A higher; however, by using the fusion method described in this embodiment, B's centrality and timeliness advantages will increase its overall weight, making the ranking more aligned with the needs of identifying key nodes in the industry chain and reducing the number of companies with "textual hits but weak actual relevance" occupying the top positions.

[0066] S7: A double-buffered update mechanism is adopted to perform full periodic updates and / or event-driven incremental updates in the computation buffer, and the online buffer is replaced by atomic switching. A term co-occurrence network is constructed based on the business scope text field and the image text field. The keyword system is dynamically expanded through Louvain community discovery and pre-trained language model semantic drift detection, and the keyword set of the semantic feature space and rule expression tree is updated accordingly. The output includes the evidence chain containing hit sub-conditions, image region coordinates and text fragment positions.

[0067] In some embodiments, the system operates with a dual-buffered dynamic update, terminology evolution expansion, and evidence chain output. The dual-buffered structure involves maintaining two parallel data and model configuration buffers: an online buffer (providing query and sorting services, including the currently active industry chain ontology version, keyword system, rule expression tree parameters, enterprise feature library, candidate enterprise list, and sorting results); and a computation buffer (performing update calculations, containing copies of the ontology configuration to be updated, keywords and threshold parameters to be updated, and intermediate results such as enterprise features and relationship graphs to be updated). The online buffer remains consistently available; all update calculations are performed only in the computation buffer, preventing the update process from affecting online response.

[0068] The full update implementation mechanism triggers a full update at a fixed period, such as once a day at midnight or once a week at a fixed time. After triggering, the following processes are executed in the computation buffer: S2 data collection is re-executed: the latest business scope text field, the latest webpage snapshot screenshot, the latest qualification certificate image, and the latest product catalog PDF page image are retrieved; S3 is re-executed: layout detection, OCR, and image-text semantic encoding are performed on the newly added or changed image data to obtain visual embeddings; S4 is re-executed: TF-IDF, semantic encoding, and visual embedding are fused to generate an enterprise feature matrix; S5 / S6 is re-executed: candidate enterprises, matching degree, enterprise relationship graph PageRank centrality, and enterprise weight ranking results are generated; the "term co-occurrence network" and "evidence chain index" for this batch are generated simultaneously. After the full update is completed, the computation buffer forms a complete online version (including ontology configuration, keyword system, rule tree parameters, enterprise library, and ranking results). The event-driven incremental update implementation involves the system subscribing to enterprise information change events or data collection events. These events include, but are not limited to: changes in the enterprise's business scope; updates to the enterprise's official website pages (such as discovering new product pages / new qualification pages); the addition of qualification certificates or the replacement of expired certificates; and the addition of product catalog PDFs or catalog version updates.

[0069] When an event is triggered by a company, the incremental process is executed only for that company in the computation buffer: If it is a change in business scope: update the business scope text field and recalculate TF-IDF / semantic encoding; if it is a new certificate / directory: perform S3 on the new image and update the visual embedding and image text fields; update the company's feature matrix and re-determine its candidate status and matching degree on the rule expression tree; if the candidate set changes, perform a partial update on the company relationship graph (only update the adjacent edges related to the company), and partially update PageRank or adopt an approximate update strategy; update the company's position in the ranking results and record the update timestamp. Through event-driven incremental updates, the system can quickly reflect changes in key company information without waiting for the full cycle. Atomic Switch is implemented when the compute buffer has completed a full update or accumulated a certain amount of incremental updates. The system performs an atomic switch: the online buffer pointer is switched to the completed version in one go. The switch does not replace data item by item, but rather switches on a version number basis, ensuring data consistency provided to the public at any given time. After the switch, the original online version is retained for a period as a rollback version for rapid revert in case of anomalies. This ensures online service stability and data consistency.

[0070] In this embodiment, the terminology co-occurrence network uses two types of text as corpus: business scope text fields and image text fields (from certificates, webpage screenshots, and product catalog OCR). The system performs word segmentation and terminology normalization on the above text for each enterprise, extracting noun phrases, product terms, and process terms as "terminology nodes"; co-occurrence edges are established between terms based on "co-occurrence within the same enterprise text window" or "co-occurrence within the same paragraph / table row", and the co-occurrence frequency or point mutual information is used as the edge weight to form a terminology co-occurrence network.

