An artificial intelligence-based archive management system

Abstract

The application discloses an archive management system based on artificial intelligence, a data preprocessing module, which is used for obtaining original archive data and preprocessing, a missing area detection module, which is used for detecting missing areas and extracting features of the preprocessed archive data to generate missing area description information, a global search optimization module, which is used for performing global search based on the missing area description information by adopting a variable scale Levy flight algorithm, a local optimization module, which is used for performing detailed optimization on local missing areas by using an improved grey wolf optimization algorithm based on global search candidate data, a data fusion and consistency matching module, which is used for fusing the filled data after local optimization and the preprocessed archive data, and a filling quality evaluation and optimization module, which is used for performing error feedback evaluation on the filled archive data and finally archiving the corrected archive data. The application improves the overall consistency and rationality of archive filling in the archive management process.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of archive management, and in particular to an archive management system based on artificial intelligence. BACKGROUND

[0002] As an important carrier of information storage and historical records, archives play an indispensable role in various enterprises, institutions, government agencies and research institutions. With the passage of time, data loss, content damage and information loss may occur in the archive management process due to physical damage, storage medium aging, improper data migration or external environmental factors, which not only affects the integrity and readability of the archives, but also has a serious impact on historical research, legal evidence preservation and enterprise operation.

[0003] At present, the repair methods for damaged archive data mainly include manual repair, filling method based on statistical model and intelligent repair technology based on deep learning. However, the existing methods still have many deficiencies in practical application.

[0004] Firstly, the manual repair method mainly depends on the experience and subjective judgment of professional archive repair personnel. For text archives, experts can fill in the missing content by manual comparison and speculation, while the repair process of image or multimedia archives is more complex and needs to rely on superb repair skills. However, the existing method not only consumes time and effort, but also is unstable in repair quality due to uncontrollable human factors. In addition, the efficiency of manual repair is extremely low in the face of large-scale archive data, which is difficult to meet the demand of modern archive management for large-scale data repair.

[0005] Secondly, the filling method based on statistical model, such as interpolation method, regression model and Markov model, usually predicts the missing data by calculating the probability distribution of the existing data. The existing method is effective in processing small-scale regularized data, but it is difficult for traditional statistical methods to accurately restore the semantic information and context logic of the data for such complex and nonlinear archive data, which often leads to insufficient accuracy of the filled content. In addition, statistical models are difficult to adapt to different types of archive data, and the scope of application is limited.

[0006] In recent years, with the development of artificial intelligence technology, deep learning methods have been gradually applied to the field of archive restoration. Methods based on convolutional neural networks, generative adversarial networks or variational autoencoders have achieved certain results in filling in image and video archives. However, there are still some limitations. On the one hand, deep learning methods usually require a large amount of high-quality training data, but in practical applications, the acquisition and labeling of archive data is often very difficult. On the other hand, existing deep learning models may produce blurred, distorted or even semantic errors when dealing with archive data with severe missing information, making it difficult to ensure that the restored archive content meets the original archive's logic and style. In addition, some deep learning-based restoration methods have high computational complexity and consume a lot of computing resources, which is not conducive to large-scale archive data restoration applications.

[0007] Therefore, there is an urgent need for an artificial intelligence-based archive management system to improve the accuracy and adaptability of archive data filling and meet the practical needs of modern archive management. SUMMARY

[0008] One object of the present application is to provide an artificial intelligence-based archive management system that improves the overall consistency and rationality of archive filling in the archive management process.

[0009] According to an embodiment of the present application, an artificial intelligence-based archive management system includes the following modules:

[0010] A data preprocessing module is used to obtain original archive data and perform denoising, invalid information removal, format unification and standardization processing on the original archive data. The data preprocessing module analyzes different types of archive data based on archive data storage formats and encoding methods.

[0011] A missing area detection module is used to detect missing areas and extract features from preprocessed archive data, determine and calibrate the boundaries and context features of damaged areas and complete areas in the archive data. The missing area detection module sets a missing area detection threshold by calculating the completeness score of the damaged area, filters archive data with a completeness score below the threshold, and generates missing area description information based on the boundary information, texture features and context semantic information of the archive data.

[0012] A global search optimization module is used to perform global search using a variable-scale Levy flight algorithm based on the missing area description information, and adaptively adjust the search step size during the global search process to identify global search candidate data that matches the missing area.

[0013] a local optimization module configured to perform detailed optimization on the local missing area based on the global search candidate data by using an improved grey wolf optimization algorithm, dynamically adjust optimization parameters in the local optimization process according to the global search result, and realize local data completion and detail recovery;

[0014] a data fusion and consistency matching module configured to fuse the filled data after local optimization and the preprocessed archive data, perform semantic, style and logical consistency matching on the fused archive data, and form preliminary corrected archive data;

[0015] a filling quality evaluation and optimization module configured to perform error feedback evaluation on the filled archive data, calculate the overall matching degree of the filled data, filter data with filling quality higher than a threshold based on the global evaluation result, perform secondary optimization on the filled data according to the global matching evaluation result, so that the filled data and the original archive data achieve best matching in content logic, format layout and semantic consistency, and finally archive the corrected archive data.

[0016] An archive management method based on artificial intelligence, configured to execute an archive management system based on artificial intelligence, comprising the following steps:

[0017] S1. Preprocessing the original archive data to obtain preprocessed archive data;

[0018] S2. Detecting and extracting features of missing areas from the preprocessed archive data, determining and marking the boundaries and context features of damaged areas and complete areas in the archive data to form missing area description information;

[0019] S3. Based on the missing area description information, performing global search on the preprocessed archive data by using a variable scale Levy flight algorithm, and generating global search candidate data matched with the missing area;

[0020] S4. Using an improved grey wolf optimization algorithm to perform detailed optimization on the local missing area, dynamically adjusting optimization parameters in the local optimization process according to the global search result, and generating filled data after local optimization;

[0021] S5. Fusing the filled data after local optimization and the preprocessed archive data, performing semantic, style and logical consistency matching on the fused archive data, and forming preliminary corrected archive data;

[0022] S6. Performing error feedback evaluation on the preliminary corrected archive data, comparing the filling result with the expected features by using an error evaluation module, dynamically adjusting global search and local optimization parameters, and performing secondary optimization on the preliminary corrected archive data to generate final completion data.

[0023] Optionally, the S1 comprises the following steps:

[0024] S11. Obtain the original archival data, identify the format of the original archival data based on the archival data storage format and encoding method, construct the format mapping relationship, and define the archival dataset:

[0025] D raw ={d1,d2,…,d N};

[0026] Among them, D raw Represents the original archive dataset, d i This represents the i-th archive data instance, where N is the total amount of original archive data;

[0027] S12. Denoise the original archive dataset and filter out invalid information from the denoised archive data, removing data items containing low-relevance information to form a valid archive dataset.

[0028] S13. Perform format standardization processing on valid archival datasets, and standardize the format of archival datasets D. formatted Standardization processing is performed, including data normalization and alignment based on the distribution characteristics of the archival content, to generate a preprocessed archival dataset D. processed .

