AI-based archive digitization whole-process management and control method

Abstract

The application discloses an AI-based file digitization whole-process management and control method, comprising the following steps: collecting file image data and text data, constructing an original file data set, and generating an initial directory set; constructing a file directory template set, performing template matching and structure adaptation, and generating an initial structure recognition result; setting a classification rule of directory deviation and structure contradiction, constructing a deviation rule library, performing rule scanning and node comparison, and generating a deviation contradiction label sequence; constructing an evolution state vector and inputting it into an improved GMN model, sequentially passing through a structure guiding unit, a deviation perception unit, a path fusion unit and a node evolution unit, and generating a structure repair state vector; performing result updating and conflict rearrangement, completing standard template alignment and semantic coverage verification, and generating a final structure recognition result. The application improves the structure integrity, semantic consistency and automatic repair capability of directory recognition, and has strong adaptability and intelligent level.

Description

Technical Field

[0001] This invention relates to the field of digital archives processing technology, and in particular to an AI-based method for full-process management and control of archive digitization. Background Technology

[0002] With the rapid growth in demand for digitized and structured management of archives, automatic identification and structure mapping technologies for heterogeneous source catalog data have received continuous attention. Most existing catalog identification methods rely on preset rules or single-path templates to determine the location and structure of catalog fields, but in practical scenarios, they generally suffer from the following problems:

[0003] The high structural complexity and diverse field distribution, especially in historical archives and manually entered data, present structural contradictions such as path intersections, hierarchical conflicts, and field redundancy. Traditional rule-based recognition algorithms struggle to cover all structural deformation types, resulting in low accuracy in structural mapping. Sequential dependencies and hierarchical relationships between directory fields are easily lost after multiple conversions and version iterations. The lack of an effective deviation modeling mechanism during structural alignment fails to adequately detect issues like field offsets, affiliation drift, and path breaks, impacting the overall structural reconstruction. Furthermore, map construction methods generally focus on node connections, neglecting the deep connections between field semantics and path evolution logic. This leads to a lack of unified confidence mechanisms and fusion judgment standards during node repair and structural correction, easily resulting in repair drift or error correction, thus reducing the reliability of the final directory recognition.

[0004] Therefore, how to provide an AI-based method for the whole-process management and control of digital archives is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI-based end-to-end management method for archival digitization. This invention employs an archival structure recognition engine, field attribution determination, structural template matching, abnormal field identification logic, and structural repair status generation to achieve structure recognition, field attribution confirmation, structural template mapping, abnormal node location, and structural repair status output for the original archival dataset. By constructing a set of candidate structural paths, a set of field attribution vectors, and a set of matching labels and difference indicators, field mapping status results are generated and input into the structural repair status generation method to obtain the final structure recognition result. Furthermore, a structural fusion graph is constructed and fed back to the archival catalog template set and deviation rule base to update template weights and rule importance. This invention possesses advantages such as high structure recognition accuracy, accurate field attribution determination, stable template matching effect, rapid anomaly detection response, and intelligent and controllable structural repair process. It can improve the structural integrity and processing efficiency in archival digitization and adapt to the intelligent analysis needs of various types of archival structures.

[0006] The AI-based end-to-end management method for digital archives according to embodiments of the present invention includes the following steps:

[0007] Step 1: Collect archival image data and text data, construct the original archival dataset, and generate an initial catalog set;

[0008] Step 2: Construct a set of archive directory templates, perform template matching and structure adaptation operations on the initial directory set, and generate initial structure recognition results;

[0009] Step 3: Set classification rules for directory deviations and structural contradictions, and build a deviation rule library; the deviation rule library includes field offset rules, hierarchy conflict rules, sequence jump rules, and path ambiguity rules;

[0010] Step 4: Based on the deviation rule base, perform rule scanning and node comparison on the initial structure identification results to generate a deviation contradiction label sequence;

[0011] Step 5: Based on the initial structure identification results and the contradictory label sequence, construct the evolutionary state vector;

[0012] Step 6: Input the evolutionary state vector into the improved GMN model and output the structural repair state vector; the improved GMN model includes a structural guidance unit, a deviation sensing unit, a path fusion unit, and a node evolution unit;

[0013] Step 7: Based on the structure repair state vector, perform result update and conflict rearrangement to generate the corrected structure result; perform standard template alignment and semantic coverage verification on the corrected structure result to generate the final structure recognition result.

[0014] Preferably, step one specifically comprises:

[0015] Set up image acquisition devices and text data acquisition interfaces, unify image format specifications and text field structure specifications, collect archival image data and text data, and construct the original archival dataset;

[0016] The archive image data undergoes a format conversion operation, uniformly converting the images to standard grayscale or binary format, and adjusting the image size ratio and resolution parameters. The converted images are then subjected to character clarity enhancement processing, generating enhanced image data through local contrast enhancement, edge sharpening, and noise suppression methods. Finally, the enhanced image data undergoes image structure region segmentation processing, extracting directory field block regions, inter-field boundary information, and image hierarchical structure information based on structural edge recognition algorithms and projection analysis methods to construct image structure data.

[0017] The text data is processed by field extraction. Based on a preset field set and an entity recognition model, the document number field, title field, responsible person field, time field, category field, and remarks field are identified to generate a field list. The field list is then processed by hierarchical restoration. Based on field name patterns, sequence number features, and punctuation patterns, the hierarchical relationship of the fields is reconstructed to build a text directory hierarchy. The text directory hierarchy is then processed by field position encoding. Based on the hierarchy depth, relative position of the fields, and adjacency relationship, position encoding vectors are generated for each field to form a position encoding matrix.

[0018] By integrating image structure data, field lists, text directory hierarchy, and location encoding matrix, an initial directory set is generated.

[0019] Preferably, step two specifically includes:

[0020] Construct a set of archive catalog templates; the set of archive catalog templates includes standard structure templates, historical archive templates and hybrid derived templates; the standard structure templates are constructed based on the standard archive catalog style, the historical archive templates are extracted based on the catalog structure accumulated in the archive system, and the hybrid derived templates are generated by the structural synthesis and deformation of the standard structure templates and historical archive templates;

[0021] Perform template matching on the initial directory set. Calculate the structural similarity between the initial directory set and each template in the archive directory template set based on structural keywords, field sequences, hierarchical relationships, and positional encoding vectors. Select the template with the highest structural similarity as the initial matching template.

[0022] The structure adaptation operation is performed based on the initial matching template. The structure adaptation operation includes node-level adjustment, missing field completion, and hierarchical structure regularization. Node-level adjustment corrects the field belonging relationship based on field type and position offset. Missing field completion inserts and completes missing fields in the initial directory set based on the structure specifications in the matching template. Hierarchical structure regularization adjusts the field hierarchy in the initial directory set according to the template hierarchical path information.

[0023] The adapted directory structure is merged with the initial directory set to generate an initial structure recognition result; the initial structure recognition result includes a structure path sequence, field level labels, template mapping index, and structure adaptation confidence.

[0024] Preferably, step three specifically includes:

[0025] The system defines directory deviation classification rules and structural contradiction classification rules. The directory deviation classification rules include field position features, field hierarchy labels, field sequence numbers, field path vectors, and template structure numbers. These rules are set based on the mapping differences between fields in the initial structure recognition results and the archive directory template set. The structural contradiction classification rules include field hierarchy relationship features, field parent-child mapping combinations, path conflict labels, and order anomaly indicators. These rules are set based on the inconsistencies between field structure nesting, path continuity, and field sorting logic.

