An AI agent-based historical archive format automatic detection and conversion method
By adopting an AI-based automatic file format detection and conversion method, the problems of low efficiency and inaccurate conversion in existing technologies have been solved, achieving efficient and accurate file format detection and conversion, and ensuring that files meet the requirements for long-term preservation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUJIAN YIRONG INFORMATION TECH
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies are inefficient and inaccurate in processing massive amounts of historical archive formats. Automated tools have limited functionality and lack effective automatic conversion solutions, making it difficult to achieve standardized conversion of archive formats.
This approach employs an AI-based intelligent agent method. By constructing an archive format feature database and a rule engine, and combining deep learning and reinforcement learning algorithms, it automatically detects and converts archive formats. It utilizes OCR and NLP technologies to parse unstructured data, constructs a format meta-model, and dynamically adjusts the rule execution order in the rule base for adaptive intelligent conversion.
It enables efficient and accurate detection and conversion of archive formats, saving time and manpower costs, ensuring that archive formats meet long-term preservation requirements, and improving the quality and reliability of archives.
Smart Images
Figure CN122432401A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of archival digitization technology, and in particular to a method for automatic detection and conversion of historical archive formats based on AI intelligent agents. Background Technology
[0002] Archives, as historical records of various forms directly created by past and present government agencies, organizations, enterprises, institutions, and individuals in their activities, possess preservation value for the nation and society. Therefore, the standardization of their format is crucial. In today's digital age, massive amounts of historical archives exist in various formats, such as different versions of Word, Excel, and PDF. However, existing technologies face numerous challenges in processing these vast amounts of historical archive formats.
[0003] On the one hand, manual inspection of archival formats is extremely inefficient. Faced with massive amounts of archives, a one-by-one manual check would consume enormous amounts of time and manpower and is prone to oversights. For example, in a large archive containing millions of historical documents, manually checking whether their formats meet long-term preservation requirements could take years to complete even a preliminary inspection, and the accuracy of the inspection is difficult to guarantee. On the other hand, existing automated inspection tools have limited functionality, only able to detect some common formatting issues, and cannot comprehensively and thoroughly check whether the archival formats meet the complex requirements of long-term preservation. Furthermore, for archives whose formats do not meet long-term preservation requirements, there is a lack of effective automatic conversion methods, making it difficult to achieve standardized conversion of archival formats to meet the needs of long-term preservation.
[0004] Therefore, how to develop an automatic detection and conversion method for historical archive formats based on AI agents is a technical problem that needs to be solved. Summary of the Invention
[0005] To address the aforementioned issues, the present invention aims to provide an automatic detection and conversion method for historical archive formats based on AI intelligent agents, thereby resolving the problems of inefficiency in manual detection of massive historical archives, limited functionality of existing automation tools, and lack of effective automatic conversion solutions.
[0006] To achieve the above objectives, the present invention adopts the following technical solution, comprising the following steps: Step S1: Collect historical archive samples from archival storage institutions, preprocess the data of the samples, extract the characteristic information of the archives, and construct an archive format characteristic database; Step S2: Develop conversion rules based on the archival storage format specifications; specifically including: Step S21: The system receives the structured format specification document and parses the unstructured standard manual using OCR and NLP technologies; Step S22: Construct a format metamodel that includes basic attributes, relational constraints describing spatial and hierarchical relationships between archive elements, and semantic constraints describing the content format and logical relationships of archive elements; Step S23: Map the completed format metamodel to the preset rule engine, and convert the attributes and constraints in the format metamodel into rules to form the initial rule base; Step S24: Based on historical archive detection data containing manually labeled error samples, train a rule induction model, automatically discover and generate implicit transformation rules through the algorithm, and supplement them to the initial rule base; Step S25: Establish a rule priority evaluation mechanism, using the rule triggering frequency, execution accuracy, and impact scope as evaluation indicators to dynamically adjust the execution order of rules in the rule base; Step S26: Use the results of manual review and the data format of manually corrected archives as new training data, and feed them back into the rule induction model to optimize the model parameters and rule base content; Step S3: Combining the archive format feature database, a deep learning algorithm is used to compare the archive to be tested with the standard format in the format feature database. The archive format is analyzed. If the archive format is qualified, the qualified archive is directly stored in the archive database. If the archive format is not qualified, the unqualified archive is adaptively and intelligently converted. The converted archive is compared with the standard format in the format feature database again. After confirming that the format is qualified, the qualified archive is stored in the archive database.