[0029] Optionally, S2 includes the following steps:

[0030] S21. For the preprocessed archive dataset D processed Perform integrity checks and calculate d for each data instance. i Completeness score C(d) i ):

[0031]

[0032] Among them, C(d) i ) represents data instance d i Completeness, |D missing,i | Represents data instance d i The number of missing contents;

[0033] S22. Based on completeness score C(d) i Set a missing region detection threshold, filter data instances with integrity scores below the threshold, and form a missing data subset D. missing ;

[0034] S23. For each data instance in the missing data subset, use structured analysis to segment the damaged regions and construct a damaged region determination function M(r i,j This ultimately forms a set of damaged areas R. missing ;

[0035] For each potential damaged region r in the data instance i,j Calculate the comprehensive damage score Λ i,j :

[0036] Λ i,j = α1·D(r i,j )+ β·S(r i,j )+ γ·C(r i,j );

[0037] wherein D(r i,j ) represents a deviation index for measuring the deviation of the data inside r i,j from the expected value, S(r i,j ) represents an index reflecting the degree of structural discontinuity inside r i,j , C(r i,j ) represents a correlation index for evaluating the contextual consistency of r i,j with its adjacent regions, and α1, β and γ are weight coefficients;

[0038] Define the damaged region determination function as:

[0039]

[0040] wherein δ th is a damage threshold value determined based on historical archive repair data experience;

[0041] S24. For each damaged region r i,j , extract its boundary information to form a boundary information set B missing , and define the boundary extraction function B(r i,j ):

[0042]

[0043] wherein represents a second-order gradient operator, W(r i,j ) is a weight function dynamically adjusted according to the size and shape of r i,j , and b i,j represents the boundary information of the damaged region r i,j ;

[0044] S25. Feature extraction is performed on the context of each damaged region r i,j to form a context feature set F context , and a context feature vector F context,i,j is constructed:

[0045] F context,i,j = [ξ1·E(r i,j ), ξ2·T(r i,j ), ξ3·L(r i,j )],

[0046] wherein E(r i,j ) represents an edge continuity measure, reflecting the edge extensibility around the damaged region r i,j , T(r i,j ) represents a texture similarity measure, reflecting the texture consistency between the damaged region r i,j and the adjacent region, and L(r i,j ) represents a layout consistency index, and ξ1, ξ2 and ξ3 are weight coefficients.

[0047] S26. Fusing the damaged region R missing , the boundary information set B missing and the context feature vector F context to construct the missing region description information D desc :

[0048] D desc,i,j = (r i,j , b i,j , F context,i,j , P(r i,j )) ;

[0049] wherein P(r i,j ) is a recovery priority prediction function:

[0050]

[0051] wherein f k is the kth feature component in F context,i,j , M represents the dimension of the context feature vector, and λ is a balance coefficient.

[0052] Finally, the missing region description information set is formed:

[0053]

[0054] Optionally, the S3 comprises the following steps:

[0055] S31. Extracting the damaged region R desc and the context feature set F missing from the missing region description information set D context , and constructing a global search space S search according to the spatial distribution of the missing region:

[0056]

[0057] wherein S search represents the search range of the damaged region of the archive data in the two-dimensional space, and (x i,j , y i,j ) is the damaged region r i,jThe coordinate point set in the search space;

[0058] S32. In the global search space S search , the variable scale Levy flight algorithm is used to perform global search, and the entropy value of the damaged area is calculated according to the context feature set F context to dynamically adjust the search step size S t :

[0059]

[0060] Where S t represents the step size of the tth iteration in the search process, S max is the initial maximum step size, H(F context,i,j ) represents the information entropy of the damaged area r i,j in the context feature set, which is used to measure the complexity of the area. The higher the information entropy value, the more serious the information loss of the area, and λ1 is a control factor.

[0061] S33. Calculate the search jump step size L α according to the variable scale Levy flight model, and update the next search point (x t+1 , y t+1 ) according to the position of the current search point:

[0062] (x t+1 , y t+1 ) = (x t , y t ) + S t · L α · (cosθ, sinθ);

[0063] Where cosθ and sinθ represent the horizontal component and vertical component of the jump direction respectively, L α obeys the Levy distribution:

[0064]

[0065] Where Γ(α) is the gamma function, which is used to normalize the calculation of the jump step size, θ is a uniform random angle, which represents the direction of the search jump, and s is a random variable, which controls the change of the jump step size.

[0066] S34. Calculate the matching degree S(x t+1 , y t+1 ) between the updated search point (x t+1 , y t+1 ) and the corresponding damaged area boundary, and judge whether to enter the local optimization mode:

[0067]

[0068] Where Bmissing b is the boundary set of the damage region k is the boundary point coordinate, γ1 is the matching degree control factor, which determines the weight of the influence of matching degree on search points;

[0069] The matching degree reflects whether the search point enters a high correlation region. When the matching degree S(x t+1 ,y t+1 ) > τ, the local optimization mode is started, and the step size is dynamically adjusted:

[0070] S local,t+1 = S max · (1-S(x t+1 ,y t+1 ));

[0071] Wherein, τ is the local optimization trigger threshold, S local,t+1 is the local optimization step size, which tends to 0 when the matching degree is high;

[0072] S35. In the local optimization mode, a direction weight vector W t is introduced to optimize the search direction, and the search direction adjustment parameter is calculated:

[0073]

[0074] Wherein, w x and w y are the matching degrees in x direction and y direction respectively;

[0075] The updated formula of the search point after optimization:

[0076] (x t+2 ,y t+2 ) = (x t+1 ,y t+1 ) + S local,t+1 · L α · W t ;

[0077] Wherein, L α is adjusted to the range of 1.5 < α < 2 in the local search mode;

[0078] S36. In the local optimization process, the local optimal point is stored. If the optimal value is not updated for K jump consecutive search rounds, a small range jump is performed to prevent local convergence;

[0079] S37. Based on the spatial distribution of the search point and its matching degree with the damage region, the target fitness function F(x t ,y t ) is calculated, which is used to screen the optimal search path:

[0080]

[0081] where S(x t+2 ,y t+2 ) is the matching degree of the optimized search point and the boundary of the damaged area, G(x t+2 ,y t+2 ) is the coverage degree of the optimized search point inside the damaged area, the normalized convergence rate, w1, w2, w3 are weight parameters for adjusting the influence factors;

[0082] S38. According to the convergence of the search trajectory, the optimal search path is screened, and a global search candidate data set D global is constructed to match the missing area:

[0083]

[0084] Optionally, the S4 comprises the following steps:

[0085] S41. Obtain a global search candidate data set D global , and construct an adaptive multi-layer local optimization population P multi according to the complexity of different damaged areas

[0086]

[0087] where P multi represents a multi-layer optimization population, and all individuals of the optimization layers constitute a complete search space, P m represents the mth optimization population, and the optimization population is divided according to the complexity of the damaged area, represents the kth search individual in the mth optimization population, i.e., a local optimization candidate point, and M is the number of optimization layers;