[0071] The Louvain community discovery and semantic drift detection (dynamically expanding the keyword system) runs the Louvain community discovery algorithm on the term co-occurrence network, resulting in multiple term communities. The system focuses on communities that are adjacent to or overlap with existing keyword communities in the target segment, considering them as candidate expansion communities. For terms within candidate expansion communities, the system uses a pre-trained language model to obtain semantic vectors and calculates their similarity to the semantic center of the target segment's keywords. If a term meets the following criteria: its semantic similarity to the target segment exceeds a drift threshold (e.g., 0.75), and its co-occurrence frequency in the nearest window shows an upward trend (indicating an emerging term), then the term is identified as an expandable term. For expandable terms, the system performs keyword system updates: the term is written into the "business scope keyword set" or "image and text keyword set" in the semantic feature space of the target segment; weights are initialized for the new term (e.g., initial weights are given based on a combination of similarity and co-occurrence strength); and the keyword set and sub-condition template of the rule expression tree are updated synchronously (e.g., the term is added to the keyword hit condition). In this way, the next round S4 / S5 can automatically cover the industrial evolution brought about by the new term.

[0072] This embodiment generates a structured evidence chain for each output candidate company to explain its selection reasons and support audit traceability. The evidence chain includes at least: 1) Hit sub-condition information: hit sub-condition number, sub-condition type (TF-IDF / semantic matching / regular expression pattern), threshold, score and confidence; 2) Text fragment position: for hits in the business scope, record the start and end positions or sentence number of the hit text fragment in the business scope field; 3) Image evidence position: for hits in image text, record the image file identifier, page number, text region coordinates (from S3 layout detection output), and OCR word / character position index; Version and time information: ontology version number, keyword system version number, data update time, used to align with the online cache version. For example, if a company is selected because "the certificate image matches the 'Hazardous Waste Management Permit' pattern", then the evidence chain records: Hit sub-condition: certificate pattern condition, threshold = hit means satisfied, score = 1; image location: page 2 of the certificate PDF, coordinates (x1, y1, x2, y2); OCR location: the kth word in the token sequence; and also includes the version number and timestamp.

[0073] Through double buffering and atomic switching, the system achieves "isolation between update computation and online service," ensuring consistency and stability. Full-scale and event-driven incremental updates balance timeliness and cost. Through terminology co-occurrence networks, Louvain community discovery, and semantic drift detection, the system can automatically capture new terms and write back rules and semantic feature spaces, dynamically optimizing the industry chain definition as the industry evolves. Through evidence chain output, the system makes the reasons for candidate company selection explainable and traceable, providing a basis for subsequent manual review, supplementary feedback, and auditing.

[0074] Preferably, in step S3, after OCR, field extraction is performed on the image text field to generate a structured field set. The structured field set includes the enterprise name, unified social credit code, certificate name, certificate number, product name, and key technical keywords, and the structured field set is incorporated into the enterprise feature matrix.

[0075] Preferably, in step S4, TF-IDF vectors and semantic encoding vectors are output for the business scope text field and the image text field, respectively, and then fused with the visual embedding in a way that involves concatenation, weighted summation, or gated fusion to form an enterprise feature matrix, wherein the fusion weight is adaptively adjusted by the matching degree feedback.

[0076] Preferably, the sub-condition nodes in step S5 include: a) keyword hitting conditions based on the TF-IDF weight threshold of the business scope text field; b) semantic matching conditions based on the semantic encoding vector similarity threshold; c) regular expression matching conditions based on the certificate number and product name pattern in the image text field; and AND nodes are used to ensure that multiple conditions are met simultaneously, and OR nodes are used to ensure that one of the multiple conditions is met.

[0077] Preferably, the differentiability gating of the expression tree in step S5 includes replacing the AND node with a multiplicative gating function, replacing the OR node with an additive gating function or a LogSumExp gating function, and using the confidence level output by the sub-condition node as the gating input, so that the keyword weight, threshold parameter and gating parameter can be jointly updated through gradient descent, thereby improving the rule convergence speed and reducing manual parameter tuning.