[0026] Construct a deviation rule library; the deviation rule library includes field offset rules, hierarchical conflict rules, sequence jump rules, and path ambiguity rules; field offset rules include field name, field original position index, standard template position index, and allowed offset range; hierarchical conflict rules include field hierarchical number, hierarchical parent-child relationship pairs, legal nested relationship identifiers, and conflict combination identifiers; sequence jump rules include field sequence number, field connection constraint pairs, allowable jump range, and consecutive field quantity threshold; path ambiguity rules include path candidate list, field attribution options, structure selection parameters, and preferred path number.

[0027] Preferably, step four specifically includes:

[0028] Based on the deviation rule base, a rule scanning operation is performed on the initial structure recognition results. The rule scanning operation includes field-level rule matching, path-level structure alignment, and hierarchical relationship retrieval. Field-level rule matching generates field offset labels and sequence anomaly labels based on field names, position indices, and sequence numbers, and compares them with field offset rules and sequence jump rules. Path-level structure alignment compares path vectors with template path numbers to check path consistency and generates path conflict labels according to path ambiguity rules. Hierarchical relationship retrieval generates hierarchical contradiction labels based on the combination structure of field hierarchical numbers and parent-child relationships, and compares them with hierarchical conflict rules.

[0029] Perform node comparison operations, which include node attribute difference detection, structural nesting difference assessment, and node combination legality judgment. Node attribute difference detection is based on the field attributes and the standard attributes of the template fields to perform field consistency comparison. Structural nesting difference assessment is based on the nesting paths in the directory template and the actual nesting paths in the initial structure identification results to perform hierarchical difference analysis. Node combination legality judgment is based on the set of legal nesting combinations set in the deviation rule base to determine whether the field combination meets the standard structure.

[0030] A sequence of discrepancies and contradiction labels is generated based on rule scanning and node comparison operations; the sequence of discrepancies and contradiction labels includes field number, label type, rule number, discrepancy value, contradiction type and confidence level.

[0031] Preferably, the evolutionary state vector includes a structure mapping vector, a semantic weight matrix, a deviation intensity parameter, and a conflict path mapping, specifically:

[0032] The structure mapping vector is based on the field number, path vector and level number in the initial structure recognition result, combined with the standard path number and node sequence number in the directory template, and performs field alignment encoding to generate structure mapping information.

[0033] The deviation intensity vector is constructed based on the deviation value, label type and field number in the deviation contradiction label sequence, and according to the preset weight mapping rule to construct an intensity scalar sequence, marking the degree of structural offset and the degree of order anomaly of each field;

[0034] The contradictory path vector is based on the path number, field combination information and rule number marked in the path conflict label and hierarchical contradiction label, and constructs the path contradiction relationship encoding between fields to generate contradictory path mapping information;

[0035] The structure mapping vector, deviation intensity vector, and contradiction path vector are concatenated in the order of their field numbers to generate the evolutionary state vector.

[0036] Preferably, step six specifically includes:

[0037] The evolutionary state vector is input into the improved GMN model; the improved GMN model includes:

[0038] The structural guidance unit receives the structural mapping vector and the standard path number, standard level number and standard field number information in the directory template, constructs the structural graph alignment matrix, and identifies the initial structural correspondence between fields.

[0039] The deviation sensing unit receives the relationship between the deviation intensity vector and the field order, combines it with the structure graph alignment matrix, performs multi-dimensional deviation attention calculation, and generates a node deviation response matrix.

[0040] The path fusion unit receives conflicting path vectors, constructs a path conflict graph structure between fields, performs path fusion calculations in conjunction with the node deviation response matrix, and generates graph attention fusion results.

[0041] Based on the structural graph alignment matrix, graph attention fusion result, and semantic weight matrix, the node evolution unit performs node state update operations and outputs a structural repair state vector. The structural repair state vector includes field recommendation position numbers, path correction suggestions, and structural evolution confidence parameters.

[0042] Preferably, the improved GMN model includes a structure guidance unit, a deviation sensing unit, a path fusion unit, and a node evolution unit, specifically:

[0043] The structure guidance unit receives the structure mapping vector and the field number, path number and level number in the directory template to construct a standard structure graph. Each field node in the standard structure graph is bound to the standard field number, path number and level number. Each structure edge represents the sequential connection relationship and hierarchical inclusion relationship between fields, and identifies the uniqueness of the field, the path dependency relationship and the level depth.

[0044] The structure guidance unit constructs the input structure graph based on the field number, path number and level number extracted from the structure mapping vector, and calculates the node matching degree and structure edge similarity between the input structure graph and the standard structure graph to generate the structure graph alignment matrix.

[0045] The deviation sensing unit receives the information on the relationship between the deviation intensity vector and the field order, extracts the field number, deviation type and deviation value from the deviation label, and performs a multi-dimensional deviation attention mechanism calculation in combination with the structure graph alignment matrix to generate the deviation attention weight distribution of each field node and construct the node deviation response matrix; the node deviation response matrix characterizes the sensitivity of different field nodes to the influence of deviation in the structure graph.

[0046] The path fusion unit receives conflicting path vectors and field number relationship pairs to construct a path conflict graph structure. Each edge in the path conflict graph structure contains a conflicting path number pair, a conflict type label, and a set of associated field numbers. The path fusion unit performs graph convolution propagation operation in conjunction with the node deviation response matrix and weights and aggregates the graph attention weights of each field node in the context of the conflicting path to generate a graph attention fusion result.

[0047] The node evolution unit constructs a set of field node state vectors based on the structure graph alignment matrix, graph attention fusion results, and semantic weight matrix, and performs node state update operations according to graph adjacency relationships and feature weights. The node evolution unit outputs a structure repair state vector. The structure repair state vector includes the field recommended position number, path correction suggestion, and structure evolution confidence parameter, which correspond to the recommended insertion position of the field node, the associated standard path number, and the confidence evaluation index of the structure correction result, respectively.

[0048] Preferably, step seven specifically comprises:

[0049] Based on the recommended position numbers and path correction suggestions of the fields in the structural repair state vector, and combined with the graph connection relationship between field nodes and the structural evolution confidence parameter, the structural connection order adjustment and field belonging path update operations are performed to generate a preliminary corrected structural graph.

[0050] Based on the path conflict information in the preliminary corrected structure graph, cross-attribution relationship identification and hierarchical conflict detection are performed. According to the frequency of field occurrence, node weight in the structure graph and semantic similarity, the field attribution position is adjusted to generate the corrected structure result.

[0051] Based on the corrected structure result and the set of field numbers, path numbers, and level numbers in the directory template, perform a number consistency comparison and level correspondence verification to generate a standard template mapping table.

[0052] Based on the content vector of each field node in the corrected structure result and the semantic vector corresponding to the template field, a semantic matching degree is calculated to determine whether each template field is uniquely covered; if there are uncovered fields or duplicate covered fields, they are marked as semantic deviation nodes.

[0053] Based on the standard template mapping table and the semantic matching results, the final structure recognition result is output; the structure recognition result includes matching pairs of field numbers and standard field numbers, belonging pairs of path numbers and standard path numbers, and identification information of semantic deviation nodes.

[0054] Preferably, the method further includes: constructing a structural fusion graph based on the final structural recognition result and the original archive dataset, wherein the graph nodes include directory item text, structural hierarchy, semantic labels, and confidence indices; and feeding the structural fusion graph back to the archive directory template set and the deviation rule base to update the template weights and rule importance, specifically:

[0055] By combining the matching relationship between field numbers and standard field numbers in the final structure recognition results, the path number to which the field belongs and the structure deviation identification information, the frequency of occurrence, location information and semantic distribution value of the corresponding fields in the original archive dataset are extracted to generate a structure fusion map containing field nodes and path nodes;

[0056] In the structural fusion graph, each field node and path node is bound to the directory item text content, structural level number, semantic label vector and structural recognition confidence score respectively; the structural connection relationship between fields and the co-occurrence relationship of path sequence are recorded in the graph edges;

[0057] Based on the proportion of field nodes covered by each standard directory template in the structural fusion graph, the accuracy of structural position matching, and the concentration of semantic deviation nodes, the usage coverage and deviation tolerance of each template are calculated, and the template weight parameters in the archive directory template set are updated.

