A data intelligent archiving method based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's material intelligent archiving method, it is related to the technical field of intelligent management of archives information, including the following steps: S1, gather to be archived material and archive associated information, generate material original data set;S2, parse material content and standardize archiving feature, generate standardization material feature set;S3, input improved LayoutLMv3 model, introduce fingerprint decoupling constraint mechanism in three branch structure, generate material multidimensional portrait;S4, construct archive knowledge graph;S5, input CompGCN network inference, and generate candidate archive location set by sorting decoder;S6, generate dynamic archives directory tree and case file aggregation result, determine target archive location;S7, match archiving control information, generate credible archive package and output intelligent archiving result.The application uses improved LayoutLMv3 model and CompGCN network, realizes material accurate archiving.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for archival information, and in particular to an intelligent archiving method for documents based on big data analysis. Background Technology

[0002] With the increasing demand for electronic records management, digital government, and centralized archiving of enterprise documents, technologies for automatic identification, classification, archiving, and traceability of archiving basis for multi-source and multi-format documents have received widespread attention. Existing document archiving systems mainly rely on manual catalog selection, keyword matching rules, or fixed field mapping methods for document classification and catalog linking. However, in practical applications, the following problems are commonly encountered: The sources of materials to be archived are complex, including electronic documents, scanned copies, form attachments, and signed documents, as well as related information generated by different business systems. Existing methods often rely solely on document names, title keywords, or business numbers for judgment, making it difficult to accurately identify the subject matter, business item attribution, and source credibility of the materials. This leads to the easy mis-archiving of materials with the same name, similar materials, and cross-system materials. During the formation process, materials exist in multiple versions, such as edited drafts, approval drafts, circulation drafts, signed and completed drafts, and final drafts. Traditional archiving methods lack the ability to automatically identify version status, easily leading to the premature inclusion of incomplete materials in the official archives, affecting the integrity and standardization of the archives. Existing automatic archiving methods mostly use single classification models or fixed rules, which are difficult to express the multiple relationship constraints between materials, business items, archive categories, and case file catalogs. This results in the scattered storage of materials under the same business item, inaccurate case file aggregation, and a lack of interpretable basis for candidate archiving positions. In addition, the retention period, access permissions, and credible verification information after archiving are usually separated from the material itself, making it difficult to trace the basis for archiving, affecting the efficiency of subsequent auditing, review, and archive management.

[0003] Therefore, how to provide an intelligent data archiving method based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an intelligent document archiving method based on big data analysis. This invention employs an improved LayoutLMv3 model, an archiving knowledge graph, and a CompGCN network. It enhances the expression of key areas of the document through layout recoding guided by archiving element anchor points, and utilizes a fingerprint decoupling constraint mechanism to generate semantic fingerprints, business fingerprints, and evidence fingerprints of the document, enabling the identification of the subject, determination of business affiliation, and assessment of source credibility of the document to be archived. Furthermore, it distinguishes between process version status and archiving completion status through a version status identification task layer, and performs multi-relational archiving reasoning based on the archiving knowledge graph and CompGCN network to generate a set of candidate archiving locations and a target archiving location. Finally, it combines a dynamic archive directory tree, case file aggregation results, and a trusted archiving package to improve the traceability and credibility of the archiving results. This method boasts advantages such as high archiving accuracy, strong version identification capability, good case file aggregation effect, and applicability to intelligent management of large volumes of documents.

[0005] According to an embodiment of the present invention, a method for intelligent data archiving based on big data analysis includes the following steps: S1. Collect the data to be archived and its archiving association information, and generate the original dataset of the data; S2. Perform data content parsing and archiving feature standardization on the original data dataset to generate a standardized data feature set; S3. Input the standardized data feature set into the improved LayoutLMv3 model. The improved LayoutLMv3 model includes a text embedding module, a layout embedding module, a multimodal Transformer encoder, a three-branch structure, and a version status recognition task layer. Introduce a fingerprint decoupling constraint mechanism in the three-branch structure to generate a multi-dimensional profile of the data. S4. Generate data nodes, business item nodes, and archive category nodes based on the multi-dimensional profile of the data, and establish business affiliation relationships and archiving mapping relationships to construct an archive knowledge graph; S5. Input the multi-dimensional profile of the data and the archived knowledge graph into the CompGCN network to perform multi-relation archived reasoning, generate a multi-relation archived reasoning representation, and generate a set of candidate archived locations through the candidate archived location sorting decoder. S6. Generate a dynamic archive directory tree and case file aggregation results based on the candidate archive location set and data association strength, and determine the target archive location; S7. Match archiving control information according to the target archiving location, generate a trusted archiving package, and output the intelligent archiving result.

[0006] Optionally, S1 specifically includes: Obtain the documents to be archived, and simultaneously obtain the archiving association information corresponding to the documents to be archived; The archiving association information is determined based on the data source and business matters of the data to be archived; The data to be archived is collected and verified to determine whether the data to be archived has missing content, missing archiving association information, or abnormal data format. The archive association information is bound to the archived data that has passed the data collection and verification, thus forming a correspondence between the archived data and the archive association information; According to the correspondence, the data to be archived and the archive association information are archived, collected, and numbered. The data to be archived and the archive association information that have been archived and numbered are then stored together to generate the original dataset.

[0007] Optionally, S2 specifically includes: Read the data to be archived, archive association information, and archive collection number from the original dataset. The data to be archived is parsed to obtain the data content parsing result, which includes the data text content and the data layout information; Based on the archive association information, archive feature fields are extracted from the data content parsing results. The archive feature fields include data subject field, business item field, formation stage field, and archive category field. The archive feature fields are subjected to archive feature standardization processing, which includes unifying field names and standardizing field values; The archive feature fields that have undergone standardized archiving feature processing, the data content parsing results, and the archive collection number are associated to generate a standardized data feature set.

[0008] Optionally, S3 specifically includes: Input the standardized data feature set into the improved LayoutLMv3 model; The archive feature fields, data content parsing results and archive collection number are read from the standardized data feature set, and the data text content and data format information are extracted from the data content parsing results. Archive element fields are generated based on the archive feature fields. The text embedding module performs text segmentation, field marking, and ordinal encoding based on the data text content to generate data text embedding features; The layout embedding module divides the page area, marks the area type, and encodes the area position based on the document layout information, and generates document layout embedding features. Determine the archived element anchor points based on the archived element fields, and convert the document layout area corresponding to the archived element anchor points into an enhanced archived layout representation. Input the data text embedding features, data layout embedding features, and archive layout enhancement representation into the multimodal Transformer encoder to generate a joint data representation; The data joint representation is input into a three-branch structure, and a fingerprint decoupling constraint mechanism is introduced into the three-branch structure to perform branch encoding on the data joint representation, generating data semantic fingerprint, data business fingerprint and data evidence fingerprint; Input the joint representation of data, the enhanced representation of archived format, and the fingerprint of data evidence into the version status recognition task layer to generate version status recognition results; By associating the semantic fingerprint, business fingerprint, evidence fingerprint, version status identification result, and archive collection number of the data, a multi-dimensional profile of the data is formed.