[0088] Calculate the complexity C damage of the damaged area, and when the complexity C damage of the damaged area is higher, it indicates that the structural complexity of the damaged area is higher, and the grey wolf optimization algorithm increases the number of optimization layers M:

[0089]

[0090] where σ G (r i,j ) represents the gradient standard deviation inside the damaged area r i,j , which measures the edge complexity inside the area, is the texture mutation degree calculated by the Laplace operator, which describes the structural discontinuity of the damaged area;

[0091] S42. In the optimization process, different layers of the optimization population adopt differentiated fitness functions to adapt to different damage characteristics of the archive data, and a multi-layer adaptive fitness function is defined:

[0092]

[0093] wherein, is the matching degree of the local search point and the boundary of the damaged area, so that the filling point is consistent with the morphological characteristics of the damaged area, is the context consistency, which measures whether the filling data conforms to the logic and content structure of the original archive data, is the texture similarity, which measures the similarity of the texture characteristics of the filling point and the adjacent recovered data:

[0094]

[0095] wherein, G(x,y) is the local gradient feature, N k is the filled data point in the neighborhood range of the search point;

[0096] is the local data smoothness, which makes the continuity of the filling data in the local area:

[0097]

[0098] wherein, I(x,y) represents the pixel value or text feature value of the archive data;

[0099] is the information entropy increment, which measures the contribution of the filling point to the overall information integrity:

[0100]

[0101] wherein, H before and H after are the entropy values of the damaged area before and after filling, respectively;

[0102] S43. In the archive filling task, the search direction is guided by the boundary information of the archive data, and the texture gradient guidance is introduced on the basis of the traditional grey wolf optimization search to construct the grey wolf optimization search based on texture guidance:

[0103]

[0104] wherein, is the search position of the kth grey wolf in the mth optimization population at the t+1th iteration, which represents the position change of the filling candidate point in the optimization process, respectively represent the positions of the optimal solution, the second optimal solution and the third optimal solution in the optimization population, A ma coefficient for controlling the search step length and convergence speed of the grey wolf individual, respectively represent the distance between the grey wolf individual and the optimal solution, the suboptimal solution and the third optimal solution position, B m is a gradient guiding term:

[0105]

[0106] wherein, represents the local gradient direction, guiding the filling point to converge to the optimal texture matching direction, and γ2 is a gradient guiding weight;

[0107] S44. In the archive data repair task, the hybrid step length strategy is used to optimize the search of the grey wolf individual, and the search step length is dynamically adjusted to balance the precision and convergence speed:

[0108]

[0109] wherein, 2·(1-t / T max ) controls the global convergence trend, λ m ·N(0,1) enhances the search diversity through hierarchical disturbance;

[0110] S45. A local convergence detection mechanism is introduced. If the fitness of a search point is not improved for K stuck consecutive iterations, a jump is performed:

[0111]

[0112] S46. After the convergence of the grey wolf optimization algorithm, the optimal filling point is screened, and the filling data after local optimization is finally formed:

[0113]

[0114] wherein, is the coordinate of the k+2 search individual in the mth layer of the optimized population, F m (p k+2 ) is the multi-level adaptive fitness function after optimization.

[0115] Optionally, the S5 comprises the following steps:

[0116] S51. The filling data after local optimization and the preprocessed archive data are obtained, the filling data is normalized according to the data format, storage structure and encoding mode, the normalized filling data and the preprocessed archive data are preliminarily fused, and an initial data fusion set is constructed;

[0117] S52. Analyze the local features of the area where the filling data is located, extract the local feature vector of the damaged area, and calculate the local similarity of the filling data and the adjacent area, including the boundary shape similarity, local structure consistency and content feature matching degree, and judge whether the filling data meets the local style characteristics of the original file according to the local similarity calculation result, and adjust the filling data when the local similarity is lower than the preset threshold value;

[0118] S53. Based on the mode characteristics of the historical file data, the style, semantic and logical consistency of the fused file data is matched, the style adaptability of the filling data is analyzed according to the language characteristics, layout style and data structure rule of the file content, and the filling data with a matching degree lower than the threshold value is corrected through a style adjustment function, so that the filling data is consistent with the original file in font format, text semantics, structure logic and context coherence;

[0119] S54. After completing the local similarity adjustment and style matching, the global evaluation of the overall filling result is carried out, the matching degree of the filling data and the entire file content is calculated, and the data with a filling quality higher than the threshold value is screened according to the global matching result, and the filling data with a low matching degree is adjusted again, so that the filling data and the entire file are best matched in content logic, format layout and semantic consistency;

[0120] S55. After completing the local and global consistency matching, the final screened filling data is archived to generate the preliminary corrected file data.

[0121] The beneficial effects of the present application are:

[0122] (1) The present application performs global search through variable scale Levy flight algorithm, and adaptively adjusts the search step length in the search process to enable the search to jump in a wide range, while performing fine-grained search in a high matching area. The variable scale Levy flight can combine large step global exploration and small step local refinement, improve search efficiency and avoid falling into local optimum, and can search for optimal filling data in a larger range, improve the overall consistency and rationality of the filling data.

[0123] (2) The present application introduces an improved grey wolf optimization algorithm to optimize and adjust the local area. The grey wolf optimization algorithm dynamically adjusts the boundary features, texture consistency and context logic of the filling data by simulating the hunting behavior of a wolf pack, ensures that the filling area and the original data are consistent in semantics, style and structure, and enables the filling data to be more naturally integrated with the surrounding area through guided search of the grey wolf optimization algorithm.

[0124] (3) After the filling is completed, an error feedback mechanism is used to automatically evaluate the repair result, and the global search and local optimization parameters are dynamically adjusted based on the feedback result, so that adaptive optimization is realized, the matching degree of the filled data and the original data is calculated through the error evaluation module, including semantic consistency, style adaptability and structure rationality, and the search step and optimization parameters are adjusted according to the feedback result, so that the filled data continuously approaches the optimal solution. BRIEF DESCRIPTION OF DRAWINGS

[0125] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application, and do not limit the application. In the drawings:

[0126] Figure 1 A flowchart of an artificial intelligence-based archive management system is provided. DETAILED DESCRIPTION

[0127] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, which only schematically show the basic structure of the application, and therefore only show the components relevant to the application.

[0128] REFERENCE Figure 1 An artificial intelligence-based archive management system comprises the following modules:

[0129] A data preprocessing module is used to obtain original archive data and perform denoising, invalid information removal, format unification and standardization processing on the original archive data. The data preprocessing module analyzes different types of archive data based on the archive data storage format and coding method;

[0130] A missing area detection module is used to detect missing areas and extract features from the preprocessed archive data, determine and calibrate the boundaries and context features of damaged areas and complete areas in the archive data. The missing area detection module sets a missing area detection threshold by calculating the completeness score of the damaged area, filters the archive data with a completeness score below the threshold, and generates missing area description information based on the boundary information, texture features and context semantic information of the archive data;

[0131] A global search optimization module is used to perform global search based on the missing area description information using a variable-scale Levy flight algorithm, and adaptively adjust the search step during the global search process to identify global search candidate data matching the missing area;

[0132] A local optimization module is used to perform detailed optimization of local missing areas based on the global search candidate data using an improved grey wolf optimization algorithm, and dynamically adjust the optimization parameters according to the global search result during the local optimization process to realize local data completion and detail recovery.