[0078] Preferably, the missed detection feedback in step S5 is generated by an active learning strategy. The active learning strategy includes: selecting companies whose matching degree is in a preset uncertainty range as supplementary candidates from unselected companies, and prioritizing the push of samples with the highest semantic similarity to the target link keywords and the largest confidence fluctuation for manual confirmation, so as to feed the confirmation results back into the loss function as missed detection feedback.

[0079] Preferably, the evidence chain in step S7 further includes a cross-modal alignment index, which includes: file identifier and page number of image data, text region coordinates, OCR character-level or word-level position index, corresponding business scope text fragment position index, and sub-condition node identifier and threshold that trigger the selection of the candidate enterprise, so that each enterprise selection can be traced back to a specific image region and a specific text fragment.

[0080] Preferably, in step S1, the three-level industry chain ontology constructed based on OWL includes a three-layer structure of industry, chain segment, and link, and a computable semantic feature space is defined in the ontology for the target link. The semantic feature space includes: (i) The set of keywords and their initial weights for the business scope text field; (ii) The set of keywords and their initial weights for the image text field; (iii) A sub-condition template for a rule expression tree, wherein the sub-condition template includes at least a TF-IDF threshold condition and a semantic similarity threshold condition; (iv) The field mapping rules of the evidence chain are used to establish a correspondence between the keywords of the target link and the text region coordinates output in step S3 and the semantic encoding vector output in step S4; The semantic feature space serves as the configuration input for constructing the rule expression tree in step S5, and as the update object when dynamically expanding the keyword system in step S7.

[0081] Preferably, the visual embedding is obtained by performing image-text semantic encoding on the image data and the image text field, including: a) Crop the image sub-regions corresponding to the text region coordinates obtained in step S3 into several ROI image blocks, and input each ROI image block into the visual encoder to obtain the ROI visual vector; b) Segment the image text field by word or phrase and input it into a pre-trained language model to obtain text vectors; c) A cross-modal alignment network is used to calculate the similarity between the ROI visual vector and the text vector and perform attention-weighted fusion to obtain a region-level visual embedding that corresponds one-to-one with the coordinates of each text region. The region-level visual embedding is then pooled to obtain an enterprise-level visual embedding. The regional visual embedding serves as the input for the enterprise feature matrix generated in step S4, and is bound to the text region coordinates when the evidence chain is output in step S7.

[0082] This application also provides a data intelligence construction and dynamic optimization system for the industrial chain, such as... Figure 2The system hardware components shown include: Data acquisition and access side hardware: a data acquisition server / acquisition node cluster, and a multi-core CPU server for web scraping, PDF downloading, image acquisition, and API retrieval of business scope text, etc. Multiple acquisition nodes can be configured for parallel scraping and distributed scheduling. Network access devices include routers / firewalls / switches for external network access to the company's official website and third-party data source interfaces; a reverse proxy or gateway (optional) is used for unified request exit, rate limiting, and auditing. Computation and inference side hardware includes a computing server cluster (CPU computing nodes) for feature computation (TF-IDF), rule expression tree determination, graph construction and PageRank, term co-occurrence network construction, Louvain community discovery, etc. AI acceleration servers (GPU / AI card nodes) are used for page area detection model inference, OCR model inference, pre-trained language model semantic encoding, and image-text semantic encoding / cross-modal alignment network inference; multiple GPU cards can be configured to support batch processing and low-latency services.

[0083] Storage and indexing hardware includes: object storage servers / distributed file storage for large files such as webpage snapshots, certificate images, PDF files, and PDF page rendering images; relational database servers (RDBMS) for storing basic enterprise information, business scope text fields, structured field sets, task queue status, version numbers, and metadata (collection timestamps, source URLs, etc.); graph database servers or graph computing storage for storing the edges and weights of the enterprise relationship graph, adjacency lists, or persistent storage using a graph computing engine; retrieval / indexing servers (search engines) for building inverted indexes of OCR text and business scope text; supporting fast evidence chain location and querying; caching servers (e.g., memory cache) for supporting fast read / write of online buffers, hot enterprise queries, and Top-K sorting result caching; and online service and operations hardware, including application service servers (API / business services) providing interfaces for enterprise selection queries, sorting queries, and evidence chain display, and carrying online buffer versions. Message queue servers for handling event-driven incremental updates: business scope change events, webpage update events, new certificate / directory events, etc.; used to decouple collection and computation. The management console / operations terminal is used to configure the industry chain entity, target link parameters, threshold ranges, task scheduling strategies, version rollback, and auditing.