[0058] Based on the path conflict sequences and semantic deviation label information marked in the structural fusion graph, the rule expression forms in the deviation rule base are called, the pattern matching operation is performed, the trigger frequency and influence level of each rule in the conflict samples are counted, and the rule importance score in the deviation rule base is updated.

[0059] The beneficial effects of this invention are:

[0060] This invention constructs a set of candidate structural paths and a set of field attribution vectors. It employs structural hierarchy mapping calculations and field feature aggregation operations to address issues such as differences in directory structure hierarchy, field attribution conflicts, and inconsistent template rule adaptation in original archival datasets. It jointly performs node matching confidence calculations and field distribution deviation assessments, combining semantic tags and contextual information to generate field mapping status results and perform structural repair status generation operations. This achieves unified parsing and attribution determination for multiple types of archival structures. During the construction of the structural fusion graph, structural hierarchy, semantic tags, and confidence indices are introduced to achieve semantic fusion expression of the final structural recognition results. Based on a graph feedback mechanism, the final structural recognition results are fed back into the archival directory template set and deviation rule base, updating template weights and rule importance, effectively improving the dynamic adjustment capability of structural template adaptation and the adaptability of rule base evolution. Ultimately, this invention achieves highly robust identification and repair of structural hierarchy mismatches, ambiguous field attributions, and template matching deviations in original archival datasets, improving the accuracy of structural recognition and the response speed of abnormal fields throughout the entire archival digitization process, and ensuring the consistency of directory hierarchy structures and the integrity of attribution determination. Attached Figure Description

[0061] 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:

[0062] Figure 1 This is a flowchart of the AI-based full-process management and control method for archive digitization proposed in this invention;

[0063] Figure 2 This is a data flow diagram of the AI-based full-process management and control method for archive digitization proposed in this invention. Detailed Implementation

[0064] 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.

[0065] refer to Figure 1 and Figure 2 The AI-based end-to-end management method for digital archives includes the following steps:

[0066] Step 1: Collect archival image data and text data to construct the original archival dataset; perform format conversion, character clarity enhancement, and image structure region segmentation on the image data; perform field extraction, hierarchical restoration, and field position encoding on the text data to generate the initial catalog set;

[0067] Step 2: Construct a set of archive catalog templates, perform template matching and structure adaptation operations on the initial catalog set, and generate initial structure recognition results; the set of archive catalog templates includes standard structure templates, historical archive templates, and hybrid derived templates;

[0068] Step 3: Set classification rules for directory deviations and structural contradictions, and build a deviation rule library; the deviation rule library includes field offset rules, hierarchy conflict rules, sequence jump rules, and path ambiguity rules;

[0069] Step 4: Based on the deviation rule base, perform rule scanning and node comparison on the initial structure identification results to generate a deviation contradiction label sequence;

[0070] Step 5: Based on the initial structure recognition results and the contradictory label sequence, construct an evolutionary state vector; the evolutionary state vector includes a structure mapping vector, a semantic weight matrix, a deviation intensity parameter, and a contradictory path mapping;

[0071] Step 6: Input the evolutionary state vector into the improved GMN model and output the structural repair state vector; the improved GMN model includes a structural guidance unit, a deviation sensing unit, a path fusion unit, and a node evolution unit;

[0072] Step 7: Based on the structure repair state vector, perform result update and conflict rearrangement to generate the corrected structure result; perform standard template alignment and semantic coverage verification on the corrected structure result to generate the final structure recognition result.

[0073] This implementation constructs an original archival dataset by collecting archival image and text data, and then performs format conversion, character clarity enhancement, structural region segmentation, field extraction, hierarchical restoration, and field position encoding to improve the structural clarity and information extraction completeness of multimodal archival data, enhancing the accuracy and hierarchical reproducibility of the initial catalog set. By constructing an archival catalog template set including standard structural templates, historical archive templates, and hybrid derived templates, and performing template matching and structural adaptation operations on the initial catalog set, a standardized and patterned expression of the catalog structure can be achieved, improving the stability of the initial structure recognition results. Furthermore, by setting field offset rules, hierarchical conflict rules, sequence jump rules, and path ambiguity rules, a deviation rule library is constructed. Based on this rule library, rule scanning and node comparison are performed on the initial structure recognition results to further identify structural deviations and contradiction types. Precise annotation improves the efficiency of identifying and locating problems. An evolutionary state vector, comprising a structural mapping vector, semantic weight matrix, deviation intensity parameters, and conflict path mapping, is constructed to comprehensively express the semantic distribution and evolutionary trend of structural problems, providing dynamic guidance information for structural repair. This evolutionary state vector is input into an improved GMN model, where structural alignment representation is completed by the structural guidance unit, difference response features are extracted by the deviation perception unit, and conflict context modeling is enhanced by the path fusion unit. Simultaneously, the node evolution unit outputs a structural repair state vector, further realizing a unified repair expression for multi-source conflicts. Based on the structural repair state vector, result updates and conflict rearrangement are performed to generate corrected structural results. Standard template alignment and semantic coverage verification are then performed, effectively improving the completeness, consistency, and credibility of the final structural identification results, achieving high-precision intelligent identification and repair of archival catalog structures.

[0074] In this embodiment, step one specifically includes:

[0075] By setting up image acquisition devices and text data acquisition interfaces, and unifying image format specifications and text field structure specifications, archival image data and text data are collected to construct the original archival dataset. Format conversion operations are performed on the archival image data to uniformly convert the images to standard grayscale or binary formats, and the image size ratio and resolution parameters are adjusted to adapt to subsequent structure recognition tasks. Character clarity enhancement processing is performed on the format-converted images, improving grayscale level response through local contrast enhancement algorithms, enhancing structural outline clarity through edge sharpening operations, and removing background stray pixels through noise suppression methods to generate enhanced image data.

[0076] Image structural region segmentation is performed on the enhanced image data. A combined strategy based on structural edge recognition algorithm and horizontal and vertical projection analysis method is called to extract the boundary coordinate information of each directory field block in the image, the vertical gap region between adjacent fields, and the hierarchical structure guide lines in the overall image to generate image structural data. To further improve the boundary continuity and inter-block independence of the segmentation results, threshold adjustment and structural posterior filtering methods are introduced. Morphological filtering and connected domain screening are used to optimize the field block contours. By analyzing the alignment relationship and vertical and horizontal density distribution between field block regions, the image hierarchical structure is constructed to form an image structural dataset.

[0077] The text data is processed by field extraction. Based on a preset field set, the target formats for the file number, title, responsible person, time, category, and remarks fields are defined. The field type labeling and extraction tasks are completed by an entity recognition model to generate a field list. The BERT and CRF joint training model is used to obtain the optimal extraction results through entity context embedding and boundary probability fitting, thereby enhancing the robustness of field recognition and cross-category generalization ability. The field list is then processed by hierarchical reconstruction. Based on the field name pattern, serial number features, and punctuation pattern, the hierarchical relationship of the fields is reconstructed through regular expression matching and hierarchical projection to construct the text directory hierarchy structure.