[0009] Optionally, the fingerprint decoupling constraint mechanism is specifically as follows: The data joint representation is divided into semantic coding branch, business coding branch and evidence coding branch; In the semantic coding branch, the title coding vector, subject word coding vector, core paragraph coding vector, and archive category coding vector in the joint representation of the data are aggregated to generate the data subject coding vector, and the data semantic fingerprint is generated based on the data subject coding vector. In the business coding branch, the business item number coding vector, project number coding vector, contract number coding vector, approval number coding vector and process node coding vector in the data joint representation are aggregated to generate the business item coding vector, and the data business fingerprint is generated based on the business item coding vector. In the evidence coding branch, the data source coding vector, upload subject coding vector, signature status coding vector, document verification result coding vector and formation stage coding vector in the data joint representation are aggregated to generate the evidence source coding vector, and the data evidence fingerprint is generated based on the evidence source coding vector. Calculate the fingerprint similarity among semantic fingerprints, business fingerprints, and evidentiary fingerprints of data; Two coding branches with fingerprint similarity higher than a preset similarity threshold are identified as branches to be decoupled, and repeated coding vectors are extracted from the branches to be decoupled. Based on the contribution ratio of the repeated coding vector in the branch to be decoupled, the repeated coding vector is weighted down to generate the decoupled coding vector; Write the decoupled encoding vector into the corresponding branch to be decoupled, and retain the encoding vectors that are not marked as branches to be decoupled, to generate the data semantic fingerprint, data business fingerprint and data evidence fingerprint after decoupling.

[0010] Optionally, S4 specifically includes: Read the multi-dimensional profile of the data, which includes the semantic fingerprint of the data, the business fingerprint of the data, the evidence fingerprint of the data, and the version status identification result; Data nodes are generated based on the data semantic fingerprint, and the data semantic fingerprint is written into the content attributes of the data nodes; Based on the data service fingerprint, a business item node is generated, and a business attribution relationship is established between the data node and the business item node; Generate archive category nodes according to preset archive classification rules, and establish an archiving mapping relationship between the business item nodes and the archive category nodes; Evidence attributes of data nodes are generated based on the data evidence fingerprint, and version attributes of data nodes are generated based on the version status identification results; The data nodes, business item nodes, archive category nodes, business affiliation relationships, and archiving mapping relationships are represented in a graph structure to construct an archiving knowledge graph.

[0011] Optionally, S5 specifically includes: Read the multi-dimensional profile of the data and the archived knowledge graph, and map the data semantic fingerprint, data business fingerprint, data evidence fingerprint and version status identification result in the multi-dimensional profile of the data to the data node in the archived knowledge graph to generate an initial representation of the data node; A node set is generated based on the data nodes, business item nodes, and archive category nodes in the archive knowledge graph, and a relationship set is generated based on the business affiliation relationship and the archive mapping relationship; The node set, relation set, and initial representation of data nodes are input into the CompGCN network, and the CompGCN network performs multi-relation archiving reasoning on the archived knowledge graph. In the multi-relation archiving reasoning process, the CompGCN network combines the initial representation of the data node with the relation representation corresponding to the relation set to obtain the updated representation of the data node under the constraints of business item attribution and archive category mapping. Generate a multi-relation archive reasoning representation based on the node update representation; The multi-relation archiving reasoning representation is input into the candidate archiving position sorting decoder, which calculates the sorting score of the data to be archived corresponding to different archiving position items and generates a candidate archiving position set based on the sorting score. The candidate archive location set includes candidate archive location, candidate archive category, candidate directory level, and archive confidence level.

[0012] Optionally, the step of calculating the ranking score of the document to be archived corresponding to different archive location items by the candidate archive location ranking decoder specifically involves: Archive location items are generated based on the candidate archive categories and the candidate directory levels, and archive location representations are generated based on the archive category nodes corresponding to the archive location items in the archive knowledge graph. The multi-relation archiving reasoning representation and the archiving location representation are input into the candidate archiving location sorting decoder, and the candidate archiving location sorting decoder calculates the business affiliation matching score and archiving rule adaptation score between the data to be archived and the archiving location item. The business affiliation matching score is determined based on the data business fingerprint, the business item node, and the business affiliation relationship. The archiving rule adaptation score is determined based on the data semantic fingerprint, the data evidence fingerprint, the version status identification result, and the archiving mapping relationship; The ranking score is obtained by fusing the business affiliation matching score and the archiving rule adaptation score. The different archiving locations are sorted in descending order based on the ranking score. The archiving locations that meet the preset filtering conditions are selected as candidate archiving locations. An archiving confidence score is generated based on the ranking score. The candidate archiving locations, archiving confidence scores, and the candidate archive categories and candidate directory levels corresponding to the candidate archiving locations are combined to generate a set of candidate archiving locations.

[0013] Optionally, S6 specifically includes: Read the candidate archiving location set, and extract the candidate archiving location, candidate archive category, candidate directory level, and archiving confidence level corresponding to the data to be archived from the candidate archiving location set; Calculate the data association strength between the data to be archived based on the multi-dimensional profile of the data and the archived knowledge graph; Based on the strength of the data association, the data to be archived that meet the preset association conditions are divided into the same case file data set, and the case file aggregation result is generated based on the case file data set; A dynamic archive directory tree is generated based on the candidate archive locations and the case file aggregation results; The candidate archive locations are checked for directory consistency based on the dynamic archive directory tree, and the target archive location is determined from the candidate archive locations based on the archive confidence score. The target archiving location is associated with the case file aggregation result to form the archiving location result of the data to be archived.

[0014] Optionally, S7 specifically includes: Read the target archive location, the dynamic archive directory tree, and the case file aggregation result; The archive category and directory level corresponding to the data to be archived are determined based on the target archiving location, and archiving control rules are matched based on the archive category and directory level. Archive control information is generated according to the archive control rules, and the archive control information includes storage control method and permission control method; The archiving criteria are generated based on the target archiving location, the dynamic archive directory tree, and the case file aggregation results. Generate a trusted verification identifier based on the data to be archived, and bind the trusted verification identifier with the archiving basis and archiving control information to generate a trusted archive package; The intelligent archiving result is output based on the trusted archive package. The intelligent archiving result includes the archive directory result and the archive description result.