[0084] Overall network connectivity: Servers are interconnected via LAN switches to form the data center's internal network; data acquisition servers connect to the external internet / third-party data source networks via routers / firewalls; application service servers provide HTTP / HTTPS interfaces externally via reverse proxy / load balancing devices. Acquisition Server → Object Storage / Database: The acquisition server retrieves business scope text, webpage screenshots, certificate images, and PDF files from the external network; text and metadata are written to a relational database; images / PDFs are written to object storage; and "add / change event messages" (for incremental updates) are written to the message queue. Computing Server / GPU Server → Storage and Indexing: The computing server reads the enterprise list and text from the database and images / PDF page images from object storage; the GPU server performs layout detection, OCR, semantic encoding / graphic-text encoding inference, and writes the OCR text, visual embeddings, and semantic vectors back to the database / vector storage (which can be categorized into the database or a dedicated vector library); the computing server performs TF-IDF, rule tree determination, self-optimizing training (gradient descent parameter updates), enterprise relationship graph construction, and PageRank, writing the results to the database and graph database; term co-occurrence network and Louvain results are written to the database, and the ontology / parameter version is updated. The evidence chain index (coordinates, page number, token position, etc.) is written to the index server or the database evidence table. Online service server ↔ Cache / Database / Object Storage; The online service server reads candidate companies, ranking results, and evidence chain indexes from the "online buffer version" (preferably from the cache); retrieves the corresponding image / PDF page through object storage; and returns the company weight ranking and evidence chain to the outside world. Hardware implementation of double buffering and atomic switching: Online buffer: deployed in the memory of the cache server / online service server, or resides in the database as an "online version table / online index"; Computation buffer: deployed on the computation server side, with computation results written to the "computation version table / computation index"; Atomic switching: the management console or scheduling service performs a "version pointer switch" in the database (e.g., switching active_version from v1 to v2), and the online service server only reads the data partition corresponding to active_version, thus achieving second-level consistent switching; rollback to the previous version is possible if necessary.

[0085] The system comprises a data acquisition server, a computing server, an AI acceleration server, an object storage server, a relational database server, a graph database server, a cache server, and an application service server. The data acquisition server communicates with external data sources via network devices and with both the object storage server and the relational database server to store image data, PDF data, and business scope text fields. The computing server and the AI ​​acceleration server are interconnected via a local area network. The computing server reads data from the object storage server and the relational database server, while the AI ​​acceleration server performs layout detection, OCR, and image-text semantic encoding on the image data and writes the results back to the database. The computing server further constructs an enterprise feature matrix and a rule expression tree, performs self-optimization, constructs an enterprise relationship graph, and calculates PageRank, storing candidate enterprises, matching degrees, centrality, and ranking results in the database and the graph database. The application service server is interconnected with the cache server and the database server and provides query interfaces. The application service server outputs enterprise ranking and evidence chains based on the online cache version. The system triggers updates in the computing cache through the management console and replaces the online cache using atomic version pointer switching.