[0078] The text directory hierarchy is subjected to field position encoding. Based on the depth of the field, the relative position index of the field in the current level, and the adjacency relationship with the context fields, the position embedding model is called to generate the field position encoding vector. The position encoding matrix is ​​generated by convolutional pooling function and relation weighting mechanism. The position embedding model adopts the structure-aware path learning method, and obtains the mapping weights and relative position sequences between levels through node path fitting.

[0079] By integrating image structural data, field lists, text directory hierarchy, and positional encoding matrices, an initial directory set is constructed based on a multi-source collaborative structural alignment method. The alignment method calls a graph structure splicing strategy and a semantic embedding alignment mechanism to complete the joint mapping of information from different sources at the field, structural, and semantic levels, outputting an initial directory set with consistent structure, complete information, and clear semantic labels.

[0080] This implementation method can achieve multimodal fusion and structural collaborative expression of archival images and text data under the conditions of unified structural standards and heterogeneous data sources, thereby improving the accuracy of directory structure reconstruction, the stability of field extraction, and the clarity of directory hierarchy expression.

[0081] In this embodiment, step two specifically includes:

[0082] A set of archival catalog templates is constructed, including standard structure templates, historical archive templates, and hybrid derived templates. The standard structure templates are built based on standardized archival catalog styles, utilizing field structures, hierarchical specifications, and field naming rules from published archival compilation standards to generate field hierarchical paths, field constraint sets, and structural positioning rules. Historical archive templates are automatically extracted from existing catalog structures in existing archival systems, using field frequency statistics and hierarchical pattern analysis to construct a set of historical style structures, and combining cluster analysis and structural pattern reduction operations to construct high-frequency structure clusters. Hybrid derived templates are constructed through structural deformation generation and path mapping combinations, building a field mapping graph based on the field similarity relationship between the standard structure template and the historical archive template, and forming derived structure sets through field replacement, path adjustment, and structural splicing.

[0083] Template matching is performed on the initial directory set. The field name sequence, field hierarchy structure, and positional encoding vector from the initial directory set are used as input features, and structural matching is performed with each template in the template set. The structural matching process calculates structural similarity scores using a field similarity model, a path sequence alignment algorithm, and a hierarchical nesting consistency rule. The field similarity model is obtained through semantic embedding space fitting, and path alignment uses a heuristic hierarchical alignment algorithm to construct a minimum-cost mapping path. The structural similarity score includes field name matching degree, hierarchical position consistency, and path order preservation rate. The optimal template is selected as the initial matching template based on the structural similarity score.

[0084] The structural adaptation operation is performed based on the initial matching template. The structural adaptation operation includes node-level adjustment, missing field completion, and hierarchical structure regularization. Node-level adjustment uses a position offset detection algorithm and type matching rules to determine whether there is a deviation in the field's belonging relationship, and adjusts the field node position through a field migration mapping matrix. Missing field completion uses the template structure path to insert field placeholders, and combines the position encoding vector and the template field semantic type to complete the field restoration. Hierarchical structure regularization constructs a hierarchical guidance graph based on the hierarchical path sequence in the matching template, and uses the path backtracking method to adjust the structural nesting relationship of the initial directory fields.

[0085] The initial structure recognition result is generated by fusing the structure adaptation result with the initial directory set. The initial structure recognition result includes the structure path sequence, field level labels, template mapping index, and structure adaptation reliability. The structure path sequence is generated by the field mapping path, the field level labels are labeled based on node depth and template level rules, the template mapping index records the matching relationship between the currently recognized structure and the template, and the structure adaptation reliability is obtained by fitting the field matching score, path retention degree, and level consistency.

[0086] In cases where the original directory structure is incomplete or inconsistent in style, this implementation method can improve the accuracy and robustness of structure identification by using standard structure templates and historical experience templates, ensuring the compatibility and adaptability of the initial identified structure with the archive directory specifications.

[0087] In this embodiment, step three specifically includes:

[0088] The system defines classification rules for directory deviation and structural contradiction. Directory deviation classification rules include field position features, field hierarchical labels, field sequence numbers, field path vectors, and template structure numbers. Field position features are generated based on the field's position in the initial structure recognition result; field hierarchical labels are generated through the hierarchical path after structure adaptation; field sequence numbers are generated by calculating the sequential relationship of fields in the initial structure recognition result; field path vectors are constructed based on the distance matrix between the field's position and adjacent fields; and template structure numbers are the structure identifier numbers in the template matched by the field. Directory deviation classification rules generate a difference mapping matrix based on the mapping difference between the field in the initial structure recognition result and the archive directory template set, and perform classification labeling using a positional similarity evaluation function.

[0089] Structural contradiction classification rules are defined. These rules include field hierarchical relationship features, field parent-child mapping combinations, path conflict labels, and order anomaly indicators. Field hierarchical relationship features are extracted based on the hierarchical jump patterns between fields in the structural path. Field parent-child mapping combinations are obtained by statistically analyzing the parent and child fields corresponding to the fields in the template. Path conflict labels are calculated using path overlap rate and node jump rate. Order anomaly indicators are estimated using field number offset rate and jump anomaly scoring function. The structural contradiction classification rules are based on the logical rationality of field structure nesting, the inconsistency between path continuity and field sorting logic. Rule mapping vectors are constructed and corresponding anomaly types are labeled.

[0090] A deviation rule base is constructed, comprising field offset rules, hierarchical conflict rules, sequence jump rules, and path ambiguity rules. Field offset rules include field name, original field position index, standard template position index, and allowed offset range. The difference between the original field position index and the template position index is the offset, and the allowed offset range is obtained by fitting the template empirical interval. Hierarchical conflict rules include field hierarchy number, hierarchy parent-child relationship pairs, legal nesting relationship identifiers, and conflict combination identifiers. Legal nesting relationship identifiers are based on structural template sample statistics, and conflict combination identifiers are extracted based on sample frequency anomalies. Sequence jump rules include field sequence number, field connection constraint pairs, allowable jump range, and a threshold for the number of consecutive fields. Field connection constraint pairs represent the logical sequential relationship between fields, and the allowable jump range is obtained by fitting the path jump frequency and the degree of semantic coupling with the context. Path ambiguity rules include a path candidate list, field attribution options, structural selection parameters, and priority path numbers. The path candidate list is summarized from historical matching samples, field attribution options are constructed based on the semantic feature distribution during the structural adaptation process, structural selection parameters are calculated using a structural scoring function, and priority path numbers are selected according to matching confidence.

[0091] This implementation method constructs a complete and comprehensive deviation rule library by modeling and classifying deviation rules and structural contradiction classification rules in detail. This provides high-confidence criteria for structural contradiction identification and repair, and improves the accuracy and robustness of complex structure catalog identification process.

[0092] In this embodiment, step four specifically includes:

[0093] The initial structure recognition results are subjected to rule scanning. Deviation and contradiction identification is achieved by setting field-level rule matching, path-level structure alignment, and hierarchical relationship retrieval methods. Field-level rule matching includes comparing field offset rules and sequence jump rules based on field name, position index, and sequence number to generate field offset labels and sequence anomaly labels. Path-level structure alignment includes comparing the structure path vector with the standard template path number and generating path conflict labels based on path ambiguity rules. Hierarchical relationship retrieval includes extracting the field hierarchy number and parent-child mapping combination information, and outputting hierarchical contradiction labels based on hierarchical conflict rules.

[0094] The system performs node comparison operations, defining three judgment methods: node attribute difference detection, structural nesting difference assessment, and node combination legality judgment. Node attribute difference detection compares the standard attributes of fields in the structural recognition results with those in the directory template, outputting attribute consistency comparison results. Structural nesting difference assessment extracts nested paths from the template and compares them with the actual nested paths in the recognition results, constructs a path comparison matrix, and calculates field level difference indices based on the path difference sequence. Node combination legality judgment includes determining whether field combinations meet structural legality requirements based on the set of nested legal combinations defined in the deviation rule base.