[0015] The beneficial effects of this invention are: This invention addresses existing problems in document archiving, such as complex data sources, inconsistent formats, mixed version statuses, difficulty in determining business affiliation, and difficulty in tracing archiving basis, by improving the LayoutLMv3 model, integrating the archived knowledge graph, and the CompGCN network. It employs data content parsing and archived feature standardization to convert documents from different sources and formats into a unified, standardized set of data features. In the data understanding stage, it introduces archived area recoding guided by archived element anchors, focusing the model on key archiving areas such as titles, business item numbers, signatures, and dates. Furthermore, it generates semantic fingerprints and business fingerprints of the data through a fingerprint decoupling constraint mechanism. Fingerprints and document evidence fingerprints enable the differentiation and expression of document themes, business matters attribution, and source credibility. Addressing the issue of mis-archiving caused by mixing process drafts, approval drafts, signed and finalized drafts, this invention sets up a version status identification task layer to determine whether documents meet the formal archiving conditions. In the archiving reasoning stage, an archiving knowledge graph is constructed, and a CompGCN network is used to perform joint reasoning on multiple relationships between documents, business matters, and archive categories, generating a set of candidate archiving locations. Finally, combining a dynamic archive directory tree, case file aggregation results, and a trusted archiving package, the target archiving location, archiving control information, and archiving basis are output, achieving accurate document archiving, automatic case file aggregation, and traceable archiving results. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a data intelligent archiving method based on big data analysis proposed in this invention; Figure 2 This is a structural diagram of the improved LayoutLMv3 model proposed in this invention; Figure 3 A graph showing the CompGCN archive reasoning and candidate archive position ranking. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A data intelligent archiving method based on big data analysis includes the following steps: S1. Collect the data to be archived and its archiving association information, and generate the original dataset of the data; S2. Perform data content parsing and archiving feature standardization on the original data dataset to generate a standardized data feature set; S3. Input the standardized data feature set into the improved LayoutLMv3 model. The improved LayoutLMv3 model includes a text embedding module, a layout embedding module, a multimodal Transformer encoder, a three-branch structure, and a version status recognition task layer. Introduce a fingerprint decoupling constraint mechanism in the three-branch structure to generate a multi-dimensional profile of the data. S4. Generate data nodes, business item nodes, and archive category nodes based on the multi-dimensional profile of the data, and establish business affiliation relationships and archiving mapping relationships to construct an archive knowledge graph; S5. Input the multi-dimensional profile of the data and the archived knowledge graph into the CompGCN network to perform multi-relation archived reasoning, generate a multi-relation archived reasoning representation, and generate a set of candidate archived locations through the candidate archived location sorting decoder. S6. Generate a dynamic archive directory tree and case file aggregation results based on the candidate archive location set and data association strength, and determine the target archive location; S7. Match archiving control information according to the target archiving location, generate a trusted archiving package, and output the intelligent archiving result.

[0019] In this embodiment, S1 specifically refers to: Retrieve the documents to be archived, and simultaneously retrieve the archiving association information corresponding to the documents to be archived; Determine the archiving-related information based on the source of the documents to be archived and the business matters involved; The data to be archived is collected and verified to determine whether there are missing data content, missing archiving information, or abnormal data format. Bind archiving association information to the data to be archived that has passed the data collection and verification, and form a correspondence between the data to be archived and the archiving association information; According to the corresponding relationship, the archived data and archived related information are collected and numbered, and the archived data and archived related information that have been collected and numbered are stored together to generate the original data dataset; In the specific implementation process, the materials to be archived can be obtained through the document management system interface, business system interface or batch import method. A single collection task is executed in units of one archiving batch, and each archiving batch contains 50 to 5000 documents to be archived. Archived information is extracted from the data source field and the business item field. The data source field includes at least the source system identifier and the uploader identifier, and the business item field includes at least the business item number and the business item name.

[0020] Data collection and verification are performed based on completeness, relevance, and format compliance. Specifically, missing data content is considered anomaly if the file size is less than 10KB, the number of parseable characters in the main text is less than 20, or the number of scanned pages is 0. Missing archiving association information is considered anomaly if the business item number or source system identifier is empty. Abnormal data format is considered anomaly if the file extension does not match the file header information, the file cannot be opened, or the file verification fails. For data that passes the data collection and verification, an archiving collection number is generated according to the rule "source system identifier + business item number + collection date + four-digit serial number". Then, using the archiving collection number as the primary key, the storage path, data source, business item, collection time, and verification result of the data to be archived are associated and stored to generate the original data dataset.

[0021] In this embodiment, S2 specifically refers to: Read the data to be archived, archive association information, and archive collection number from the original dataset; The archived materials are analyzed to obtain the content analysis results, which include the text content and format information of the materials. Based on the archive association information, archive feature fields are extracted from the data content parsing results. Archive feature fields include data subject field, business item field, formation stage field, and archive category field. The archival feature fields are standardized, which includes unifying field names and standardizing field values. The archive feature fields that have completed the standardization of archive features, the data content parsing results, and the archive collection number are associated to generate a standardized data feature set; In the specific implementation process, when reading the original dataset, the archive collection number is used as an index to call up the archived documents and archived related information one by one; for electronic text documents, the main text, title and table text are directly extracted; for scanned documents, OCR recognition is used to obtain the document text content, and the layout information of the document is obtained by detecting the page area, paragraph position, table position and signature position. The data content analysis results consist of the data text content and the data layout information. The data text content is saved in UTF-8 encoding, and the data layout information is recorded in a structured manner according to page number, area coordinates, area type and area order. The area coordinates adopt normalized coordinates with the upper left corner of the page as the origin, and the coordinate values ​​range from 0 to 1. The archive feature fields are extracted jointly based on the archive association information and the data content parsing results. When the business item number exists in the archive association information, the business item number is directly written into the archive feature field. When the business item number is missing, the contract number, project number, or approval number is matched from the data text content, and the number with a matching confidence level of not less than 0.85 is used as the business item field. The data subject is extracted based on the title text and the first 300 characters of the body text. The formation stage is determined based on the signature position, approval opinion area, and date area in the data layout information. The standardization of archival features is performed using a field mapping table and a value specification table. The field mapping table unifies synonymous fields such as "project number" and "item number" into business item numbers, and the value specification table unifies dates into YYYY-MM-DD format and maps archival categories to preset archive category codes. Finally, the archival feature fields, data content parsing results, and archival collection numbers are written into the standardized data feature set according to a one-to-one correspondence.