[0133] a data fusion and consistency matching module for fusing the locally optimized filling data with the preprocessed archive data, and performing semantic, style and logical consistency matching on the fused archive data;

[0134] a filling quality evaluation and optimization module for performing error feedback evaluation on the filled archive data, calculating the overall matching degree of the filling data, and based on the global evaluation result, screening data with filling quality higher than a threshold, and performing secondary optimization on the filling data according to the global matching evaluation result, so that the filling data and the original archive achieve best matching in content logic, format layout and semantic consistency, and finally archive the corrected archive data.

[0135] An archive management method based on artificial intelligence, for executing an archive management system based on artificial intelligence, comprising the following steps:

[0136] S1. Preprocessing the original archive data to obtain preprocessed archive data;

[0137] S2. Detecting missing areas and extracting features from the preprocessed archive data, determining and marking the boundaries and context features of damaged areas and complete areas in the archive data, and forming missing area description information;

[0138] S3. Based on the missing area description information, using a variable scale Levy flight algorithm to perform global search on the preprocessed archive data, and generating global search candidate data matching the missing area;

[0139] S4. Using an improved grey wolf optimization algorithm to perform detailed optimization on the local missing area, and dynamically adjusting the optimization parameters during the local optimization process according to the global search result, to generate locally optimized filling data;

[0140] S5. Fusing the locally optimized filling data with the preprocessed archive data, and performing semantic, style and logical consistency matching on the fused archive data to form preliminary corrected archive data;

[0141] S6. Error feedback evaluation is performed on the preliminary corrected archive data, the filling result is compared with the expected features through the error evaluation module, the global search and local optimization parameters are dynamically adjusted, and the preliminary corrected archive data is optimized again to generate the final filling data.

[0142] In this embodiment, S1 includes the following steps:

[0143] S11. Obtain the original archive data, identify the format of the original archive data based on the archive data storage format and encoding method, construct the format mapping relationship, and define the archive data set:

[0144] D raw = {d1, d2, …, di, …, dN}; N};

[0145] wherein D raw represents the original archive data set, di i represents the i-th archive data instance, and N is the total amount of original archive data;

[0146] S12. The original archive data set is denoised, and the denoised archive data is filtered for invalid information, and data items containing low correlation information are removed to form an effective archive data set;

[0147] S13. The effective archive data set is subjected to format unification processing, and the format-unified archive data set D formatted is subjected to standardization processing, and according to the distribution characteristics of the archive content, the data is normalized and aligned to generate a preprocessed archive data set D processed .

[0148] In this embodiment, S2 includes the following steps:

[0149] S21. The preprocessed archive data set D processed is subjected to integrity detection, and the completeness score C(di i ) of each data instance di i is calculated:

[0150]

[0151] wherein C(di i ) represents the completeness of the data instance di i , and |D missing,i | represents the number of missing contents in the data instance di i ;

[0152] S22. The missing region detection threshold is set based on the completeness score C(di i ), and the data instances with a completeness score lower than the threshold are screened to form a missing data subset D missing ;

[0153] S23. A structured analysis method is used to segment each data instance in the missing data subset, a damaged region judgment function M(r i,j ) is constructed, and finally a damaged region set R missing is formed;

[0154] The comprehensive damage score Λ i,j is calculated for each potential damaged region r i,j in the data instance:

[0155] Λi,j = a1D(r i,j ) + βS(r i,j ) + γC(r i,j ) ;

[0156] wherein D(r i,j ) represents a deviation index for measuring the deviation of r i,j from the expected value, S(r i,j ) represents an index reflecting the degree of structural discontinuity of r i,j , C(r i,j ) represents a correlation index for evaluating the consistency of r i,j with the context of its adjacent regions, a1, β and γ are weight coefficients;

[0157] Define the damaged area judgment function as:

[0158]

[0159] wherein δ th is a damage threshold value determined based on historical archive repair data experience;

[0160] S24. Extract the boundary information of each damaged area r i,j to form a boundary information set B missing , and define the boundary extraction function B(r i,j ):

[0161]

[0162] wherein represents a second-order gradient operator, W(r i,j ) is a weight function dynamically adjusted according to the size and shape of r i,j , and b i,j represents the boundary information of the damaged area r i,j ;

[0163] S25. Feature extraction is performed on the context of each damaged area r i,j to form a context feature set F context , and a context feature vector F context,i,j is constructed:

[0164] F context,i,j = [ξ1E(r i,j ), ξ2T(r i,j ), ξ3L(r i,j )],

[0165] wherein E(r i,j ) represents an edge continuity measure, reflecting the damaged area r i,jSurrounding edge extensibility, T(r) i,j ) represents a texture similarity measure, reflecting the damaged area r i,j Texture consistency with neighboring regions, L(r) i,j ) represents the layout consistency index, where ξ1, ξ2, and ξ3 are weighting coefficients;

[0166] S26. Damaged area R missing Boundary information set B missing and context feature vector F context Merge the data to construct the missing region description information D. desc :

[0167] D desc,i,j =(r i,j ,b i,j ,F context,i,j ,P(r i,j ));

[0168] Wherein, P(r) i,j For the recovery priority prediction function:

[0169]

[0170] Among them, f k For F context,i,j The k-th feature component in the context feature vector, where M represents the dimension of the context feature vector and λ is the balance coefficient;

[0171] The final set of missing region description information is formed:

[0172]

[0173] In this embodiment, S3 includes the following steps:

[0174] S31. Describe information set D from missing regions desc Extract damaged area R missing and context feature set F context A global search space S is constructed based on the spatial distribution of the missing regions. search :

[0175]

[0176] Among them, S search This represents the search range in two-dimensional space for the damaged area of ​​the archive data, (x i,j ,y i,j ) represents the damaged area r i,j The set of coordinate points in the search space;

[0177] S32. In the global search space S searchIn the present application, the variable scale Levy flight algorithm is used to perform global search, and the information entropy of the context feature set F context The entropy value of the damaged area is calculated, and the search step S is dynamically adjusted t :

[0178]

[0179] Where S t represents the step size of the tth iteration in the search process, S max is the initial maximum step size, H(F context,i,j ) represents the information entropy of the damaged area r i,j in the context feature set, which is used to measure the complexity of the area. The higher the information entropy value, the more serious the information loss of the area, and λ1 is the control factor.

[0180] S33. Calculate the search jump step L according to the variable scale Levy flight model α , and update the next search point (x t+1 , y t+1 ) according to the position of the current search point:

[0181] (x t+1 , y t+1 ) = (x t , y t ) + S t · L α · (cosθ, sinθ);

[0182] Where cosθ and sinθ represent the horizontal component and vertical component of the jump direction respectively, L α obeys the Levy distribution:

[0183]

[0184] Where Γ(α) is the gamma function, which is used to normalize the calculation of the jump step, θ is a uniform random angle, which represents the direction of the search jump, and s is a random variable, which controls the change of the jump step.