[0095] The results of field-level rule matching, path-level structure alignment, and node comparison are fused to construct a deviation and contradiction label sequence. The deviation and contradiction label sequence includes field number, label type, matching rule number, deviation value, contradiction type classification, and confidence level. The deviation value is obtained by fitting the similarity between the original path vector of the field and the template path vector. The confidence level is obtained by comprehensively evaluating the rule hit count and label repetition rate.

[0096] This implementation generates a traceable sequence of discrepancies and contradictions by using multi-dimensional rule matching and structural consistency comparison. This enables fine-grained quality labeling and structural compliance determination of the initial structural identification results, improving the accuracy of understanding the state before structural repair and the precision of repair path planning.

[0097] In this embodiment, the evolutionary state vector includes a structure mapping vector, a semantic weight matrix, a deviation intensity parameter, and a conflict path mapping, specifically:

[0098] The structure mapping vector is based on the field number, path vector, and level number in the initial structure recognition result. Combined with the standard path number and node sequence number in the directory template, a field alignment encoding operation is performed. The field alignment encoding operation includes field position mapping construction, path number mapping combination, and level number difference encoding. Field position mapping construction is based on the difference between the field's position index in the original recognition structure and its position index in the standard template. Path number mapping combination is based on a one-to-one match between the path vector and the standard path number. Level number difference encoding is based on the difference between the level number and the standard level number of the corresponding field in the template, and then normalization is performed. Finally, a structure mapping information vector is generated with the field number as the index.

[0099] The deviation intensity parameter construction process is based on the deviation value, label type and field number in the deviation contradiction label sequence. According to the preset weight mapping rules, an intensity scalar sequence is constructed. The deviation value includes position offset, path offset and hierarchical difference. The label type includes offset type, sequence type and structure type. Each label type corresponds to a set mapping weight coefficient. The label type and deviation value are fitted with corresponding weights. The fitting method adopts a multi-parameter weighted method to combine and map, generating a continuous deviation intensity vector that reflects the degree of field structure offset and the degree of sequence anomaly.

[0100] The contradiction path mapping information is based on the path number, field combination information and rule number recorded in the path conflict label and hierarchical contradiction label. The path contradiction relationship encoding between fields is constructed. The path contradiction relationship encoding includes path conflict matrix generation, field conflict node positioning and structural conflict label index mapping. The path conflict matrix generation establishes a binary conflict mapping table based on the path pairs extracted from the conflict label. The field conflict node positioning determines the structural conflict center based on the intersection of field numbers. The structural conflict label index mapping points to the specific conflict rule content and template number through the structural rule number.

[0101] The structural mapping vector, deviation intensity vector, and contradictory path mapping information are concatenated in the order of field numbers. The concatenation operation includes field number alignment, feature dimension merging, and contradictory information fusion. Field number alignment ensures the consistency of the same field dimension. Feature dimension merging uses a sequence connection method to merge multiple vector features. Contradictory information fusion strengthens the structural anomaly signal through position weight superposition processing, and finally generates a high-dimensional dense evolutionary state vector.

[0102] This implementation generates an evolutionary state vector through multi-dimensional structure mapping, multi-label bias strength calculation, and path contradiction coding operations. This vector can comprehensively characterize the differences and conflicting structures between the initial structure identification results and the template standard, thereby improving the accuracy of structure understanding in map fusion and dynamic diagnostic models.

[0103] In this embodiment, step six specifically includes:

[0104] The structure guidance unit receives the structure mapping vector and the standard path number, standard level number, and standard field number information from the directory template; it establishes a standard structure graph node set based on the field numbers; it constructs a standard structure graph edge set based on the path number and level number; it performs field alignment relationship encoding by combining the field position and path correspondence in the structure mapping vector; the alignment encoding operation uses a matching metric function between the path sequence and the field number, and calculates the structure correspondence probability between each field through a structure similarity score fitting function; it converts the structure correspondence probability matrix into a structure graph alignment matrix to identify the initial structure correspondence between fields;

[0105] The deviation sensing unit receives the deviation intensity vector and the field order relationship; encodes the field order relationship into an order vector matrix; calculates the structural distance weight between field pairs based on the structural graph alignment matrix; constructs a deviation weighting matrix based on the field deviation values ​​and order anomaly identifiers in the deviation intensity vector; performs multi-dimensional deviation attention calculation, which includes a combined weighting of structural distance attention, deviation intensity attention, and order consistency attention; the combined weighting coefficients are obtained by fitting the deviation sample labels and structural sample labels in the training data samples; and outputs a node deviation response matrix, representing the attention intensity of each field node under the background of structural offset and order anomaly.

[0106] The system receives conflicting path vectors through a path fusion unit; constructs a path conflict graph structure based on path number, field combination, and rule number; each edge in the path conflict graph structure represents a path conflict relationship triggered by a rule; introduces path conflict weights and sets path conflict coefficients according to conflict type and rule level; combines the field attention weights in the node deviation response matrix to perform path fusion calculation; the path fusion calculation includes a path conflict propagation mechanism and a node response modulation mechanism; the path conflict propagation mechanism diffuses path anomaly signals based on graph convolution operations; the node response modulation mechanism adjusts the conflict propagation weights based on field deviation intensity; the fusion calculation result forms a graph attention fusion matrix, reflecting the multidimensional response relationship interwoven with structure and deviation between fields;

[0107] The node evolution unit receives the structure graph alignment matrix, graph attention fusion result, and semantic weight matrix. The semantic weight matrix is ​​generated based on the previous semantic distribution calculation module and includes field semantic importance weights and contextual semantic constraint factors. Node state update operations are performed, including structure position adjustment estimation, path correction direction prediction, and semantic consistency enhancement. Structure position adjustment estimation estimates the reasonable position of fields based on the structure graph alignment relationship and graph attention fusion value. Path correction direction prediction calculates the field path suggestion vector based on the path conflict graph and node deviation response. The semantic consistency enhancement operation introduces semantic weights to correct the semantic constraints of field repair suggestions. The above results are integrated to form a structure repair state vector.

[0108] The structural repair state vector includes a field recommended position number, a path correction suggestion, and a structural evolution confidence parameter. The field recommended position number indicates the most likely standard position of the field to be repaired in the directory template. The path correction suggestion is a field movement direction hint based on path fusion. The structural evolution confidence parameter is used to measure the credibility level of the recommended repair results.

[0109] This implementation improves the accuracy and semantic consistency of structural evolution by introducing a structural map alignment mechanism, a multi-dimensional deviation perception mechanism, and a path conflict fusion mechanism, thereby achieving adaptive repair of structural recognition results and improved accuracy of directory recognition evolution.

[0110] In this embodiment, the improved GMN model includes a structure guidance unit, a deviation sensing unit, a path fusion unit, and a node evolution unit, specifically:

[0111] The structure guidance unit receives the structure mapping vector and the field number, path number and level number in the directory template to construct a standard structure graph. Each field node in the standard structure graph is bound to the standard field number, path number and level number. Each structure edge represents the sequential connection relationship and hierarchical inclusion relationship between fields, clarifying the uniqueness of fields, path subordination relationship and hierarchical depth.

[0112] The structure guidance unit constructs an input structure graph based on the field number, path number, and level number extracted from the structure mapping vector. Based on the topological relationship and semantic distribution between field nodes, it obtains the node matching degree and structural edge similarity between the input structure graph and the standard structure graph through a structure similarity function fitting, and generates a structure graph alignment matrix.