[0007] Furthermore, in step S1, the preprocessing process is as follows: the collected historical archive samples are processed to unify their format, and noise data in the samples is removed to ensure the accuracy of feature extraction and the integrity of the archive format feature database.
[0008] Furthermore, step S24 specifically includes the following steps: Step S241: Standardize the marked error samples and extract the error type and archive metadata; where the error type includes missing fields, format misalignment and encoding error, and the archive metadata includes format type, year and storage medium; Step S242: Use the FP-Growth algorithm to mine erroneous features and correct strong association rules for actions; use the C4.5 decision tree to generate conditional branch rules; specifically, the core indicator calculation formula for association rule mining is as follows: Support is calculated as Support(A→B) = count(A∩B) / count(D); The confidence level is Confidence(A→B) = count(A∩B) / count(A); The lift is calculated as Lift(A→B) = Confidence(A→B) / Support(B). Where A is the set of erroneous features, B is the corrective action, and D is the total number of samples; Support(A→B) is the coverage of rule A→B in the overall sample; count(A∩B) is the number of samples in the intersection of A and B; count(D) is the total number of valid samples participating in the association rule analysis; Confidence(A→B) is the probability that corrective action B is executed when erroneous feature A occurs; count(A) is the number of samples in the total sample set D containing erroneous feature set A; Lift(A→B) is the ratio of the confidence of rule A→B to the support of corrective action B itself; Support(B) is the proportion of samples in the total sample set D containing corrective action B to the total number of samples; where Lift>1 indicates that the rule is valid. Step S243: Extract candidate rules from the hybrid model. The format of the candidate rules is: IF condition set THEN correction action. Use three indicators, support, confidence and lift, to screen effective rules and eliminate redundant or inefficient rules.
[0009] Furthermore, step S25 specifically includes the following steps: Step S251: Insert points in the rule engine to collect the running data of each rule in real time and summarize the indicator data on a regular basis. The transportation data specifically includes the number of triggers Nt, the number of successful executions Ns, and the number of affected files Na. Step S242: Standardize the transportation data and construct a weighted scoring model to calculate the priority score of each rule and set the weight coefficients; Step S243: Update the rule execution order of the rule engine according to priority scores.
[0010] Furthermore, in step 23, the rule engine is Drools or a custom graph rule engine.
[0011] Further, step S31: Construct a multimodal feature fusion network, including a visual branch and a structural branch, and fuse the features extracted from the visual branch and the structural branch to form a joint feature representation that is sensitive to the file format and file content; Step S32: Compare the joint features of the file to be inspected with the standard features in the file format feature database, and determine whether the file format is qualified based on the comparison results; if the file format is qualified, the qualified file is directly stored in the archive database; if the file format is unqualified, a structured fine-grained defect report is output. Furthermore, in step S3, if the file format is unqualified, an adaptive intelligent conversion is performed on the unqualified file. The specific steps are as follows: Step S33: Construct a reinforcement learning environment, define the state of the reinforcement learning environment as the fine-grained defect report of the current file and the current format state of the file, define the action space as the conversion operation that the AI agent can execute, and construct a multi-dimensional reward function, and associate the reward value of the multi-dimensional reward function with the repair accuracy, execution efficiency and fidelity of the original content of the conversion operation. Step S34: The AI agent learns and builds an optimal policy network through a reinforcement learning environment, enabling it to predict and output the sequence of transformation actions that can obtain the maximum cumulative reward based on the current defect status of the file, and execute the corresponding transformation operations in sequence. Step S35: Perform a format check on the corrected file again; if the check fails, the new defect status is fed back to the AI agent, which makes decisions and repairs based on the new status until the file format is approved.
[0012] Furthermore, in step S31, a multimodal feature fusion network is constructed, including a visual branch and a structural branch, as detailed below: The visual branch uses a Transformer-based document understanding model, taking the document image, text content, and two-dimensional layout information of the archive to be detected as input, and capturing the global format features and semantic relationships of the archive through a self-attention mechanism; the structural branch uses a graph neural network to model archives with a clear hierarchical structure, analyzes the parent-child and sibling relationships between archive elements, and judges the compliance of the archive's logical structure.
[0013] Furthermore, in step S32, the structured fine-grained defect report specifically includes defect ID, defect type, pixel-level visual marker location, document structure tree path, problem description, and confidence level.