[0022] In this embodiment, S3 specifically refers to: Input the standardized data feature set into the improved LayoutLMv3 model; The archive feature fields, data content parsing results and archive collection number are read from the standardized data feature set, and the data text content and data format information are extracted from the data content parsing results. Archive element fields are generated based on the archive feature fields. The text embedding module performs text segmentation, field tagging, and ordinal encoding based on the data text content to generate data text embedding features; The layout embedding module divides the page area, marks the area type, and encodes the area location based on the document layout information, and generates document layout embedding features. Determine the archived element anchor points based on the archived element fields, and convert the document layout area corresponding to the archived element anchor points into an enhanced archived layout representation. Input the data text embedding features, data layout embedding features, and archive layout enhancement representation into the multimodal Transformer encoder to generate a joint data representation; The data joint representation is input into a three-branch structure, and a fingerprint decoupling constraint mechanism is introduced into the three-branch structure to perform branch encoding on the data joint representation, generating data semantic fingerprint, data business fingerprint and data evidence fingerprint; Input the joint representation of data, the enhanced representation of archived format, and the fingerprint of data evidence into the version status recognition task layer to generate version status recognition results; By associating the semantic fingerprint, business fingerprint, evidence fingerprint, version status identification result, and archive collection number of the data, a multi-dimensional profile of the data is formed. In the specific implementation process, the improved LayoutLMv3 model inherits the document multimodal joint coding framework of the traditional LayoutLMv3 model in its overall structure, including text embedding, layout embedding, two-dimensional positional coding, and Transformer coding process. However, in the multimodal coding process, it introduces archived area recoding guided by archived element anchor points to generate archived layout enhanced representation and constructs a three-branch structure to extract document semantic fingerprints, document business fingerprints, and document evidence fingerprints respectively. At the same time, a fingerprint decoupling constraint mechanism is introduced into the three-branch structure. Through fingerprint similarity constraints and weighting of repeated coding vectors, the redundant coupling between document semantic fingerprints, document business fingerprints, and document evidence fingerprints is reduced, so that document theme, business item attribution, and evidence source form mutually distinguishable fingerprint expressions. Combined with the version status recognition task layer, version status recognition results are generated, thereby forming a multi-dimensional profile of the document. First, the archived feature fields, data content parsing results, and archived collection numbers are read from the standardized data feature set. The data text content is then segmented according to paragraphs and layout areas. The maximum length of each text segment is set to 512 characters. Data exceeding 512 characters is segmented using a sliding window with a window overlap rate of 20%. The page area coordinates in the document layout information are uniformly converted into two-dimensional position coordinates, with the coordinate range normalized to 0 to 1000, and the multimodal input data of the document is constructed according to the correspondence of "text fragment, area coordinates, area type, and archive collection number". Then, the archived element anchor points are determined based on the archived feature fields. The archived element anchor points include the area where the business item number is located, the area where the document title is located, the area where the signature status is located, and the area where the date field is located. For fields that cannot be directly located, the anchor point score is determined by combining keyword matching confidence, field completeness, layout salience, and source credibility. Among these factors, keyword matching confidence accounts for 35%, field completeness accounts for 25%, layout salience accounts for 20%, and source credibility accounts for 20%. When the overall anchor point score is not lower than 0.75, the area where the corresponding field is located is determined as the archived element anchor point.

[0023] Then, the archived format information is recoded in the archived area with the archived element anchor point as the center. The area weight of the anchor point area is set to 1.0, the area weight of the two adjacent areas on the same page or the area with a normalized coordinate distance of less than 0.15 is set to 0.65, the weight of the ordinary area is set to 0.25, and the area weight is added to the two-dimensional position embedding to form the archived format enhanced representation. Subsequently, the data text embedding features, data layout embedding features, and archive layout enhancement representation are input into the improved LayoutLMv3 model. The number of Transformer encoding layers in the model is set to 12, the hidden layer dimension is set to 768, and the number of multi-head attention heads is set to 12. The output is the joint representation of the data. Unlike the traditional LayoutLMv3 model, which relies solely on text, images, and layout positions for general document understanding, the improved LayoutLMv3 model prioritizes key archiving areas such as business item numbers, titles, signatures, and dates by recoding the archiving area, thus avoiding interference from ordinary text areas in archiving judgment. After the data is combined into a three-branch structure, the semantic fingerprint output branch generates a data semantic fingerprint based on the data topic-related coding vector, the business fingerprint output branch generates a data business fingerprint based on the business item-related coding vector, and the evidence fingerprint output branch generates a data evidence fingerprint based on the source, signature, and formation stage-related coding vector. All three fingerprint dimensions are set to 256. The fingerprint decoupling constraint mechanism controls the redundant representation between semantic fingerprints, business fingerprints, and evidence fingerprints through fingerprint similarity. When the fingerprint similarity between any two fingerprints is higher than 0.75, the corresponding two coding branches are marked as branches to be decoupled, and the repeated coding vectors in the branches to be decoupled are weighted to generate decoupled coding vectors. Finally, the joint representation of data, the enhanced representation of archived format, and the fingerprint of data evidence are input into the version status identification task layer. The version status identification task layer generates version status identification results based on the content change characteristics, the distribution characteristics of archived elements, and the credibility characteristics of the source. The classification confidence threshold is set to 0.80. Data to be archived that is lower than 0.80 is marked as requiring manual review. Compared to the unimproved LayoutLMv3 model, this invention can not only identify the content of the data, but also distinguish the business ownership, evidence source and version status of the data in the archiving process, providing a more stable structured input for subsequent archiving knowledge graph construction and candidate archiving location reasoning.

[0024] In this embodiment, the fingerprint decoupling constraint mechanism is specifically as follows: The data joint representation is divided into semantic coding branch, business coding branch and evidence coding branch; In the semantic coding branch, the title coding vector, subject word coding vector, core paragraph coding vector, and archive category coding vector in the joint representation of the data are aggregated to generate the data subject coding vector, and the data semantic fingerprint is generated based on the data subject coding vector. In the business coding branch, the business item number coding vector, project number coding vector, contract number coding vector, approval number coding vector and process node coding vector in the data joint representation are aggregated to generate the business item coding vector, and the data business fingerprint is generated based on the business item coding vector. In the evidence coding branch, the data source coding vector, upload subject coding vector, signature status coding vector, document verification result coding vector and formation stage coding vector in the data joint representation are aggregated to generate the evidence source coding vector, and the data evidence fingerprint is generated based on the evidence source coding vector. Calculate the fingerprint similarity among semantic fingerprints, business fingerprints, and evidentiary fingerprints of data; Two coding branches with fingerprint similarity higher than a preset similarity threshold are identified as branches to be decoupled, and repeated coding vectors are extracted from the branches to be decoupled. Based on the contribution ratio of the repeated coding vector in the branch to be decoupled, the repeated coding vector is weighted down to generate the decoupled coding vector; Write the decoupled encoding vector into the corresponding branch to be decoupled, and retain the encoding vectors that are not marked as branches to be decoupled, to generate the data semantic fingerprint, data business fingerprint and data evidence fingerprint after decoupling.

[0025] In practical implementation, the fingerprint decoupling constraint mechanism is set within the three-branch structure. The semantic coding branch, business coding branch, and evidence coding branch all receive the same joint representation of the data. The semantic coding branch prioritizes selecting the coding vectors corresponding to the title area, keyword area, core paragraph area of ​​the main text, and archive category field, and generates the data topic coding vector using an attention-weighted aggregation method. The business coding branch prioritizes selecting the coding vectors corresponding to the business item number, project number, contract number, approval number, and process node field, and generates the business item coding vector according to the field completeness and field matching confidence. The evidence coding branch prioritizes selecting the coding vectors corresponding to the data source, uploading entity, signature status, document verification result, and formation stage, and generates the evidence source coding vector according to the source credibility and signature completeness.