[0185] S34. Calculate the matching degree S(x t+1 , y t+1 ) between the updated search point (x t+1 , y t+1 ) and the corresponding damaged area boundary, and judge whether to enter the local optimization mode:

[0186]

[0187] Where B missing is the boundary set of the damaged area, and b kis the coordinate of the boundary point, γ1is the matching degree control factor, which determines the weight of the influence of matching degree on the search point;

[0188] The matching degree reflects whether the search point enters the high correlation area. When the matching degree S(x t+1 ,y t+1 ) of the boundary of the damaged area is greater than τ, the local optimization mode is started, and the step size is dynamically adjusted:

[0189] S local,t+1 = S max ·(1-S(x t+1 ,y t+1 ));

[0190] Wherein, τ is the local optimization trigger threshold, S local,t+1 is the local optimization step size, which tends to 0 when the matching degree is high;

[0191] S35. In the local optimization mode, a direction weight vector W t is introduced to optimize the search direction, and the search direction adjustment parameter is calculated:

[0192]

[0193] Wherein, w x and w y are the matching degrees in x direction and y direction respectively;

[0194] The updated formula of the search point after optimization:

[0195] (x t+2 ,y t+2 ) = (x t+1 ,y t+1 ) + S local,t+1 ·L α ·W t ;

[0196] Wherein, L α is adjusted to the range of 1.5 < α < 2 in the local search mode;

[0197] S36. In the local optimization process, the local optimal point is stored. If the optimal value is not updated for K jump consecutive search rounds, a small range jump is performed to prevent local convergence;

[0198] S37. Based on the spatial distribution of the search point and its matching degree with the damaged area, the target fitness function F(x t ,y t ) is calculated, which is used to select the optimal search path:

[0199]

[0200] where S(x t+2 ,y t+2 ) is the matching degree of the optimized search point and the boundary of the damaged area, G(x t+2 ,y t+2 ) is the coverage degree of the optimized search point inside the damaged area, the normalized convergence rate, w1, w2, and w3 are weight parameters for adjusting the influence factors;

[0201] S38. According to the convergence of the search trajectory, the optimal search path is screened, and a global search candidate data set D global is constructed to match the missing area.

[0202]

[0203] In this embodiment, S4 includes the following steps:

[0204] S41. The global search candidate data set D global is obtained, and an adaptive multi-layer local optimization population P multi is constructed according to the complexity of different damaged areas. Different levels of optimization populations undertake different optimization tasks in the search process.

[0205]

[0206] where P multi represents the multi-layer optimization population, and the individuals of all optimization layers constitute a complete search space. P m represents the mth optimization population, and the optimization population is divided according to the complexity of the damaged area, represents the kth search individual in the mth optimization population, i.e., the local optimization candidate point, and M is the number of optimization layers.

[0207] The complexity C damage of the damaged area is calculated. When the complexity C damage of the damaged area is higher, it indicates that the structural complexity of the damaged area is higher, and the grey wolf optimization algorithm increases the number of optimization layers M:

[0208]

[0209] where σ G (r i,j ) represents the gradient standard deviation inside the damaged area r i,j , which measures the edge complexity inside the area, is the degree of texture mutation calculated by the Laplace operator, which describes the structural discontinuity of the damaged area.

[0210] S42. In the optimization process, different layers of the optimization population adopt differentiated fitness functions to adapt to different damage characteristics of the archive data, and a multi-layer adaptive fitness function is defined:

[0211]

[0212] wherein, is the matching degree of the local search point and the boundary of the damaged area, so that the filling point is consistent with the morphological characteristics of the damaged area, is the context consistency, which measures whether the filling data conforms to the logic and content structure of the original archive data, is the texture similarity, which measures the similarity of the texture characteristics of the filling point and the adjacent recovered data:

[0213]

[0214] wherein, G(x,y) is the local gradient feature, N k is the filled data point in the neighborhood range of the search point;

[0215] is the local data smoothness, which makes the continuity of the filling data in the local area:

[0216]

[0217] wherein, I(x,y) represents the pixel value or text feature value of the archive data;

[0218] is the information entropy increment, which measures the contribution of the filling point to the overall information integrity:

[0219]

[0220] wherein, H before and H after are the entropy values of the damaged area before and after filling, respectively;

[0221] S43. In the archive filling task, the search direction is guided by the boundary information of the archive data, and the texture gradient guidance is introduced on the basis of the traditional grey wolf optimization search to construct a grey wolf optimization search based on texture guidance:

[0222]

[0223] wherein, is the search position of the kth grey wolf in the mth optimization population at the t+1th iteration, which represents the position change of the filling candidate point in the optimization process, respectively represent the positions of the optimal solution, the second optimal solution and the third optimal solution in the optimization population, A ma coefficient for controlling the search step length and convergence speed of the grey wolf individual, respectively represent the distance between the grey wolf individual and the optimal solution, the suboptimal solution and the third optimal solution position, B m is a gradient guiding term:

[0224]

[0225] wherein, represents the local gradient direction, guiding the filling point to converge to the optimal texture matching direction, and γ2 is the gradient guiding weight;

[0226] S44. The hybrid step length strategy is used to optimize the search of the grey wolf individual in the archive data repair task, and the search step length is dynamically adjusted to balance the precision and convergence speed:

[0227]

[0228] wherein, 2·(1-t / T max ) controls the global convergence trend, λ m ·N(0,1) enhances the search diversity through hierarchical disturbance;

[0229] S45. A local convergence detection mechanism is introduced. If the fitness of a search point is not improved for K stuck consecutive iterations, a jump is performed:

[0230]

[0231] S46. After the convergence of the grey wolf optimization algorithm, the optimal filling point is screened, and the filling data after local optimization is finally formed:

[0232]

[0233] wherein, is the coordinate of the k+2 search individual in the mth layer of the optimized population, F m (p k+2 ) is the multi-level adaptive fitness function after optimization.

[0234] In the embodiment, S5 includes the following steps:

[0235] S51. The filling data after local optimization and the preprocessed archive data are obtained, the filling data is normalized according to the data format, storage structure and encoding method, the normalized filling data and the preprocessed archive data are preliminarily fused, and an initial data fusion set is constructed;

[0236] S52. Analyze the local features of the area where the filling data is located, extract the local feature vector of the damaged area, and calculate the local similarity between the filling data and the adjacent area, including boundary shape similarity, local structure consistency and content feature matching degree, and determine whether the filling data meets the local style characteristics of the original file according to the local similarity calculation result. When the local similarity is lower than the preset threshold, adjust the filling data;

[0237] S53. Based on the pattern characteristics of historical archive data, the style, semantic and logical consistency of the fused archive data is matched, and the style adaptability of the filling data is analyzed according to the language characteristics, layout style and data structure rule of the archive content. For filling data with a matching degree lower than the threshold, the filling data is modified by a style adjustment function to make the filling data consistent with the original file in font format, text semantics, structure logic and context coherence;

[0238] S54. After completing the local similarity adjustment and style matching, the overall filling result is globally evaluated, the matching degree of the filling data with the entire archive content is calculated, and the filling data with a quality higher than the threshold is selected according to the global matching result, and the filling data with a low matching degree is adjusted again to make the filling data and the entire archive in content logic, format layout and semantic consistency achieve the best matching;

[0239] S55. After completing the local and global consistency matching, the final selected filling data is archived to generate the preliminary corrected archive data.