[0113] The deviation sensing unit receives the deviation intensity vector and the field order relationship information, extracts the field number, deviation type, and deviation value from the deviation label, and performs deviation weighting on the structure graph alignment matrix based on a multi-channel attention mechanism to construct a multi-dimensional attention mapping. The multi-dimensional attention mapping models the deviation type and field order relationship on each attention channel, and obtains the deviation attention weight distribution of each field node through graph neural network feature propagation path fitting, forming a node deviation response matrix. The node deviation response matrix describes the deviation sensitivity and local deviation amplification effect of the field node in the overall structure.

[0114] The path fusion unit receives conflicting path vectors and field number relationship pairs to construct a path conflict graph structure. Each edge in the path conflict graph structure contains conflicting path number pairs, conflict type labels, and a set of associated field numbers, forming a learnable path conflict representation space. The path fusion unit performs conflict context-aware modeling based on graph convolution operations combined with node bias response matrices. It obtains the aggregation weights of field nodes in the conflicting path context through graph attention mechanisms, and aggregates them to generate graph attention fusion results, expressing the multi-source fusion perception capability of structurally abnormal regions.

[0115] The node evolution unit receives the structure graph alignment matrix, graph attention fusion result, and semantic weight matrix, constructs a set of field node state vectors, performs multiple rounds of state propagation based on graph adjacency relationships, and updates the node states by fusing feature weights to generate a structure repair state vector. The structure repair state vector includes the field recommended position number, path correction suggestion, and structure evolution confidence parameter. The field recommended position number is generated by fitting the optimal insertion position function, the path correction suggestion is constructed based on the conflict path similarity mapping, and the structure evolution confidence parameter is calculated by combining the node consistency distribution and the repair stability index.

[0116] This implementation method enables dynamic adjustment of field positions, path logic, and hierarchical consistency under complex deviation structures, effectively improving the accuracy and stability of directory structure repair.

[0117] In this embodiment, step seven specifically includes:

[0118] Based on the recommended position numbers and path correction suggestions of the fields in the structural repair state vector, and combined with the graph connection relationships between field nodes and the structural evolution confidence parameters, structural connection order adjustment and field belonging path update operations are performed to generate a preliminary corrected structural graph. The structural connection order adjustment is obtained by fitting the structural edge sorting weights between field nodes, and the field belonging path update is obtained by minimizing the structural distance fitting function between the recommended path number and the original path number, ensuring that the field nodes have clear path belonging and sequence positions in the corrected graph.

[0119] Based on the path conflict information in the preliminary corrected structure graph, cross-attribution relationship identification and hierarchical conflict detection are performed. Cross-attribution relationship identification is performed by extracting the intersection region of multiple path numbers belonging to the same field node. Hierarchical conflict detection is based on the deviation judgment of the hierarchical difference between the hierarchical number of the field node and the hierarchical difference of its parent path node. According to the frequency of field occurrence, node weights in the structure graph, and semantic similarity, the field attribution position is adjusted to generate the corrected structure result. Semantic similarity is obtained by fitting the cosine similarity between the field node content vector and the path context vector. Node weights are obtained by weighting the historical recognition stability and semantic matching degree. The field attribution position adjustment is implemented using the minimum cost path reconstruction algorithm.

[0120] Based on the corrected structure results and the field number set, path number set, and level number set in the directory template, a number consistency comparison and level correspondence verification are performed to generate a standard template mapping table. The number consistency comparison generates a one-to-one mapping relationship based on the field number alignment rules. The level correspondence verification uses the node level distance fitting function to calculate the structural level overlap. The standard template mapping table records the standard field number, standard path number, and standard level number corresponding to each field node.

[0121] Based on the content vector of each field node in the corrected structure result and the semantic vector corresponding to the template field, the semantic matching degree is calculated to determine whether each template field is uniquely covered. The semantic matching degree extracts the deep representation of the field content and the template semantics through a bidirectional attention mechanism, and fits the coverage probability through a matching probability function. If there are uncovered fields or repeated covered fields, they are marked as semantic deviation nodes. The semantic deviation nodes are marked with the deviation type and the influence path, and form a structure fallback flag.

[0122] Based on the standard template mapping table and semantic matching results, the final structure recognition result is output; the structure recognition result includes matching pairs of field numbers and standard field numbers, attribution pairs of path numbers and standard path numbers, and identification information of semantic deviation nodes; the structure recognition result provides three-dimensional evaluation indicators of structure attribution accuracy, node matching completeness, and semantic calibration accuracy.

[0123] This implementation method achieves accurate generation of structure recognition results and automatic annotation of deviation nodes by adjusting the structural connection order, updating the path attribution relationship, and calibrating the semantic matching degree. This improves the consistency, completeness, and semantic accuracy of field and path recognition under complex directory structures.

[0124] It also includes: constructing a structural fusion graph based on the final structural recognition result and the original archive dataset, wherein the graph nodes include catalog item text, structural hierarchy, semantic labels, and confidence index; and feeding the structural fusion graph back to the archive catalog template set and the deviation rule base to update the template weights and rule importance, specifically:

[0125] By combining the matching relationship between field numbers and standard field numbers in the final structure recognition results, the path number to which the field belongs and the structure deviation identification information, the frequency of occurrence, location information and semantic distribution value of the corresponding fields in the original archive dataset are extracted to generate a structure fusion map containing field nodes and path nodes;

[0126] In the structural fusion graph, each field node and path node is bound to the directory item text content, structural level number, semantic label vector and structural recognition confidence score respectively; the structural connection relationship between fields and the co-occurrence relationship of path sequence are recorded in the graph edges;

[0127] Based on the proportion of field nodes covered by each standard directory template in the structural fusion graph, the accuracy of structural position matching, and the concentration of semantic deviation nodes, the usage coverage and deviation tolerance of each template are calculated, and the template weight parameters in the archive directory template set are updated.

[0128] Based on the path conflict sequences and semantic deviation label information marked in the structural fusion graph, the rule expression forms in the deviation rule base are called, the pattern matching operation is performed, the trigger frequency and influence level of each rule in the conflict samples are counted, and the rule importance score in the deviation rule base is updated.

[0129] Example 1:

[0130] To verify the feasibility of this invention in practice, it was applied to a provincial-level archives management institution during the process of promoting archives digitization reform. The following typical problems were found: In the original digitization process, the accuracy of archive type identification was insufficient, the control of circulation nodes was loose, the content quality verification method relied on manual labor, the process traceability efficiency was low, and some document archives experienced phenomena such as mismatched identification tags, abnormal loss of task status, and redundant document copies during the circulation process, which seriously affected the progress and quality of archives digitization.

[0131] To address the aforementioned issues, this embodiment introduces an AI-based end-to-end management and control method for digital archives. It constructs an archive management and control process with a unified identification-driven mechanism, a graph-guided modeling mechanism, a multi-state fusion decision-making mechanism, a task process correction mechanism, and a multi-dimensional collection and evaluation mechanism. This enables intelligent operation and closed-loop control in the stages of original archive collection, task modeling, node control, content quality inspection, and result collection.

[0132] In the application process, the standard for inputting original archives is first unified, and recognition-driven instructions are constructed for the original archive information. The instructions include task identification information, archive metadata structure and preset target labels. Archive images and structured content are scheduled, and archive type identification is performed through content matching, style matching and context comparison algorithms. The credibility calculation is performed using the recognition scoring mechanism to generate standardized archive element vectors.

[0133] An archival task graph is constructed based on element vectors. This graph embeds task execution status and workflow information. Control rule vectors are dynamically generated by combining task tags and structure tags, enabling node status judgment, task status prediction, and anomaly linkage control. In actual operation, the system tracks the status of each archival task. For tasks with identification errors, lost status, or execution timeouts, the system invokes the correction module to perform tag backtracking and node reconstruction, significantly reducing result deviations caused by error propagation.