[0026] After the three types of encoded vectors are generated, they are normalized into 256-dimensional semantic fingerprints, 256-dimensional business fingerprints, and 256-dimensional evidence fingerprints, respectively. Fingerprint similarity is calculated using cosine similarity, with a similarity threshold set to 0.75. When the similarity between any two fingerprints is higher than 0.75, the corresponding two encoded branches are designated as branches to be decoupled. Encoded vectors with higher attention weights from these branches are selected as repeated encoded vectors. The contribution ratio of repeated encoded vectors is determined based on their proportion of attention weights in the two branches to be decoupled, and they are weighted according to this ratio, with a weighting factor set between 0.3 and 0.6. After weighting, the decoupled encoded vectors are written into the corresponding branches to be decoupled. Encoded branches not designated as branches to be decoupled retain their original encoded vectors, ultimately generating the decoupled semantic fingerprints, business fingerprints, and evidence fingerprints.

[0027] In this embodiment, S4 specifically refers to: Read the multi-dimensional profile of the data, which includes the data semantic fingerprint, data business fingerprint, data evidence fingerprint, and version status recognition results; Data nodes are generated based on data semantic fingerprints, and the data semantic fingerprints are written into the content attributes of the data nodes; Business item nodes are generated based on data business fingerprints, and the business attribution relationship between data nodes and business item nodes is established. Generate archive category nodes according to preset archive classification rules, and establish an archiving mapping relationship between business item nodes and archive category nodes; Evidence attributes of data nodes are generated based on data evidence fingerprints, and version attributes of data nodes are generated based on version status identification results; The document nodes, business item nodes, archive category nodes, business affiliation relationships, and archiving mapping relationships are expressed in a graph structure to construct an archiving knowledge graph; In the specific implementation process, the archive knowledge graph is constructed with data nodes, business item nodes and archive category nodes as core node types, and with business affiliation relationship and archive mapping relationship as core edge types. Data nodes are uniquely generated based on the archive collection number. The node identifier uses a combination of "data node + archive collection number" coding. The data semantic fingerprint is written as the content attribute of the data node. The data theme similarity can be calculated by the cosine similarity between the data semantic fingerprints. When the similarity is not less than 0.80, the corresponding data is marked as data with similar themes. Business item nodes are generated based on data business fingerprints. When the cosine similarity between the data business fingerprint and the existing business item node is not less than 0.85, the data to be archived is associated with the existing business item node; when the similarity is less than 0.85, a new business item node is created based on the business item number in the archive feature field. The archive category node is generated based on the preset archive classification rules, which include the archive category code, category name, applicable business scope, and retention period rules. For example, contract documents are mapped to the contract archive category, and approval documents are mapped to the approval archive category. The business attribution relationship is established based on the matching results between data nodes and business item nodes. The relationship weight is calculated jointly by the data business fingerprint similarity and the business item number matching results, where the data business fingerprint similarity weight is set to 0.70 and the business item number matching weight is set to 0.30. The archiving mapping relationship is established based on the archiving classification rules between business item nodes and archiving category nodes. When the business item type matches the preset archiving classification rules and the source credibility of the data evidence fingerprint is not less than 0.75, the archiving mapping relationship is established. Finally, the data nodes, business item nodes, archive category nodes, and their corresponding relationships are converted into triples for storage. The triple format is "head node - relation type - tail node". The content attributes, evidence attributes, and version attributes of the data nodes are written as additional attributes of the nodes to form an archived knowledge graph.

[0028] In this embodiment, S5 specifically refers to: Read the multi-dimensional profile of the data and the archived knowledge graph, and map the data semantic fingerprint, data business fingerprint, data evidence fingerprint and version status recognition results in the multi-dimensional profile of the data to the data nodes in the archived knowledge graph to generate the initial representation of the data nodes; A set of nodes is generated based on the data nodes, business item nodes, and archive category nodes in the archive knowledge graph, and a set of relationships is generated based on the business affiliation relationship and the archive mapping relationship; The node set, relation set, and initial representation of data nodes are input into the CompGCN network, which then performs multi-relation archiving reasoning on the archived knowledge graph. In the multi-relation archiving reasoning process, the CompGCN network combines the initial representation of the data node with the relation representation corresponding to the relation set to obtain the updated representation of the data node under the constraints of business item attribution and archive category mapping. Generate a multi-relation archive reasoning representation based on the node update representation; The multi-relation archiving reasoning representation is input into the candidate archiving position sorting decoder. The candidate archiving position sorting decoder calculates the sorting score of the data to be archived for different archiving position items, and generates a candidate archiving position set based on the sorting score. The candidate archive location set includes candidate archive location, candidate archive category, candidate directory level, and archive confidence level; In practice, the CompGCN network is used to perform joint reasoning on multiple types of nodes and multiple types of relationships in the archived knowledge graph. Compared with the ordinary GCN, which only aggregates the features of adjacent nodes, the CompGCN network simultaneously updates the node representation and the relationship representation, which can more accurately express the archiving constraint relationships between the data to be archived, business matters and archive categories. First, the multi-dimensional profile of the data and the archived knowledge graph are read, and the semantic fingerprint, business fingerprint, evidence fingerprint and version status recognition results of the data are mapped to the data nodes in the archived knowledge graph. Among them, the dimensions of the semantic fingerprint, business fingerprint and evidence fingerprint are uniformly set to 256 dimensions, and the version status recognition results are converted into 32-dimensional state embedding vectors. Then, the four types of vectors are concatenated and input into the linear mapping layer to map to the initial representation of the 256-dimensional data node. Then, data nodes, business item nodes, and archive category nodes are extracted from the archive knowledge graph to form a node set; a relationship set is formed based on business affiliation and archive mapping relationships, and a relationship representation is configured for each relationship, with the relationship representation dimension set to 256 dimensions; the initial weight of the business affiliation relationship is determined based on the similarity of the data business fingerprint, and the initial weight of the archive mapping relationship is determined based on the hit rate of the preset archive classification rules and the source credibility of the data evidence fingerprint; In the multi-relation archiving reasoning process, the CompGCN network combines the initial representation of data nodes with the relation representation, using vector multiplication or cyclic correlation operations. When a data node is connected to a business item node through a business affiliation relation, the network updates the affiliation strength of the data node to the business item based on the business affiliation relation representation. When a business item node is connected to an archive category node through an archiving mapping relation, the network updates the matching strength of the data node to the archive category based on the archiving mapping relation representation. The CompGCN network has 2 to 3 layers, with a hidden dimension of 256 per layer and Dropout set to 0.2. The node update representation is calculated by combining the neighbor node representation, the relation representation, and the relation weight. Edges with a relation weight lower than 0.50 do not participate in this round of aggregation to avoid low-confidence relations interfering with archiving reasoning. After completing multi-level aggregation, the final node update representation corresponding to the data node is extracted and used as the multi-relation archiving reasoning representation to characterize the archiving matching relationship between the data to be archived and different archive categories. Finally, the multi-relation archiving reasoning representation is input into the candidate archiving location sorting decoder. The candidate archiving location sorting decoder calculates the matching score between the data to be archived and each archiving location item. The sorting score is generated by weighting the node matching score, relation path score, and evidence credibility score, with the weights of the three set to 0.50, 0.30, and 0.20, respectively. Archiving location items with a sorting score of not less than 0.70 are selected as candidate archiving locations, and archiving confidence is generated based on the sorting score, thus forming a set of candidate archiving locations that includes candidate archiving locations, candidate file categories, candidate directory levels, and archiving confidence.