[0014] The present invention has the following beneficial effects: 1- This invention utilizes big data and intelligent technology to efficiently complete the detection and conversion of tens of thousands of files, significantly saving time and manpower costs.
[0015] 2. This invention, through big data analysis and intelligent algorithms, can comprehensively and accurately detect document format issues and convert them according to strict standards. This avoids omissions and errors that may occur with manual inspection, ensuring that document formats meet long-term preservation requirements and improving the quality and reliability of the documents. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the present invention. Figure 1 ; Detailed Implementation
[0017] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: See Figure 1 As shown, the solution includes the following steps: Step S1: Collect historical archive samples from archival storage institutions, preprocess the data of the samples, extract the characteristic information of the archives, and construct an archive format characteristic database; Step S2: Develop conversion rules based on the archival storage format specifications; specifically including: Step S21: The system receives the structured format specification document and parses the unstructured standard manual using OCR and NLP technologies; Step S22: Construct a format metamodel that includes basic attributes, relational constraints describing spatial and hierarchical relationships between archive elements, and semantic constraints describing the content format and logical relationships of archive elements. Relational constraints describe spatial and hierarchical relationships between elements, such as "titles must be above the main text", "page numbers must be in the lower right corner of the page footer", and "table headers and content must be aligned". Semantic constraints describe the content and logical relationships of elements, such as "date format must be YYYY-MM-DD" and "the 'serial number' field of the volume's file directory must be consecutive numbers".
[0018] Step S23: Map the completed format meta-model to the preset rule engine, and transform the attributes and constraints in the format meta-model into rules to form an initial rule base; further, the rule induction model adopts a decision tree algorithm or an association rule mining algorithm. Specifically, the decision tree algorithm extracts rules from the association of "error type-correction method" by constructing a classification tree; the association rule mining algorithm (such as the Apriori algorithm) discovers hidden association patterns from a large number of "error-correction" cases (such as "no blank line after table" is often accompanied by "no indentation of the first line of the main text", thus generating implicit rules that are difficult to exhaustively enumerate by experience, such as "the first line indentation needs to be checked synchronously after fixing the table spacing".
[0019] Step S24: Based on historical archive detection data containing manually labeled error samples, train a rule induction model, automatically discover and generate implicit transformation rules through the algorithm, and supplement them to the initial rule base; Step S25: Establish a rule priority evaluation mechanism, using the triggering frequency, execution accuracy, and impact scope of rules as evaluation indicators to dynamically adjust the execution order of rules in the rule base; when conflicts are detected between rules, use reinforcement learning algorithms to test the execution effect of different rule combinations in a simulated environment, learn and determine the optimal conflict resolution strategy; Step S26: Use the results of manual review and the data format of manually corrected archives as new training data, and feed them back into the rule induction model to optimize the model parameters and rule base content; Step S3: Combining the archive format feature database, a deep learning algorithm is used to compare the archive to be tested with the standard format in the format feature database. The archive format is analyzed. If the archive format is qualified, the qualified archive is directly stored in the archive database. If the archive format is not qualified, the unqualified archive is adaptively and intelligently converted. The converted archive is compared with the standard format in the format feature database again. After confirming that the format is qualified, the qualified archive is stored in the archive database.
[0020] Furthermore, in step S1, the preprocessing process is as follows: the collected historical archive samples are processed to unify their format, and noise data in the samples is removed to ensure the accuracy of feature extraction and the integrity of the archive format feature database.
[0021] Furthermore, step S24 specifically includes the following steps: Step S241: Standardize the marked error samples and extract the error type and archive metadata; where the error type includes field missing, format misalignment and encoding error, and the archive metadata includes format type, year and storage medium; specifically, align the error feature vector of the error sample with the correction action label of the corrected sample to construct a mapping dataset D={(Xi,Yi)}, where Xi is the error feature vector containing error type, error location and archive metadata, and Yi is the correction action label.
[0022] Step S242: Employ the FP-Growth algorithm to mine error features and strong association rules for corrective actions. Specifically, firstly, the mapping dataset is transformed into a Boolean transaction-itemset matrix, and Boolean one-hot encoding is performed on the error features and corrective actions, forcing the corrective action item to be a fixed suffix item of the transaction. Secondly, an optimized FP-Tree for archival data is constructed, with the corrective action item given the highest priority in the item header table, and the remaining error features arranged in descending order of support. Extremely low-frequency error feature items with support below min_support / 2 are pre-pruned. Finally, frequent itemsets containing corrective action items are recursively mined from the optimized FP-Tree, and candidate rules with antecedents being error features and consequents being corrective actions are generated in a targeted manner. Coarse-grained strong association rules that satisfy minimum support min_support=0.02, minimum confidence=0.9, and minimum lift=1.2 are retained.