[0134] During three months of actual operation, the original management model and the digital control process after the implementation of this invention were compared, and statistics were compiled on four core indicators: task identification accuracy, task execution closure rate, efficiency of status anomaly repair, and accuracy of result collection. The relevant data is summarized below:

[0135] Table 1. Comparison of Data on Task Identification and Execution in Archival Digitization

[0136] Indicator Name Traditional methods Method of the present invention Increase File type identification accuracy 83.7% 96.4% +12.7% Archive task closed-loop execution rate 75.2% 93.8% +18.6% Status anomaly repair response time 3.6 hours 41 minutes shortened by 79% Tag backtracking success rate 68.5% 92.3% +23.8%

[0137] Table 1 shows that, after adopting the method of this invention, in the digitization of 5,000 archives within three months, the accuracy rate of archive type identification significantly improved from 83.7% to 96.4%, the task closed-loop execution rate increased from 75.2% to 93.8%, the response time for anomaly repair was reduced from an average of 3.6 hours to 41 minutes, and the tag backtracking success rate increased to 92.3%. These results fully demonstrate that the unified identification-driven mechanism and the graph-guided modeling mechanism significantly improve the accuracy of archive type identification and the closed-loop nature of task scheduling, especially achieving significant breakthroughs in task execution continuity and anomaly tracing capabilities.

[0138] Furthermore, in the results collection and quality inspection stage, a multi-dimensional collection and evaluation mechanism is introduced to construct a dynamic quality evaluation index system, which combines dimensions such as process completeness, document consistency, and content logical consistency for quantitative evaluation. A comparative analysis of specific evaluations is as follows:

[0139] Table 2. Data Table for Quality Assessment of Digital Outputs

[0140] Evaluation Dimensions Traditional methods Method of the present invention Increase File structure integrity 82.4 points 95.7 points +13.3 points Content tag consistency 76.1 points 94.5 points +18.4 points Attachment matching accuracy 71.6 points 90.2 points +18.6 points Settlement state stability 79.8 points 96.1 points +16.3 points

[0141] Table 2 reflects the significant improvement in all evaluation dimensions of the digitized output quality after the implementation of the method of this invention. The score for document structure integrity increased from 82.4 to 95.7, the score for content tag consistency jumped from 76.1 to 94.5, and the scores for attachment matching accuracy and collection status stability improved by 18.6 and 16.3 points, respectively. This fully demonstrates that the synergistic effect of the multi-state fusion decision-making mechanism and the collection evaluation mechanism effectively ensures the comprehensive consistency and stability of the digitized archival output in terms of content, structure, and collection.

[0142] The successful deployment and application of the method of this invention in real archive digitization projects not only solves the problems of identification deviation, abnormal flow and content disconnection in the original method, but also realizes refined, closed-loop and intelligent management of the entire task process through AI model construction. It has significant feasibility and practicality and is suitable for promotion and application in archive digitization transformation scenarios at all levels.

[0143] 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. An AI-based method for end-to-end management and control of digital archives, characterized in that: Includes the following steps: Step 1: Collect archival image data and text data, construct the original archival dataset, and generate an initial catalog set; Step 2: Construct a set of archive directory templates, perform template matching and structure adaptation operations on the initial directory set, and generate initial structure recognition results; Step 3: Set classification rules for directory deviations and structural contradictions, and build a deviation rule library; the deviation rule library includes field offset rules, hierarchy conflict rules, sequence jump rules, and path ambiguity rules; Step 4: Based on the deviation rule base, perform rule scanning and node comparison on the initial structure identification results to generate a deviation contradiction label sequence; Step 5: Based on the initial structure identification results and the contradictory label sequence, construct the evolutionary state vector; Step 6: Input the evolutionary state vector into the improved GMN model and output the structural repair state vector; the improved GMN model includes a structural guidance unit, a deviation sensing unit, a path fusion unit, and a node evolution unit; The structural guidance unit receives the structural mapping vector and the standard path number, standard level number and standard field number information in the directory template, constructs the structural graph alignment matrix, and identifies the initial structural correspondence between fields. The deviation sensing unit receives the relationship between the deviation intensity vector and the field order, combines it with the structure graph alignment matrix, performs multi-dimensional deviation attention calculation, and generates a node deviation response matrix. The path fusion unit receives conflicting path vectors, constructs a path conflict graph structure between fields, performs path fusion calculations in conjunction with the node deviation response matrix, and generates graph attention fusion results. Based on the structural graph alignment matrix, graph attention fusion result, and semantic weight matrix, the node evolution unit performs node state update operations and outputs a structural repair state vector. The structural repair state vector includes field recommendation position numbers, path correction suggestions, and structural evolution confidence parameters. Step 7: Based on the structure repair state vector, perform result update and conflict rearrangement to generate the corrected structure result; The corrected structure result is then subjected to standard template alignment and semantic coverage verification to generate the final structure recognition result.

2. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, Step one specifically involves: Set up image acquisition devices and text data acquisition interfaces, unify image format specifications and text field structure specifications, collect archival image data and text data, and construct the original archival dataset; The archive image data undergoes a format conversion operation, uniformly converting the images to standard grayscale or binary format, and adjusting the image size ratio and resolution parameters. The converted images are then subjected to character clarity enhancement processing, generating enhanced image data through local contrast enhancement, edge sharpening, and noise suppression methods. Finally, the enhanced image data undergoes image structure region segmentation processing, extracting directory field block regions, inter-field boundary information, and image hierarchical structure information based on structural edge recognition algorithms and projection analysis methods to construct image structure data. The text data is subjected to field extraction processing. Based on a preset field set and an entity recognition model, the document number field, title field, responsible person field, time field, category field, and remarks field are identified, and a field list is generated. Perform a hierarchical restoration operation on the field list, reconstruct the hierarchical relationship of the fields based on the field name pattern, serial number feature and punctuation pattern, and construct the text directory hierarchical structure; Perform field position encoding on the text directory hierarchy, and generate position encoding vectors for each field based on the hierarchy depth, relative position of the fields and adjacency relationship, forming a position encoding matrix; By integrating image structure data, field lists, text directory hierarchy, and location encoding matrix, an initial directory set is generated.

3. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, Step two specifically involves: Construct a set of archive catalog templates; the set of archive catalog templates includes standard structure templates, historical archive templates and hybrid derived templates; the standard structure templates are constructed based on the standard archive catalog style, the historical archive templates are extracted based on the catalog structure accumulated in the archive system, and the hybrid derived templates are generated by the structural synthesis and deformation of the standard structure templates and historical archive templates; Perform template matching on the initial directory set. Calculate the structural similarity between the initial directory set and each template in the archive directory template set based on structural keywords, field sequences, hierarchical relationships, and positional encoding vectors. Select the template with the highest structural similarity as the initial matching template. Based on the initial matching template, structural adaptation operations are performed, including node-level adjustment, missing field completion, and hierarchical structure regularization. Node-level adjustments correct field affiliation relationships based on field type and position offset; missing field completion inserts and completes missing fields in the initial directory set based on the structural specifications in the matching template. The hierarchical structure is standardized by adjusting the field hierarchy in the initial directory set based on the template hierarchical path information; The adapted directory structure is merged with the initial directory set to generate the initial structure recognition result. The initial structure identification results include structure path sequence, field level labels, template mapping index, and structure adaptation confidence.

4. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, Step three specifically involves: Define directory deviation classification rules and structural contradiction classification rules; the directory deviation classification rules include field position features, field level labels, field sequence numbers, field path vectors, and template structure numbers; The directory deviation classification rule is set based on the mapping difference between the field in the initial structure recognition result and the archive directory template set; the structural contradiction classification rule includes field hierarchical relationship features, field parent-child mapping combination, path conflict label and order anomaly index; the structural contradiction classification rule is set based on the inconsistency relationship between field structure nesting, path continuity and field sorting logic; Construct a deviation rule library; the deviation rule library includes field offset rules, hierarchical conflict rules, sequence jump rules, and path ambiguity rules; field offset rules include field name, field original position index, standard template position index, and allowed offset range; hierarchical conflict rules include field hierarchical number, hierarchical parent-child relationship pairs, legal nested relationship identifiers, and conflict combination identifiers; sequence jump rules include field sequence number, field connection constraint pairs, allowable jump range, and consecutive field quantity threshold; path ambiguity rules include path candidate list, field attribution options, structure selection parameters, and preferred path number.

5. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, Step four specifically involves: Based on the deviation rule base, a rule scanning operation is performed on the initial structure recognition result. The rule scanning operation includes field-level rule matching, path-level structure alignment, and hierarchical relationship retrieval. Field-level rule matching generates field offset labels and order anomaly labels by comparing field offset rules and order jump rules with field offset rules and order anomaly labels. Path-level structural alignment is based on comparing the path vector with the template path number to ensure path consistency, and generates path conflict labels according to path ambiguity rules. Hierarchical relationship retrieval is based on the combination structure of field hierarchical number and parent-child relationship, and generates hierarchical contradiction tags by referring to hierarchical conflict rules; Perform node comparison operations, which include node attribute difference detection, structural nesting difference evaluation, and node combination legality judgment. Node attribute difference detection performs a field consistency comparison based on field attributes and template field standard attributes; structural nesting difference assessment performs a hierarchical difference analysis based on the nested paths in the directory template and the actual nested paths in the initial structure identification results; node combination legality judgment judges whether the field combination meets the standard structure based on the set of legal nested combinations set in the deviation rule base. A sequence of discrepancies and contradiction labels is generated based on rule scanning and node comparison operations; the sequence of discrepancies and contradiction labels includes field number, label type, rule number, discrepancy value, contradiction type and confidence level.

6. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, The evolutionary state vector includes a structure mapping vector, a semantic weight matrix, a deviation intensity parameter, and a conflict path mapping, specifically: The structure mapping vector is based on the field number, path vector and level number in the initial structure recognition result, combined with the standard path number and node sequence number in the directory template, and performs field alignment encoding to generate structure mapping information. The deviation intensity vector is constructed based on the deviation value, label type and field number in the deviation contradiction label sequence, and according to the preset weight mapping rule to construct an intensity scalar sequence, marking the degree of structural offset and the degree of order anomaly of each field; The contradictory path vector is based on the path number, field combination information and rule number marked in the path conflict label and hierarchical contradiction label, and constructs the path contradiction relationship encoding between fields to generate contradictory path mapping information; The structure mapping vector, deviation intensity vector, and contradiction path vector are concatenated in the order of their field numbers to generate the evolutionary state vector.

7. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, The improved GMN model includes a structure guidance unit, a deviation sensing unit, a path fusion unit, and a node evolution unit, specifically: The structure guidance unit receives the structure mapping vector and the field number, path number and level number in the directory template to construct a standard structure graph. Each field node in the standard structure graph is bound to the standard field number, path number and level number. Each structure edge represents the sequential connection relationship and hierarchical inclusion relationship between fields, and identifies the uniqueness of the field, the path dependency relationship and the level depth. The structure guidance unit constructs the input structure graph based on the field number, path number and level number extracted from the structure mapping vector, and calculates the node matching degree and structure edge similarity between the input structure graph and the standard structure graph to generate the structure graph alignment matrix. The deviation sensing unit receives the deviation intensity vector and the field order relationship information, extracts the field number, deviation type and deviation value from the deviation label, and performs multi-dimensional deviation attention mechanism calculation in combination with the structure graph alignment matrix to generate the deviation attention weight distribution of each field node and construct the node deviation response matrix. The node deviation response matrix characterizes the sensitivity of nodes in different fields to deviations in the structure graph; The path fusion unit receives the relationship pairs between conflicting path vectors and field numbers, and constructs a path conflict graph structure. Each edge in the path conflict graph structure contains a pair of conflict path numbers, a conflict type label, and a set of associated field numbers. The path fusion unit performs graph convolution propagation operation in conjunction with the node deviation response matrix, and weights and aggregates the graph attention weights of each field node in the context of the conflict path to generate the graph attention fusion result. The node evolution unit constructs a set of field node state vectors based on the structure graph alignment matrix, graph attention fusion results, and semantic weight matrix, and performs node state update operations according to graph adjacency relationships and feature weights. The node evolution unit outputs a structure repair state vector. The structure repair state vector includes the field recommended position number, path correction suggestion, and structure evolution confidence parameter, which correspond to the recommended insertion position of the field node, the associated standard path number, and the confidence evaluation index of the structure correction result, respectively.

8. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, Step seven specifically involves: Based on the recommended position numbers and path correction suggestions of the fields in the structural repair state vector, and combined with the graph connection relationship between field nodes and the structural evolution confidence parameter, the structural connection order adjustment and field belonging path update operations are performed to generate a preliminary corrected structural graph. Based on the path conflict information in the preliminary corrected structure graph, cross-attribution relationship identification and hierarchical conflict detection are performed. According to the frequency of field occurrence, node weight in the structure graph and semantic similarity, the field attribution position is adjusted to generate the corrected structure result. Based on the corrected structure result and the set of field numbers, path numbers, and level numbers in the directory template, perform a number consistency comparison and level correspondence verification to generate a standard template mapping table. Based on the content vector of each field node in the corrected structure result and the semantic vector corresponding to the template field, the semantic matching degree is calculated to determine whether each template field is uniquely covered. If there are uncovered fields or duplicate covered fields, mark them as semantic deviation nodes; Based on the standard template mapping table and the semantic matching results, the final structure recognition result is output; the structure recognition result includes matching pairs of field numbers and standard field numbers, belonging pairs of path numbers and standard path numbers, and identification information of semantic deviation nodes.

9. The AI-based full-process management and control method for archival digitization according to claim 1, characterized in that, Also includes: Based on the final structure recognition result and the original archive dataset, a structure fusion graph is constructed. Graph nodes include directory item text, structural hierarchy, semantic labels, and confidence indices. The structure fusion graph is then fed back to the archive directory template set and the deviation rule base to update template weights and rule importance. Specifically: By combining the matching relationship between field numbers and standard field numbers in the final structure recognition results, the path number to which the field belongs and the structure deviation identification information, the frequency of occurrence, location information and semantic distribution value of the corresponding fields in the original archive dataset are extracted to generate a structure fusion map containing field nodes and path nodes; In the structural fusion graph, each field node and path node is bound to the directory item text content, structural level number, semantic label vector and structural recognition confidence score respectively; the structural connection relationship between fields and the co-occurrence relationship of path sequence are recorded in the graph edges; Based on the proportion of field nodes covered by each standard directory template in the structural fusion graph, the accuracy of structural position matching, and the concentration of semantic deviation nodes, the usage coverage and deviation tolerance of each template are calculated, and the template weight parameters in the archive directory template set are updated. Based on the path conflict sequences and semantic deviation label information marked in the structural fusion graph, the rule expression forms in the deviation rule base are called, the pattern matching operation is performed, the trigger frequency and influence level of each rule in the conflict samples are counted, and the rule importance score in the deviation rule base is updated.

Images