[0029] In this embodiment, the candidate archive position sorting decoder calculates the sorting score for different archive position items corresponding to the data to be archived, specifically as follows: Archive location items are generated based on candidate archive categories and candidate directory levels, and archive location representations are generated based on the archive category nodes corresponding to the archive location items in the archive knowledge graph. The multi-relation archiving inference representation and archiving location representation are input into the candidate archiving location sorting decoder, which calculates the business affiliation matching score and archiving rule adaptation score between the data to be archived and the archiving location item. Among them, the business attribution matching score is determined based on the data business fingerprint, business item node, and business attribution relationship; The archiving rule adaptation score is determined based on the semantic fingerprint of the data, the evidence fingerprint of the data, the version status identification result, and the archiving mapping relationship. The business affiliation matching score and the archiving rule adaptation score are combined to obtain the ranking score; The different archiving location items are sorted in descending order according to the ranking score. The archiving location items that meet the preset filtering conditions are taken as candidate archiving locations. The archiving confidence score is generated according to the ranking score. The candidate archiving locations, the archiving confidence scores, and the candidate archive categories and candidate directory levels corresponding to the candidate archiving locations are combined to generate a set of candidate archiving locations. In the specific implementation process, the candidate archiving location sorting decoder is used to score the archiving adaptation of different archiving location items. The archiving location item is generated by combining the candidate archive category and the candidate directory level. The encoding format can be set to "archiving category code + directory level code + year identifier". For example, HT-02-2026 represents the 2026 archiving location of the second-level directory under the contract archive category. First, an archive location representation is generated based on the archive category node corresponding to the archive location item in the archive knowledge graph. The archive location representation is obtained by linear mapping after concatenation of archive category node embedding, directory hierarchy embedding and archive rule embedding, with the dimension set to 256. Then, the multi-relation archive reasoning representation and archive location representation are input into the candidate archive location sorting decoder, which calculates the business affiliation matching score and the archive rule adaptation score, respectively. The business attribution matching score is determined based on the data business fingerprint, business item node, and business attribution relationship. Specifically, the similarity between the data business fingerprint and the business item node representation is used as the primary judgment factor, and the weight of the business attribution relationship is used as a secondary judgment factor. Among them, the similarity between the data business fingerprint and the business item node representation accounts for 70%, and the weight of the business attribution relationship accounts for 30%. When the business item number matches completely, the weight of the business attribution relationship is 1.0; when only the business type matches, the weight of the business attribution relationship is 0.6; and when neither the business item number nor the business type matches, the weight of the business attribution relationship is 0.

[0030] The archiving rule adaptation score is determined based on the semantic fingerprint of the document, the evidence fingerprint of the document, the version status identification result, and the archiving mapping relationship. Specifically, it can be calculated by weighting the document topic matching degree, source credibility, version status validity, and archiving mapping relationship, with weights set to 0.35, 0.25, 0.20, and 0.20, respectively. Among them, when the version status is signed and stamped, finalized, or ready for archiving, the version status validity is 1.0; when the version status is approved or circulating, it is 0.5; and when the version status is edited, it is 0.2. Finally, when integrating the business affiliation matching score and the archiving rule adaptation score, the candidate archiving position sorting decoder uses the business affiliation matching score as the main criterion and the archiving rule adaptation score as the auxiliary criterion. The business affiliation matching score accounts for a higher proportion of the sorting results than the archiving rule adaptation score. In the specific integration process, the business affiliation matching score accounts for 55%, and the archiving rule adaptation score accounts for 45%. The two scores are weighted together to form the ranking score. When the ranking score is not lower than 0.70, the corresponding archive position item is taken as a candidate archive position, and they are arranged in descending order of ranking score. The top 3 archive position items are selected to enter the candidate archive position set. Archive confidence is generated based on the normalized ranking score; the higher the ranking score, the higher the archive confidence. When the difference between the highest ranking score and the second highest ranking score is not less than 0.15, it indicates that the candidate archive position ranked first has a significant advantage over other candidate positions. The candidate archive position corresponding to the highest ranking score is then given priority recommendation, thereby improving the accuracy and interpretability of candidate archive position selection.

[0031] In this embodiment, S6 specifically refers to: Read the candidate archive location set, and extract the candidate archive location, candidate archive category, candidate directory level, and archive confidence level of the data to be archived from the candidate archive location set; Calculate the strength of data associations between the data to be archived based on the multidimensional profiles of the data and the archived knowledge graph; Based on the strength of the data association, the data to be archived that meet the preset association conditions are divided into the same case file data set, and the case file aggregation result is generated based on the case file data set; A dynamic archive directory tree is generated based on the candidate archive locations and case file aggregation results; Based on the dynamic archive directory tree, the candidate archive locations are checked for directory consistency, and the target archive location is determined from the candidate archive locations based on the archive confidence score. The target archiving location is associated with the case file aggregation results to form the archiving location result of the data to be archived; In the specific implementation process, when reading the candidate archive location set, the candidate archive location, candidate archive category, candidate directory level and archive confidence are extracted using the archive collection number as the index. The data association strength is calculated based on the data semantic fingerprint, data business fingerprint, and version status identification results. The data semantic fingerprint is used to compare the similarity of data topics, the data business fingerprint is used to compare the consistency of business matters, and the version status identification results are used to determine whether the data are in the same archiving stage. The similarity of data topics accounts for 45%, the consistency of business matters accounts for 40%, and the consistency of version status accounts for 15%, and the data association strength is obtained by combining the results. When the correlation strength of the data is not less than 0.75 and the candidate archives are of the same category, the corresponding data to be archived will be assigned to the same case file data set; when the correlation strength of the data is between 0.60 and 0.75, the corresponding data to be archived will be marked as data to be reviewed and will not be directly included in the case file data set. The case file aggregation results are generated based on the case file data set. Specifically, they are formed by statistically analyzing the number of data in the case file data set, the common business item number, the candidate file categories, and the 3 to 5 most frequently occurring thematic keywords. The case file number is generated according to the rule of "file category code + year + three-digit serial number". The dynamic archive catalog tree is generated based on the candidate filing location and the case file aggregation results. Specifically, the catalog nodes are established according to the hierarchy of "archive category - year - case file - data item", and the case file number, data item and candidate catalog level are attached to the corresponding catalog node. The target archiving location is determined based on the catalog consistency check results and archiving confidence. When a candidate archiving location is consistent with the archive category, year information and case subject in the dynamic archive catalog tree, and the archiving confidence is not lower than 0.80, the corresponding candidate archiving location is determined as the target archiving location. The archiving location result is formed by linking the target archiving location and the case file aggregation result, and is used to record the archive category, directory level, case file number and data entry location of the data to be archived.