[0023] Specifically, the core metrics for association rule mining are calculated using the following formulas: Support is calculated as Support(A→B) = count(A∩B) / count(D); The confidence level is Confidence(A→B) = count(A∩B) / count(A); The lift is calculated as Lift(A→B) = Confidence(A→B) / Support(B). Where A is the set of erroneous features, B is the corrective action, and D is the total number of samples; Support(A→B) is the coverage of rule A→B in the overall sample; count(A∩B) is the number of samples in the intersection of A and B; count(D) is the total number of valid samples participating in the association rule analysis; Confidence(A→B) is the probability that corrective action B is executed when erroneous feature A occurs; count(A) is the number of samples in the total sample set D containing erroneous feature set A; Lift(A→B) is the ratio of the confidence of rule A→B to the support of corrective action B itself; Support(B) is the proportion of samples in the total sample set D containing corrective action B to the total number of samples; where Lift>1 indicates that the rule is valid. The C4.5 decision tree is used to generate conditional branching rules. Specifically, firstly, for each coarse-grained strong association rule, fine-grained conditional features are expanded. These fine-grained conditional features include features from the OCR recognition parameter dimension, the archive metadata extension dimension, and the file system attribute dimension. Redundant features with a correlation less than a preset value to the "whether to perform correction action" label are removed using the Pearson correlation coefficient. Secondly, the samples containing the error feature set of the coarse-grained strong association rule are divided into independent datasets. The information gain ratio is used as the splitting criterion. If the difference between the feature with the highest information gain ratio and the second highest information gain ratio is less than a preset value, the feature with a clearer business meaning is selected first. A pessimistic pruning strategy is used to prune the decision tree, pruning branches with fewer than 5 leaf node samples, with a pruning threshold of 0.1. Finally, the path corresponding to "execute correction action" generated by the C4.5 decision tree is used as a fine condition set, connected with the antecedent of the coarse-grained strong association rule using "AND". If multiple paths are generated, the path with the highest accuracy on the validation set is retained. If multiple paths have the same accuracy, the path with the fewest conditions is retained.
[0024] Step S243: Extract candidate rules from the hybrid model. The format of the candidate rules is: IF condition set THEN correction action. Use three indicators, namely support, confidence and lift, to screen effective rules and eliminate redundant or inefficient rules.
[0025] Furthermore, step S25 specifically includes the following steps: Step S251: Insert points in the rule engine to collect the running data of each rule in real time and summarize the indicator data on a regular basis. The transportation data specifically includes the number of triggers Nt, the number of successful executions Ns, and the number of affected files Na. Step S242: Standardize the transportation data and construct a weighted scoring model to calculate the priority score of each rule, setting weight coefficients. Specifically, the trigger frequency weight is w1=0.3, the execution accuracy weight is w2=0.5, and the impact range weight is w3=0.2, and w1+w2+w3=1. Specifically, Min-Max standardization is used to standardize the trigger frequency, execution accuracy, and impact range indicators to eliminate differences in units. Specifically, the core calculation formula for indicator standardization is: X norm =(XX min ) / (X max -X min ), Where X is the original index value, X min X max These are the minimum and maximum values of the indicator across all rules, respectively. The core formula for calculating rule priority score is: S=w1×F norm +w2×A norm +w3×R norm ; Where F norm The standardized trigger frequency is F = Nt / ∑Nt; A norm For the standardized execution accuracy, A = Ns / Nt; R norm The range of influence after standardization is R = Na / ∑Na.
[0026] Step S243: Update the rule execution order of the rule engine according to priority scores. Specifically, the rules are divided into three priority levels: the top 20% are high priority, the middle 60% are medium priority, and the bottom 20% are low priority; before executing low priority rules, check for conflicts with high priority rules, and if a conflict exists, execution is suspended.