[0240] Example 1:

[0241] In July 2023, in the digital archive management project of an archive in East China, the staff found that some electronic archive files of historical documents were abnormal when checking the digital storage of the collected archives. When checking the database, the technical personnel found that the content of the archive file numbered AG1952-001 had serious missing, and some text characters were damaged due to aging of the storage medium. The completeness score of the file was lower than the set safety threshold. Further inspection found that about 18% of the content was missing in the lower right corner of the second and fifth pages of the scanned image of the archive, and some content of the document could not be automatically recognized due to blurring and noise interference.

[0242] In order to repair the archive file, the technical team decided to use the archive damaged content filling method based on variable scale Levy flight and gray wolf optimization to repair it, and compared the repair results with traditional methods to verify the effectiveness of the method.

[0243] The technician first inputs the scanned image of file No. AG1952-001 into the system, which automatically detects the damaged areas of the document and calculates a completeness score of 0.82 (below the threshold of 0.9). It is found that the bottom right corner of page 2 has 5 lines of text completely missing, and the right side of page 5 has 2 paragraphs that cannot be read due to blurring. The system automatically performs denoising, contrast enhancement, and binarization on the file image to improve the clarity of the text.

[0244] After preprocessing is complete, the system begins to detect missing areas, calculates a text damage score of 0.24 (the lower the value, the more severe the damage), and extracts the context features of the damaged areas, recording the coordinate information of the damaged areas to prepare for subsequent filling.

[0245] After entering the filling stage, the system calls the variable-scale Levy flight algorithm to perform global search. The system analyzes historical archives from 1951 to 1955, calculates the most similar text fragments to the missing part, and in the search process, the system adaptively adjusts the search step size to find content fragments with a similarity greater than 85% in the global range. The system filters out three possible matching candidates:

[0246] 1. File No. AG1952-002: document content similarity 86.3%, but there is a deviation in the layout structure;

[0247] 2. File No. AG1951-007: document content similarity 82.1%, but the language style is older;

[0248] 3. File No. AG1953-004: document content similarity 91.5%, layout and semantics are well matched.

[0249] The system finally selects the matching fragment of AG1953-004 as the basis for filling data and fills it into the missing part, and passes the filled data to the local optimization module.

[0250] After filling is complete, the system uses the grey wolf optimization algorithm for local adjustment. The system analyzes the boundary consistency score of the filled text, which is 0.76 (the target value is 0.85 or higher), indicating that the boundary transition is not natural enough, so it enters the optimization process:

[0251] 1. Adjust font alignment: based on the comparison between the filled text and the original text, adjust the line spacing and word spacing to improve the format matching degree of the filled part by 12.7%;

[0252] 2. Semantic logic optimization: use context consistency analysis to compare the semantics of the filled content with the preceding and following text, eliminate 1 grammatical error, and correct 2 phenomena of mismatched subject-predicate relationship;

[0253] 3. Text style adjustment: adjust the wording of the filled part to make it more consistent with the 1950s official language, ensuring that the repaired content is consistent in style;

[0254] Finally, the boundary consistency score of the locally optimized filled text is improved to 0.89, and the semantic matching degree reaches 96.4%, and the repair is completed.

[0255] In order to verify the repair effect of the method of the present application, the technical team simultaneously uses the traditional OCR+rule matching filling method for repair, and compares the repair results of the two methods, and the experimental data are as follows:

[0256]

[0257] The test results show that the method of the present application has the characteristics of higher filling accuracy, more consistent semantics and logic, more natural boundary transition, and higher repair efficiency in the repair of all types of archives, especially in the repair of ancient literature, the effect is most obvious.

[0258] After the method of the present application in this experiment, it is formally applied to the digital repair project of historical archives in the archives, and from September to December 2023, 3127 historical archives are repaired, and more than 215,000 pages of archives are repaired in total, greatly improving the efficiency and quality of archive repair, and the repaired archives are re-stored in the collection database for researchers and the public to consult.

[0259] This embodiment demonstrates in detail the application of the archive damage content filling method based on variable scale Levy flight and grey wolf optimization in real scene through the actual archive repair case of an archive in East China, and the experimental results show that the method of the present application is significantly better than the traditional method in filling accuracy, semantic consistency and repair efficiency, which can effectively improve the quality and automation degree of archive repair, and has important practical application value.

[0260] The present application performs global search through variable scale Levy flight algorithm, and adaptively adjusts the search step length in the search process to enable the search to jump in a wide range, while performing fine-grained search in high matching areas. Variable scale Levy flight can combine large step global exploration and small step local refinement to improve search efficiency and avoid falling into local optimum, while searching for optimal filling data in a larger range to improve the overall consistency and rationality of the filling data.

[0261] The present application introduces an improved grey wolf optimization algorithm to optimize and adjust the local area. The grey wolf optimization algorithm dynamically adjusts the boundary features, texture consistency and context logic of the filling data by simulating the hunting behavior of a wolf pack, ensuring that the semantic, style and structure of the filling area are consistent with the original data. Through guided search by the grey wolf optimization algorithm, the filling data can be more naturally integrated with the surrounding area.

[0262] After the filling is completed, the error feedback mechanism is used to automatically evaluate the repair result, and the global search and local optimization parameters are dynamically adjusted based on the feedback result, so as to realize adaptive optimization, the matching degree of the filling data and the original data is calculated through the error evaluation module, including semantic consistency, style adaptability and structure rationality, and the search step and the optimization parameter are adjusted according to the feedback result, so that the filling data continuously approaches the optimal solution.

[0263] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can make equivalent replacement or change within the technical range disclosed by the present application according to the technical scheme and the inventive concept of the present application, which should be covered within the protection scope of the present application.

Claims

1. An artificial intelligence-based archive management system, characterized in that, Includes the following modules: The data preprocessing module is used to acquire raw archival data and perform noise reduction, removal of invalid information, format unification and standardization on the raw archival data. The data preprocessing module parses different types of archival data based on the archival data storage format and encoding method. The missing region detection module is used to detect and extract features from the preprocessed archival data, determine and mark the boundaries and contextual features of damaged and complete regions in the archival data. The missing region detection module calculates the integrity score of the damaged region, sets the missing region detection threshold, filters archival data with integrity scores below the threshold, and generates missing region description information based on the boundary information, texture features and contextual semantic information of the archival data. The global search optimization module is used to perform a global search based on the missing region description information using the variable-scale Lévy flight algorithm. During the global search process, the search step size is adaptively adjusted to identify global search candidate data that matches the missing region. The local optimization module is used to perform detailed optimization of local missing areas based on global search candidate data and an improved gray wolf optimization algorithm. During the local optimization process, the optimization parameters are dynamically adjusted according to the global search results to achieve local data completion and detail restoration. The data fusion and consistency matching module is used to fuse the locally optimized fill data with the preprocessed archival data, and to perform semantic, style and logical consistency matching on the fused archival data; The filling quality assessment and optimization module is used to evaluate the error feedback of the filled archival data, calculate the overall matching degree of the filled data, and filter the data with filling quality higher than the threshold based on the global assessment results. Based on the global matching assessment results, the filled data is further optimized to achieve the best matching with the original archive in terms of content logic, format layout and semantic consistency, and finally the corrected archival data is archived.