[0032] In this embodiment, S7 specifically refers to: Read the target archive location, dynamic archive directory tree, and case file aggregation results; Determine the archive category and catalog level corresponding to the materials to be archived based on the target archiving location, and match the archiving control rules according to the archive category and catalog level; Archive control information is generated based on the archive control rules. The archive control information includes the storage control method and the access control method. Generate archiving criteria based on the target archiving location, dynamic archive directory tree, and case file aggregation results; Generate a trusted verification identifier based on the documents to be archived, and bind the trusted verification identifier with the archiving basis and archiving control information to generate a trusted archive package; Based on the trusted archive package, intelligent archiving results are output, including archive directory results and archive description results. In the specific implementation process, after reading the target archiving location, dynamic archive directory tree and case file aggregation results, the archiving control rules are matched according to the archive category code and directory level number in the target archiving location. The archiving control rules are called from the preset archive management rule table. The rule table includes at least the archive category code, retention period, access role, download permission and review requirements; for example, the retention period for ordinary business data is set to 10 years, the retention period for important contract data is set to 30 years, and the retention period for long-term retention data is set to permanent. Archiving control information is generated based on the matched archiving control rules. Specifically, the retention period, access permissions, download permissions, and review requirements are written into the same control record, forming management constraints after the materials to be archived enter the target archiving location. The archiving basis is generated based on the target archiving location, dynamic archive directory tree, and case file aggregation results. It specifically records the target archive category, directory level, case file number, archiving confidence level, and directory consistency verification results, which are used to explain the source of the determination that the data to be archived is classified into the corresponding location. The trusted verification identifier is formed based on the file hash value, archiving acquisition number and generation timestamp of the data to be archived. When the file hash value is consistent with the verification value before archiving, the trusted verification identifier is marked as valid. A trusted archive package is generated by binding together the data to be archived, a trusted verification identifier, the basis for archiving, and archiving control information. The intelligent archiving results are output based on the trusted archiving package, including the archiving directory results and the archiving description results. The archiving directory results record the final directory location of the data, and the archiving description results record the archiving basis, retention period, and access permissions.

[0033] Example 1: To verify the feasibility of this invention in practice, it was applied to a centralized data archiving scenario. The data to be archived included electronic documents, scanned copies, form attachments, signed documents, and related data generated by different business systems. During implementation, the data to be archived was first parsed and its archiving features were standardized to form a standardized data feature set. Then, the semantic fingerprint, business fingerprint, evidence fingerprint, and version status identification results of the data were extracted using an improved LayoutLMv3 model to form a multi-dimensional profile of the data. Subsequently, an archiving knowledge graph was constructed based on the multi-dimensional profile, and multi-relational archiving reasoning was performed using the CompGCN network to generate a set of candidate archiving locations. Finally, a dynamic archive directory tree and case file aggregation results were generated based on the data association strength to determine the target archiving location and generate a trusted archiving package and intelligent archiving results. The implementation results show that this invention can comprehensively judge the data subject, business matter attribution, evidence source, and version status, reducing the problem of incorrect archiving caused by relying solely on file names, title keywords, or fixed directory rules. It can also output archiving basis, archiving control information, and trusted verification identifiers, improving the standardization, interpretability, and traceability of the data archiving process.

[0034] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent data archiving based on big data analysis, characterized in that, Includes the following steps: S1. Collect the data to be archived and its archiving association information, and generate the original dataset of the data; S2. Perform data content parsing and archiving feature standardization on the original data dataset to generate a standardized data feature set; S3. Input the standardized data feature set into the improved LayoutLMv3 model. The improved LayoutLMv3 model includes a text embedding module, a layout embedding module, a multimodal Transformer encoder, a three-branch structure, and a version status recognition task layer. Introduce a fingerprint decoupling constraint mechanism in the three-branch structure to generate a multi-dimensional profile of the data. S4. Generate data nodes, business item nodes, and archive category nodes based on the multi-dimensional profile of the data, and establish business affiliation relationships and archiving mapping relationships to construct an archive knowledge graph; S5. Input the multi-dimensional profile of the data and the archived knowledge graph into the CompGCN network to perform multi-relation archived reasoning, generate a multi-relation archived reasoning representation, and generate a set of candidate archived locations through the candidate archived location sorting decoder. S6. Generate a dynamic archive directory tree and case file aggregation results based on the candidate archive location set and data association strength, and determine the target archive location; S7. Match archiving control information according to the target archiving location, generate a trusted archiving package, and output the intelligent archiving result.

2. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S1 is: Obtain the documents to be archived, and simultaneously obtain the archiving association information corresponding to the documents to be archived; The archiving association information is determined based on the data source and business matters of the data to be archived; The data to be archived is collected and verified to determine whether the data to be archived has missing content, missing archiving association information, or abnormal data format. The archive association information is bound to the archived data that has passed the data collection and verification, thus forming a correspondence between the archived data and the archive association information; According to the correspondence, the data to be archived and the archive association information are archived, collected, and numbered. The data to be archived and the archive association information that have been archived and numbered are then stored together to generate the original dataset.

3. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S2 is: Read the data to be archived, archive association information, and archive collection number from the original dataset. The data to be archived is parsed to obtain the data content parsing result, which includes the data text content and the data layout information; Based on the archive association information, archive feature fields are extracted from the data content parsing results. The archive feature fields include data subject field, business item field, formation stage field, and archive category field. The archive feature fields are subjected to archive feature standardization processing, which includes unifying field names and standardizing field values; The archive feature fields that have undergone standardized archiving feature processing, the data content parsing results, and the archive collection number are associated to generate a standardized data feature set.

4. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S3 is: Input the standardized data feature set into the improved LayoutLMv3 model; The archive feature fields, data content parsing results and archive collection number are read from the standardized data feature set, and the data text content and data format information are extracted from the data content parsing results. Archive element fields are generated based on the archive feature fields. The text embedding module performs text segmentation, field marking, and ordinal encoding based on the data text content to generate data text embedding features; The layout embedding module divides the page area, marks the area type, and encodes the area position based on the document layout information, and generates document layout embedding features. Determine the archived element anchor points based on the archived element fields, and convert the document layout area corresponding to the archived element anchor points into an enhanced archived layout representation. Input the data text embedding features, data layout embedding features, and archive layout enhancement representation into the multimodal Transformer encoder to generate a joint data representation; The data joint representation is input into a three-branch structure, and a fingerprint decoupling constraint mechanism is introduced into the three-branch structure to perform branch encoding on the data joint representation, generating data semantic fingerprint, data business fingerprint and data evidence fingerprint; Input the joint representation of data, the enhanced representation of archived format, and the fingerprint of data evidence into the version status recognition task layer to generate version status recognition results; By associating the semantic fingerprint, business fingerprint, evidence fingerprint, version status identification result, and archive collection number of the data, a multi-dimensional profile of the data is formed.