[0027] Further, step S31: Construct a multimodal feature fusion network, including a visual branch and a structural branch, and fuse the features extracted from the visual branch and the structural branch to form a joint feature representation that is sensitive to the file format and file content; Step S32: Compare the joint features of the file to be inspected with the standard features in the file format feature database, and determine whether the file format is qualified based on the comparison results; if the file format is qualified, the qualified file is directly stored in the archive database; if the file format is unqualified, a structured fine-grained defect report is output. Furthermore, in step S3, if the file format is unqualified, an adaptive intelligent conversion is performed on the unqualified file. The specific steps are as follows: Step S33: Construct a reinforcement learning environment. Define the state of the reinforcement learning environment as the fine-grained defect report of the current file and the current format state of the file. Define the action space as the conversion operations that the AI agent can execute. At the same time, construct a multi-dimensional reward function and associate the reward value of the multi-dimensional reward function with the repair accuracy, execution efficiency and fidelity of the original content of the conversion operation. When an action is executed, if the defect is repaired, a positive reward is given; if a new error is introduced or the repair effect is poor, a negative reward is given.
[0028] Step S34: The AI agent learns and builds an optimal policy network through a reinforcement learning environment, enabling it to predict and output the sequence of transformation actions that can obtain the maximum cumulative reward based on the current file's defect status, and execute the corresponding transformation operations in sequence. Specifically, when the two problems of "too small spacing between title and body text" and "no indentation of the first line of body text" are detected, the agent may first execute "increase the spacing after paragraphs" and then execute "set the first line indentation", and evaluate the effect of each step, rather than simply executing them in a fixed order.
[0029] Step S35: Perform a format check on the corrected file again; if the check fails, the new defect status is fed back to the AI agent, which makes decisions and repairs based on the new status until the file format is approved.
[0030] Furthermore, in step S31, a multimodal feature fusion network is constructed, including a visual branch and a structural branch, as detailed below: The visual branch uses a Transformer-based document understanding model (such as LayoutLM, Donut) as input, taking the document image, text content, and two-dimensional layout information of the archive to be detected as input, and capturing the global format features and semantic relationships of the archive through a self-attention mechanism; the structural branch uses a graph neural network to model archives with a clear hierarchical structure, analyze the parent-child and sibling relationships between archive elements, and determine the compliance of the archive's logical structure.
[0031] Furthermore, in step S32, the structured fine-grained defect report specifically includes a defect ID (unique identifier), defect type (such as "font error", "alignment deviation", "missing element", "semantic mismatch"), pixel-level visual marker position, document structure tree path, problem description (such as "the font of the title 'Chapter 1' should be 'Heiti No. 2', but it is actually 'Songti No. 2'") and confidence level (the certainty of the model's judgment on the defect).
[0032] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0033] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0034] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0035] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for automatic detection and conversion of historical archive formats based on AI intelligent agents, characterized in that, Includes the following steps: Step S1: Collect historical archive samples from archival storage institutions, preprocess the data of the samples, extract the characteristic information of the archives, and construct an archive format characteristic database; Step S2: Develop conversion rules based on the archival storage format specifications; specifically including: Step S21: The system receives the structured format specification document and parses the unstructured standard manual using OCR and NLP technologies; Step S22: Construct a format metamodel that includes basic attributes, relational constraints describing spatial and hierarchical relationships between archive elements, and semantic constraints describing the content format and logical relationships of archive elements; Step S23: Map the completed format metamodel to the preset rule engine, and convert the attributes and constraints in the format metamodel into rules to form the initial rule base; Step S24: Based on historical archive detection data containing manually labeled error samples, train a rule induction model, automatically discover and generate implicit transformation rules through the algorithm, and supplement them to the initial rule base; Step S25: Establish a rule priority evaluation mechanism, using the rule triggering frequency, execution accuracy, and impact scope as evaluation indicators to dynamically adjust the execution order of rules in the rule base; Step S26: Use the results of manual review and the data format of manually corrected archives as new training data, and feed them back into the rule induction model to optimize the model parameters and rule base content; Step S3: Combining the archive format feature database, a deep learning algorithm is used to compare the archive to be tested with the standard format in the format feature database. The archive format is analyzed. If the archive format is qualified, the qualified archive is directly stored in the archive database. If the archive format is not qualified, the unqualified archive is adaptively and intelligently converted. The converted archive is compared with the standard format in the format feature database again. After confirming that the format is qualified, the qualified archive is stored in the archive database.
2. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 1, characterized in that, In step S1, the preprocessing process is as follows: the collected historical archive samples are processed to unify the format, and noise data in the samples is removed to ensure the accuracy of feature extraction and the integrity of the archive format feature database.
3. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 1, characterized in that, Step S24 specifically includes the following steps: Step S241: Standardize the marked error samples and extract the error type and archive metadata; where the error type includes missing fields, format misalignment and encoding error, and the archive metadata includes format type, year and storage medium; Step S242: Use the FP-Growth algorithm to mine erroneous features and correct strong association rules for actions; use the C4.5 decision tree to generate conditional branch rules; specifically, the core indicator calculation formula for association rule mining is as follows: Support is calculated as Support(A→B) = count(A∩B) / count(D); The confidence level is Confidence(A→B) = count(A∩B) / count(A); The lift is calculated as Lift(A→B) = Confidence(A→B) / Support(B). Where A is the set of erroneous features, B is the corrective action, and D is the total number of samples; Support(A→B) is the coverage of rule A→B in the overall sample; count(A∩B) is the number of samples in the intersection of A and B; count(D) is the total number of valid samples participating in the association rule analysis; Confidence(A→B) is the probability that corrective action B is executed when erroneous feature A occurs; count(A) is the number of samples in the total sample set D containing erroneous feature set A; Lift(A→B) is the ratio of the confidence of rule A→B to the support of corrective action B itself; Support(B) is the proportion of samples in the total sample set D containing corrective action B to the total number of samples; where Lift>1 indicates that the rule is valid. Step S243: Extract candidate rules from the hybrid model. The format of the candidate rules is: IF condition set THEN correction action. Use three indicators, support, confidence and lift, to screen effective rules and eliminate redundant or inefficient rules.
4. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 1, characterized in that: Step S25 specifically includes the following steps: Step S251: Insert points in the rule engine to collect the running data of each rule in real time and summarize the indicator data on a regular basis. The transportation data specifically includes the number of triggers Nt, the number of successful executions Ns, and the number of affected files Na. Step S242: Standardize the transportation data and construct a weighted scoring model to calculate the priority score of each rule and set the weight coefficients; Step S243: Update the rule execution order of the rule engine according to priority scores.
5. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 1, characterized in that, In step 23, the rule engine is Drools or a custom graph rule engine.
6. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 1, characterized in that, In step S3, combining the archive format feature database, a deep learning algorithm is used to compare the archive to be detected with the standard formats in the format feature database to analyze the archive format. This specifically includes the following steps: Step S31: Construct a multimodal feature fusion network, including a visual branch and a structural branch, and fuse the features extracted from the visual branch and the structural branch to form a joint feature representation that is sensitive to the file format and file content; Step S32: Compare the joint features of the file to be tested with the standard features in the file format feature database, and determine whether the file format is qualified based on the comparison results; if the file format is qualified, the qualified file is directly stored in the archive database; if the file format is unqualified, a structured fine-grained defect report is output.
7. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 6, characterized in that, In step S3, if the file format is not up to standard, an adaptive intelligent conversion is performed on the unqualified file. The specific steps are as follows: Step S33: Construct a reinforcement learning environment, define the state of the reinforcement learning environment as the fine-grained defect report of the current file and the current format state of the file, define the action space as the conversion operation that the AI agent can execute, and construct a multi-dimensional reward function, and associate the reward value of the multi-dimensional reward function with the repair accuracy, execution efficiency and fidelity of the original content of the conversion operation. Step S34: The AI agent learns and builds an optimal policy network through a reinforcement learning environment, enabling it to predict and output the sequence of transformation actions that can obtain the maximum cumulative reward based on the current defect status of the file, and execute the corresponding transformation operations in sequence. Step S35: Perform a format check on the corrected file again; If the inspection fails, the new defect status is fed back to the AI agent, which then makes decisions and repairs based on the new status until the file format is approved.
8. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 6, characterized in that, In step S31, a multimodal feature fusion network is constructed, including a visual branch and a structural branch, as detailed below: The visual branch uses a Transformer-based document understanding model, taking the document image, text content, and two-dimensional layout information of the archive to be detected as input, and capturing the global format features and semantic relationships of the archive through a self-attention mechanism; the structural branch uses a graph neural network to model archives with a clear hierarchical structure, analyzes the parent-child and sibling relationships between archive elements, and judges the compliance of the archive's logical structure.
9. The method for automatic detection and conversion of historical archive formats based on AI intelligent agents according to claim 6, characterized in that, In step S32, the structured fine-grained defect report specifically includes defect ID, defect type, pixel-level visual marker location, document structure tree path, problem description, and confidence level.