[0086] The module constructs a three-tier industry chain ontology based on OWL and defines the semantic feature space of the target links; The data collection module collects enterprise data, including the business scope text field, as well as image data such as screenshots of enterprise web pages, images of enterprise qualification certificates, and PDF pages of product catalogs. The extraction module performs layout region detection on the image data to obtain text region coordinates, performs OCR on the text region to obtain image text fields, and performs image-text semantic encoding on the image data and image text fields to obtain visual embedding; The enterprise feature matrix generation module performs TF-IDF feature extraction and pre-trained language model semantic encoding on the business scope text field and image text field respectively, and then fuses them with visual embedding to generate an enterprise feature matrix. A rule expression tree module containing AND and OR nodes is constructed. The expression tree consists of sub-condition nodes based on TF-IDF weights and semantic similarity scores. The enterprise feature matrix is ​​input into the rule expression tree for sub-condition judgment, and candidate enterprises are output. The matching degree is calculated based on the sub-condition hit confidence. The expression tree is made differentiable and gated. Within a sliding window, a loss function is constructed based on the deviation between the number of candidate enterprises and the target size, as well as the missed detection feedback formed by supplementary entries. The keyword weights and threshold parameters are adaptively updated through gradient descent. The ranking module constructs an enterprise relationship graph based on candidate enterprises and calculates PageRank centrality. It then integrates centrality, matching degree, and time decay coefficient based on enterprise data update time to obtain enterprise weights and outputs the ranking. The output module employs a double-buffered update mechanism to perform periodic full updates and / or event-driven incremental updates in the computation buffer, and replaces the online buffer with an atomic switching method. It constructs a term co-occurrence network based on the business scope text field and the image text field, dynamically expands the keyword system through Louvain community discovery and pre-trained language model semantic drift detection, and updates the keyword set of the semantic feature space and rule expression tree accordingly. The output includes a chain of evidence containing hit sub-conditions, image region coordinates, and text fragment positions.

[0087] Compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) This application uses the OWL three-level industry chain ontology as a unified semantic framework, explicitly solidifying the keyword set, sub-condition templates, and evidence mapping rules of the target link, making the enterprise aggregation criteria configurable, reusable, and auditable. Compared with single-source matching that relies solely on business scope or industry code, this application also introduces image data such as webpage snapshots, qualification certificates, and product catalog PDFs. It locates text regions through layout area detection, extracts image text fields through OCR, and further performs image-text semantic encoding to obtain visual embedding. This is then fused with the TF-IDF features of the business scope text and the semantic vectors of the pre-trained language model to generate an enterprise feature matrix. As a result, it can reliably identify enterprises on the chain even when unstructured evidence such as certificate numbers, product names, and key technical terms are present, significantly reducing missed detections caused by missing text, synonym substitution, or non-standard expressions. At the same time, it provides an evidence chain for each hit, outputting the coordinates of the image region and the position of the text fragment, enhancing the interpretability and traceability of the results.

[0088] (2) This application constructs a rule expression tree containing AND / OR nodes, expressing sub-conditions such as TF-IDF weight threshold, semantic similarity threshold, and regular pattern matching in a combinable form, realizing collaborative judgment between "interpretable rules" and "semantic models"; further, the expression tree is made differentiable and gated, mapping the AND / OR logic to a trainable gate function, so that keyword weights, threshold parameters, and gate parameters can be jointly updated through gradient descent within a sliding window. The system uses the deviation between the number of candidate enterprises and the target size as constraints, and introduces the missed detection feedback formed by supplementary recording as a supervision signal to construct a loss function, thereby automatically finding the optimal balance between scale control and identification accuracy, avoiding the problems of long cycles and poor reusability caused by traditional rule schemes relying on repeated trial and error by experts and manual parameter tuning. Combined with the active learning supplementary recording mechanism that prioritizes the confirmation of samples in uncertain intervals, higher quality feedback can be obtained with less manual annotation, further reducing maintenance costs and improving rule convergence efficiency and stability.

[0089] (3) This application constructs an enterprise relationship graph based on the candidate enterprise set and calculates PageRank centrality. It integrates centrality, matching degree, and time decay coefficient based on update time to form enterprise weight, realizing a three-dimensional ranking of "hit degree - network position - timeliness". Compared with ranking based on a single matching score or static indicators, this scheme can highlight key node enterprises in the industrial chain, reduce the risk of "weakly related but text-hit" noisy enterprises ranking high, and improve the reliability of key enterprise discovery. At the same time, a double-buffer update mechanism is adopted to perform full-cycle updates and / or event-driven incremental updates in the calculation buffer, and the online buffer is replaced by atomic switching to ensure decoupling between online query and back-end calculation, avoiding inconsistencies or service jitter introduced during the update process. Combined with the dynamic expansion of the keyword system of term co-occurrence network, Louvain community discovery and semantic drift detection, it can continuously track the evolution of industrial technology and write back the expansion results to the semantic feature space and rule expression tree to ensure that the industrial chain caliber and data results maintain timeliness and consistency in the long term.