5. The intelligent data archiving method based on big data analysis according to claim 4, characterized in that, The fingerprint decoupling constraint mechanism is as follows: The data joint representation is divided into semantic coding branch, business coding branch and evidence coding branch; In the semantic coding branch, the title coding vector, subject word coding vector, core paragraph coding vector, and archive category coding vector in the joint representation of the data are aggregated to generate the data subject coding vector, and the data semantic fingerprint is generated based on the data subject coding vector. In the business coding branch, the business item number coding vector, project number coding vector, contract number coding vector, approval number coding vector and process node coding vector in the data joint representation are aggregated to generate the business item coding vector, and the data business fingerprint is generated based on the business item coding vector. In the evidence coding branch, the data source coding vector, upload subject coding vector, signature status coding vector, document verification result coding vector and formation stage coding vector in the data joint representation are aggregated to generate the evidence source coding vector, and the data evidence fingerprint is generated based on the evidence source coding vector. Calculate the fingerprint similarity among semantic fingerprints, business fingerprints, and evidentiary fingerprints of data; Two coding branches with fingerprint similarity higher than a preset similarity threshold are identified as branches to be decoupled, and repeated coding vectors are extracted from the branches to be decoupled. Based on the contribution ratio of the repeated coding vector in the branch to be decoupled, the repeated coding vector is weighted down to generate the decoupled coding vector; Write the decoupled encoding vector into the corresponding branch to be decoupled, and retain the encoding vectors that are not marked as branches to be decoupled, to generate the data semantic fingerprint, data business fingerprint and data evidence fingerprint after decoupling.

6. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S4 is: Read the multi-dimensional profile of the data, which includes the semantic fingerprint of the data, the business fingerprint of the data, the evidence fingerprint of the data, and the version status identification result; Data nodes are generated based on the data semantic fingerprint, and the data semantic fingerprint is written into the content attributes of the data nodes; Based on the data service fingerprint, a business item node is generated, and a business attribution relationship is established between the data node and the business item node; Generate archive category nodes according to preset archive classification rules, and establish an archiving mapping relationship between the business item nodes and the archive category nodes; Evidence attributes of data nodes are generated based on the data evidence fingerprint, and version attributes of data nodes are generated based on the version status identification results; The data nodes, business item nodes, archive category nodes, business affiliation relationships, and archiving mapping relationships are represented in a graph structure to construct an archiving knowledge graph.

7. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S5 is: Read the multi-dimensional profile of the data and the archived knowledge graph, and map the data semantic fingerprint, data business fingerprint, data evidence fingerprint and version status identification result in the multi-dimensional profile of the data to the data node in the archived knowledge graph to generate an initial representation of the data node; A node set is generated based on the data nodes, business item nodes, and archive category nodes in the archive knowledge graph, and a relationship set is generated based on the business affiliation relationship and the archive mapping relationship; The node set, relation set, and initial representation of data nodes are input into the CompGCN network, and the CompGCN network performs multi-relation archiving reasoning on the archived knowledge graph. In the multi-relation archiving reasoning process, the CompGCN network combines the initial representation of the data node with the relation representation corresponding to the relation set to obtain the updated representation of the data node under the constraints of business item attribution and archive category mapping. Generate a multi-relation archive reasoning representation based on the node update representation; The multi-relation archiving reasoning representation is input into the candidate archiving position sorting decoder, which calculates the sorting score of the data to be archived corresponding to different archiving position items and generates a candidate archiving position set based on the sorting score. The candidate archive location set includes candidate archive location, candidate archive category, candidate directory level, and archive confidence level.

8. The intelligent data archiving method based on big data analysis according to claim 7, characterized in that, The specific steps for calculating the sorting score of the document to be archived corresponding to different archive positions using the candidate archive position sorting decoder are as follows: Archive location items are generated based on the candidate archive categories and the candidate directory levels, and archive location representations are generated based on the archive category nodes corresponding to the archive location items in the archive knowledge graph. The multi-relation archiving reasoning representation and the archiving location representation are input into the candidate archiving location sorting decoder, and the candidate archiving location sorting decoder calculates the business affiliation matching score and archiving rule adaptation score between the data to be archived and the archiving location item. The business affiliation matching score is determined based on the data business fingerprint, the business item node, and the business affiliation relationship. The archiving rule adaptation score is determined based on the data semantic fingerprint, the data evidence fingerprint, the version status identification result, and the archiving mapping relationship; The ranking score is obtained by fusing the business affiliation matching score and the archiving rule adaptation score. The different archiving locations are sorted in descending order based on the ranking score. The archiving locations that meet the preset filtering conditions are selected as candidate archiving locations. An archiving confidence score is generated based on the ranking score. The candidate archiving locations, archiving confidence scores, and the candidate archive categories and candidate directory levels corresponding to the candidate archiving locations are combined to generate a set of candidate archiving locations.

9. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S6 is: Read the candidate archiving location set, and extract the candidate archiving location, candidate archive category, candidate directory level, and archiving confidence level corresponding to the data to be archived from the candidate archiving location set; Calculate the data association strength between the data to be archived based on the multi-dimensional profile of the data and the archived knowledge graph; Based on the strength of the data association, the data to be archived that meet the preset association conditions are divided into the same case file data set, and the case file aggregation result is generated based on the case file data set; A dynamic archive directory tree is generated based on the candidate archive locations and the case file aggregation results; The candidate archive locations are checked for directory consistency based on the dynamic archive directory tree, and the target archive location is determined from the candidate archive locations based on the archive confidence score. The target archiving location is associated with the case file aggregation result to form the archiving location result of the data to be archived.

10. The intelligent data archiving method based on big data analysis according to claim 1, characterized in that, Specifically, S7 is: Read the target archive location, the dynamic archive directory tree, and the case file aggregation result; The archive category and directory level corresponding to the data to be archived are determined based on the target archiving location, and archiving control rules are matched based on the archive category and directory level. Archive control information is generated according to the archive control rules, and the archive control information includes storage control method and permission control method; The archiving criteria are generated based on the target archiving location, the dynamic archive directory tree, and the case file aggregation results. Generate a trusted verification identifier based on the data to be archived, and bind the trusted verification identifier with the archiving basis and archiving control information to generate a trusted archive package; The intelligent archiving result is output based on the trusted archive package. The intelligent archiving result includes the archive directory result and the archive description result.

Images