2. An artificial intelligence-based archive management method, used to execute the artificial intelligence-based archive management system described in claim 1, characterized in that, Includes the following steps: S1. Preprocess the original archival data to obtain preprocessed archival data; S2. Perform missing region detection and feature extraction on the preprocessed archival data, determine and mark the boundaries and contextual features of damaged and complete regions in the archival data, and form missing region description information; S3. Based on the missing region description information, the variable-scale Levy flight algorithm is used to perform a global search on the preprocessed archive data and generate global search candidate data that matches the missing regions; S4. The improved Grey Wolf optimization algorithm is used to perform detailed optimization on the local missing areas. During the local optimization process, the optimization parameters are dynamically adjusted based on the global search results to generate the locally optimized filling data. S5. The locally optimized fill data is merged with the preprocessed archival data, and the merged archival data is matched for semantic, stylistic and logical consistency to form the initially corrected archival data; S6. Implement error feedback evaluation on the initially revised archival data. Compare the filling results with the expected features through the error evaluation module, dynamically adjust the global search and local optimization parameters, and further optimize the initially revised archival data to generate the final completed data.

3. The artificial intelligence-based record management method according to claim 2, characterized in that, S1 includes the following steps: S11. Obtain the original archival data, identify the format of the original archival data based on the archival data storage format and encoding method, construct the format mapping relationship, and define the archival dataset: D raw ={d1,d2,…,d N }; Among them, D raw Represents the original archive dataset, d i This represents the i-th archive data instance, where N is the total amount of original archive data; S12. Denoise the original archive dataset and filter out invalid information from the denoised archive data, removing data items containing low-relevance information to form a valid archive dataset. S13. Perform format standardization processing on valid archival datasets, and standardize the format of archival datasets D. formatted Standardization processing is performed, including data normalization and alignment based on the distribution characteristics of the archival content, to generate a preprocessed archival dataset D. processed .

4. The artificial intelligence-based record management method according to claim 2, characterized in that, S2 includes the following steps: S21. For the preprocessed archive dataset D processed Perform integrity checks and calculate d for each data instance. i Completeness score C(d) i ): Among them, C(d) i ) represents data instance d i Completeness, |D missing,i | Represents data instance d i The number of missing contents; S22. Based on completeness score C(d) i Set a missing region detection threshold, filter data instances with integrity scores below the threshold, and form a missing data subset D. missing ; S23. For each data instance in the missing data subset, use structured analysis to segment the damaged regions and construct a damaged region determination function M(r i,j This ultimately forms a set of damaged areas R. missing ; For each potentially damaged area r in the data instance i,j Calculate the overall damage score Λ i,j : L i,j =α1·D(r i,j )+β·S(r i,j )+γ·C(r i,j ); Wherein, D(r) i,j ) indicates that it is used to measure r i,j The deviation indicator of internal data from expected values, S(r) i,j ) indicates reflecting r i,j An index of the degree of discontinuity in internal structure, C(r) i,j ) indicates that it is used to evaluate r i,j The relevance index of its contextual consistency with its neighboring regions, with α1, β and γ as weighting coefficients; The function for determining the damaged area is defined as follows: Where, δ th The damage threshold is determined based on experience with historical archive restoration data; S24. For each damaged area r i,j Extract its boundary information to form a boundary information set B. missing Define the boundary extraction function B(r) i,j ): in, Let W(r) represent the second-order gradient operator. i,j Based on r i,j The weighting function for dynamically adjusting the size and shape, b i,j Indicates the damaged area r i,j Boundary information; S25. For each damaged area r i,j Feature extraction is performed on the context to form a context feature set F. context Construct context feature vector F context,i,j : F context,i,j =[ξ1·E(r i,j ),ξ2·T(r i,j ),ξ3·L(r i,j )], Wherein, E(r) i,j () represents the edge continuity measure, reflecting the damaged area r i,j Surrounding edge extensibility, T(r) i,j ) represents a texture similarity measure, reflecting the damaged area r i,j Texture consistency with neighboring regions, L(r) i,j ) represents the layout consistency index, where ξ1, ξ2, and ξ3 are weighting coefficients; S26. Damaged area R missing Boundary information set B missing and context feature vector F context To integrate.

5. The artificial intelligence-based record management method according to claim 4, characterized in that, The missing region description information D desc Specifically, it includes: D desc,i,j =(r i,j ,b i,j ,F context,i,j ,P(r i,j )); Wherein, P(r) i,j ) is the recovery priority prediction function: Among them, f k For F context,i,j The k-th feature component in the context feature vector, where M represents the dimension of the context feature vector and λ is the balance coefficient; The final set of missing region description information is formed:

6. The artificial intelligence-based record management method according to claim 2, characterized in that, S3 includes the following steps: S31. Describe information set D from missing regions desc Extract damaged area R missing and context feature set F context A global search space S is constructed based on the spatial distribution of the missing regions. search : Among them, S search This represents the search range in two-dimensional space for the damaged area of ​​the archive data, (x i,j ,y i,j ) represents the damaged area r i,j The set of coordinate points in the search space; S32. In the global search space S search In this study, the variable-scale Lévy flight algorithm is used to perform a global search based on the context feature set F. context Calculate the entropy value of the damaged area and dynamically adjust the search step size S. t : Among them, S t S represents the step size of the t-th iteration in the search process. max Let H(F) be the initial maximum step size. context,i,j ) represents the damaged area r i,j Information entropy in the context feature set is used to measure the complexity of the region. The higher the information entropy value, the more severe the information loss in the region. λ1 is a control factor. S33. Calculate the search jump step length L based on the variable-scale Levy flight model. α And update the next search point (x) based on the current search point's location. t+1 ,y t+1 ): (x t+1 ,y t+1 )=(x t ,y t )+S t ·L α ·(cosθ,sinθ); Where cosθ and sinθ represent the horizontal and vertical components of the jump direction, respectively, L α To follow the Lévy distribution: Where Γ(α) is the gamma function, used to normalize the calculation of the jump step size, θ is a uniform random angle, representing the direction of the search jump, and s is a random variable that controls the change of the jump step size; S34. Calculate the updated search point (x) t+1 ,y t+1 The matching degree S(x) between the boundary of the corresponding damaged area and the boundary of the damaged area. t+1 ,y t+1 ), determine whether to enter local optimization mode, and obtain the local optimization step size S. local,t+1 ; S35. Introduce a directional weight vector W in the local optimization mode. t Optimize the search direction and calculate the search direction adjustment parameters: Among them, w x and w y These represent the matching degree in the x and y directions, respectively. Optimized search point update formula: (x t+2 ,y t+2 )=(x t+1 ,y t+1 )+S local,t+1 ·L α ·W t ; Among them, L α In local search mode, the value range is adjusted to 1.5 < α < 2; S36. During the local optimization process, store the local optimum. If K consecutively... jump If the optimal value is not updated in a round of search, a small-range jump is performed to prevent local convergence; S37. Based on the spatial distribution of search points and their matching degree with damaged areas, calculate the target fitness function F(x). t ,y t ), used to filter the optimal search path: Where S(x) t+2 ,y t+2 G(x) represents the matching degree between the optimized search point and the boundary of the damaged area. t+2 ,y t+2 The optimized search point coverage within the damaged area is represented by ( ). The normalized convergence rate, where w1, w2, and w3 are the weight parameters for adjusting various influencing factors; S38. Based on the convergence of the search trajectory, select the optimal search path and construct a global search candidate dataset D that matches the missing region. global :