[0090] Those skilled in the art will understand that embodiments of this application may be provided as methods, systems, or computer program products, and therefore this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0091] While the present invention has been disclosed above, it is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A method for intelligent construction and dynamic optimization of supply chain data, characterized in that, Including the following steps: S1: Construct a three-level industry chain ontology based on OWL and define the semantic feature space of the target links; S2: Collect enterprise data, which includes the business scope text field, as well as image data such as enterprise webpage snapshots, enterprise qualification certificate images, and product catalog PDF page images; S3: Perform layout area detection on the image data to obtain text area coordinates, perform OCR on the text area to obtain image text fields, and perform image-text semantic encoding on the image data and image text fields to obtain visual embedding; S4: Extract TF-IDF features from the business scope text field and the image text field respectively, and encode them semantically using a pre-trained language model. Then, fuse them with the visual embedding to generate an enterprise feature matrix. S5: Construct a rule expression tree containing AND and OR nodes. The expression tree consists of sub-condition nodes based on TF-IDF weights and semantic similarity scores. Input the enterprise feature matrix into the rule expression tree to perform sub-condition judgment, output candidate enterprises, and calculate the matching degree based on the sub-condition hit confidence. Make the expression tree differentiable and construct a loss function within a sliding window based on the deviation between the number of candidate enterprises and the target size, as well as the missed detection feedback formed by supplementary entries. Adaptively update the keyword weights and threshold parameters through gradient descent. S6: Construct an enterprise relationship graph based on candidate enterprises and calculate PageRank centrality. Combine centrality, matching degree and time decay coefficient based on enterprise data update time to obtain enterprise weights and output the ranking. S7: A double-buffered update mechanism is adopted to perform full periodic updates and / or event-driven incremental updates in the computation buffer, and the online buffer is replaced by atomic switching. A term co-occurrence network is constructed based on the business scope text field and the image text field. The keyword system is dynamically expanded through Louvain community discovery and pre-trained language model semantic drift detection, and the keyword set of the semantic feature space and rule expression tree is updated accordingly. The output includes the evidence chain containing hit sub-conditions, image region coordinates and text fragment positions.

2. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, In step S3, after OCR, field extraction is performed on the image text field to generate a structured field set. The structured field set includes the enterprise name, unified social credit code, certificate name, certificate number, product name, and key technical keywords, and the structured field set is incorporated into the enterprise feature matrix.

3. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, In step S4, TF-IDF vectors and semantic encoding vectors are output for the business scope text field and image text field, respectively, and then fused with the visual embedding in the manner of concatenation, weighted summation or gated fusion to form an enterprise feature matrix, wherein the fusion weight is adaptively adjusted by the matching degree feedback.

4. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, The sub-condition nodes in step S5 include: a) keyword matching conditions based on the TF-IDF weight threshold of the business scope text field; b) semantic matching conditions based on the semantic encoding vector similarity threshold; c) regular expression matching conditions based on the certificate number and product name pattern in the image text field; and AND nodes are used to ensure that multiple conditions are met simultaneously, and OR nodes are used to ensure that one of the multiple conditions is met.

5. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, Step S5, which involves making the expression tree differentiable, includes replacing the AND node with a multiplicative gate function, the OR node with an additive gate function or a LogSumExp gate function, and using the confidence score output by the sub-condition node as the gate input. This allows the keyword weights, threshold parameters, and gate parameters to be jointly updated through gradient descent, thereby improving the rule convergence speed and reducing manual parameter tuning.

6. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, The missed detection feedback mentioned in step S5 is generated by an active learning strategy. The active learning strategy includes: selecting companies whose matching degree is in a preset uncertainty range from the unselected companies as supplementary candidates, and prioritizing the push of samples with the highest semantic similarity to the target link keywords and the largest confidence fluctuation for manual confirmation, so as to feed the confirmation results back into the loss function as missed detection feedback.

7. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, The evidence chain in step S7 further includes a cross-modal alignment index, which includes: file identifier and page number of image data, text region coordinates, OCR character-level or word-level position index, corresponding business scope text fragment position index, and sub-condition node identifier and threshold that trigger the selection of the candidate enterprise, so that each enterprise selection can be traced back to a specific image region and a specific text fragment.

8. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, In step S1, the three-tiered industry chain ontology constructed based on OWL includes a three-layered structure of industry, chain segment, and link, and a computable semantic feature space is defined in the ontology for the target link. The semantic feature space includes: (i) The set of keywords and their initial weights for the business scope text field; (ii) The set of keywords and their initial weights for the image text field; (iii) A sub-condition template for a rule expression tree, wherein the sub-condition template includes at least a TF-IDF threshold condition and a semantic similarity threshold condition; (iv) The field mapping rules of the evidence chain are used to establish a correspondence between the keywords of the target link and the text region coordinates output in step S3 and the semantic encoding vector output in step S4; The semantic feature space serves as the configuration input for constructing the rule expression tree in step S5, and as the update object when dynamically expanding the keyword system in step S7.

9. The method for intelligent construction and dynamic optimization of industrial chain data according to claim 1, characterized in that, Visual embeddings are obtained by performing image-text semantic encoding on image data and image text fields, including: a) Crop the image sub-regions corresponding to the text region coordinates obtained in step S3 into several ROI image blocks, and input each ROI image block into the visual encoder to obtain the ROI visual vector; b) Segment the image text field by word or phrase and input it into a pre-trained language model to obtain text vectors; c) A cross-modal alignment network is used to perform similarity calculation and attention-weighted fusion on the ROI visual vector and the text vector to obtain a region-level visual embedding that corresponds one-to-one with the coordinates of each text region, and the region-level visual embedding is pooled to obtain an enterprise-level visual embedding. The regional visual embedding serves as the input for the enterprise feature matrix generated in step S4, and is bound to the text region coordinates when the evidence chain is output in step S7.

10. A data-driven intelligent construction and dynamic optimization system for the industrial chain, characterized in that: include: The module constructs a three-tier industry chain ontology based on OWL and defines the semantic feature space of the target links; The data collection module collects enterprise data, including the business scope text field, as well as image data such as screenshots of enterprise web pages, images of enterprise qualification certificates, and images of product catalog PDF pages. The extraction module performs layout region detection on the image data to obtain text region coordinates, performs OCR on the text region to obtain image text fields, and performs image-text semantic encoding on the image data and image text fields to obtain visual embedding; The enterprise feature matrix generation module performs TF-IDF feature extraction and pre-trained language model semantic encoding on the business scope text field and image text field respectively, and then fuses them with visual embedding to generate an enterprise feature matrix. A rule expression tree module containing AND and OR nodes is constructed. The expression tree consists of sub-condition nodes based on TF-IDF weights and semantic similarity scores. The enterprise feature matrix is ​​input into the rule expression tree for sub-condition judgment, and candidate enterprises are output. The matching degree is calculated based on the sub-condition hit confidence. The expression tree is made differentiable and gated. Within a sliding window, a loss function is constructed based on the deviation between the number of candidate enterprises and the target size, as well as the missed detection feedback formed by supplementary entries. The keyword weights and threshold parameters are adaptively updated through gradient descent. The ranking module constructs an enterprise relationship graph based on candidate enterprises and calculates PageRank centrality. It then integrates centrality, matching degree, and time decay coefficient based on enterprise data update time to obtain enterprise weights and outputs the ranking. The output module employs a double-buffered update mechanism to perform periodic full updates and / or event-driven incremental updates in the computation buffer, and replaces the online buffer with an atomic switching method. It constructs a term co-occurrence network based on the business scope text field and the image text field, dynamically expands the keyword system through Louvain community discovery and pre-trained language model semantic drift detection, and updates the keyword set of the semantic feature space and rule expression tree accordingly. The output includes a chain of evidence containing hit sub-conditions, image region coordinates, and text fragment positions.

Images