7. The artificial intelligence-based record management method according to claim 6, characterized in that, The matching degree S(x) of the boundary of the damaged area in S34 t+1 ,y t+1 )for: Among them, B missing Let b be the set of boundaries of the damaged area. k γ1 represents the coordinates of the boundary point, and γ1 is the matching degree control factor, which determines the weight of the matching degree's influence on the search point. The matching degree reflects whether the search point enters a highly relevant region. When the boundary of the damaged region is broken, the matching degree S(x) t+1 ,y t+1 If τ > 0, then local optimization mode is activated, and the step size is dynamically adjusted. S local,t+1 =S max ·(1-S(x t+1 ,y t+1 )); Where τ is the local optimization trigger threshold, S local,t+1 To optimize the step size locally, the step size tends to be 0 when the matching degree is high.

8. The artificial intelligence-based record management method according to claim 2, characterized in that, S4 includes the following steps: S41. Obtain the global search candidate dataset D global To address the complexity of different damaged areas in archives, an adaptive multi-level local optimization population P is constructed. multi Different levels of optimization populations undertake different optimization tasks during the search process: Among them, P multi This represents a multi-level optimization population, where individuals from all optimization levels constitute a complete search space, P. m This represents the m-th layer optimization group, which divides the damaged regions based on their complexity. This represents the k-th search individual in the m-th layer of the optimization population, i.e., the local optimization candidate point, where M is the optimization layer number; The computational complexity of calculating the damaged area is C. damage When the complexity of the damaged area is C damage The higher the value, the greater the structural complexity of the damaged area. The Gray Wolf optimization algorithm increases the number of optimization layers M accordingly. Where, σ G (r i,j ) indicates the damaged area r i,j The standard deviation of the gradient within a region measures the marginal complexity within that region. The degree of abrupt changes in regional texture is calculated for the Laplacian operator, describing the structural discontinuities in the damaged region; S42. During the optimization process, different optimization populations at different levels adopt differentiated fitness functions to adapt to different damage characteristics of the archival data. A multi-level adaptive fitness function is defined as follows: in, To ensure the matching degree between local search points and the boundaries of damaged areas, the filling points should match the morphological characteristics of the damaged areas. To ensure contextual consistency, measure whether the supplementary data conforms to the logic and content structure of the original archival data. Texture similarity measures the similarity between the filled-in point and the texture features of the neighboring recovered data. To ensure local data smoothness and maintain the continuity of the filled data within local regions, It represents the increment of information entropy, measuring the contribution of the fill-in point to the overall information integrity. S43. In the archive filling task, the search direction is guided by the boundary information of the archive data. Based on the traditional gray wolf optimization search, texture gradient guidance is introduced to construct a texture-guided gray wolf optimization search: in, Let be the search position of the k-th gray wolf in the m-th layer of the optimization population at iteration t+1, representing the positional change of this filling point during the optimization process. These represent the positions of the three wolves representing the optimal, second-best, and third-best solutions in the optimization population, respectively. m To control the search step size and convergence speed coefficients of individual gray wolves, B represents the distance between the individual gray wolf and the positions of the optimal, second-best, and third-best solutions, respectively. m For gradient-guided terms: in, γ2 represents the local gradient direction, guiding the filling point to converge towards the optimal texture matching direction; γ2 is the gradient guidance weight. S44. In the archival data restoration task, a hybrid step-size strategy is adopted to optimize the search for individual gray wolves, dynamically adjusting the search step-size to balance accuracy and convergence speed: Among them, 2·(1-t / T max ) controls the global convergence trend, λ m • N(0,1) enhances search diversity through hierarchical perturbation; S45. Introduce a local convergence detection mechanism. If a search point is continuously K... stuck If the fitness is not improved in a round of iteration, then a jump is performed: S46. After the Grey Wolf optimization algorithm converges, the optimal filling point is selected, and the final locally optimized filling data is formed: in, To determine the coordinates of the k+2 search individuals in the optimized population at layer m, F m (p k+2 ) represents the optimized multi-level adaptive fitness function.

9. The artificial intelligence-based record management method according to claim 8, characterized in that, The texture similarity The calculation is as follows: Where G(x,y) is the local gradient feature, N k For the filled data points within the neighborhood of the search point; Local data smoothness The calculation is as follows: Where I(x,y) represents the pixel value or text feature value of the archive data; The information entropy increment The calculation is as follows: Among them, H before and H after These are the entropy values ​​of the damaged areas before and after filling in the damage, respectively.

10. The artificial intelligence-based record management method according to claim 2, characterized in that, S5 includes the following steps: S51. Obtain the locally optimized fill data and the preprocessed archive data. According to the data format, storage structure and encoding method, normalize the fill data. Perform preliminary fusion of the normalized fill data and the preprocessed archive data to construct an initial data fusion set. S52. Analyze the local features of the area where the data is filled, extract the local feature vector of the damaged area, and calculate the local similarity between the data and the neighboring areas, including boundary morphology similarity, local structural consistency and content feature matching degree. Based on the local similarity calculation results, determine whether the data filled conforms to the local style features of the original archive. When the local similarity is lower than the preset threshold, adjust the data filled. S53. Based on the pattern characteristics of historical archive data, the style, semantics and logical consistency of the merged archive data are matched. The style adaptability of the fill data is analyzed according to the language characteristics, layout style and data structure rules of the archive content. For fill data with a matching degree lower than the threshold, the style adjustment function is used to correct it so that the fill data is consistent with the original archive in terms of font format, text semantics, structural logic and contextual coherence. S54. After completing the local similarity adjustment and style matching, the overall filling result is evaluated globally, the matching degree between the filling data and the entire archive content is calculated, and based on the global matching result, the filling quality of the data is selected to be higher than the threshold, and the filling data with low matching degree is adjusted again to make the filling data and the entire archive achieve the best matching in terms of content logic, format layout and semantic consistency. S55. After completing the local and global consistency matching, the final selected fill data is archived to generate the preliminary corrected archive data.

Images