An intelligent monitoring management method and system based on file digitization

Abstract

The application discloses an intelligent monitoring management method and system based on file digitization, relates to the technical field of data management, and comprises the following steps: collecting multi-source file data and preprocessing, adopting a multi-modal BERT model to perform feature fusion on different data sources, and generating unified semantic representation; performing semantic tagging on the file data through a multi-label classification model, and utilizing a graph database to construct a semantic graph of file content, representing the correlation between files, constructing a semantic index tree based on the semantic graph, and optimizing the rapid positioning and calling of file content; recording the changes of each file version, positioning the change position based on the semantic index tree, identifying the semantic changes of the file through a semantic difference comparison algorithm and recording the hash, and performing granularity division on the file content through the semantic boundaries of the nodes at all levels in the index tree and generating a user access strategy. The application realizes dynamic perception and risk early warning on user behavior, and effectively improves the intelligence and security of the file management system.

Description

Technical Field

[0001] This invention relates to the field of data management technology, and in particular to an intelligent monitoring and management method and system based on archive digitization. Background Technology

[0002] With the rapid development of information technology, cloud computing, and artificial intelligence, traditional archival management methods are gradually evolving towards digitalization, networking, and intelligence. Archival digitization systems not only handle the archiving, storage, and retrieval of structured and unstructured information, but also undertake the tasks of recording key business processes, knowledge accumulation, and compliance auditing. Currently, numerous documents and products have implemented basic functions such as centralized management, keyword retrieval, full-text retrieval, and access control for digital archives. In recent years, the rise of semantic technology, natural language processing (NLP), and graph neural networks (GNN) has provided new possibilities for deep semantic modeling of archival content. Utilizing deep learning models such as BERT to extract features from text content and perform classification, clustering, and labeling has become a mainstream approach. Simultaneously, with the large-scale presence of multimodal data (such as images, text, and structured forms) in archives, cross-modal fusion processing and unified representation have become technical challenges and research hotspots in intelligent archival management systems.

[0003] However, existing technologies still have several shortcomings. On the one hand, most systems only extract and classify surface-level information from document content, lacking the ability to model the semantic hierarchy and relationships between archives, making it difficult to support archive retrieval, comparison, and update tracking in complex contexts. On the other hand, current access control mechanisms mostly rely on static role assignment or rule-driven access control, unable to refine authorization based on the semantic granularity of archive content, let alone dynamically adapt to user behavior characteristics. Furthermore, the process of changing archive content versions lacks detailed tracking and comparison mechanisms, making it difficult to accurately record and identify semantic changes. More importantly, current archive systems mainly focus on recording access logs for monitoring user behavior, lacking deep modeling based on semantic paths and behavioral graphs, thus failing to effectively identify potential risky behaviors. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an intelligent monitoring and management method and system based on archive digitization, which solves the problems of insufficient semantic understanding of content, weak version change recognition ability, coarse-grained permission management, and limited user behavior monitoring ability in existing archive digitization technologies.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an intelligent monitoring and management method based on archive digitization, comprising,

[0008] Collect and preprocess multi-source archival data, and use a multimodal BERT model to fuse features from different data sources to generate a unified semantic representation;

[0009] The archival data is semantically labeled using a multi-label classification model, and a semantic graph of the archival content is constructed using a graph database to represent the relationships between archives. A semantic index tree is then built based on the semantic graph to optimize the rapid location and retrieval of archival content.

[0010] Record changes for each archive version, locate the change location based on the semantic index tree, identify the semantic changes of the archive through the semantic difference comparison algorithm and record the hash, and divide the archive content into granular parts and generate user access policies through the semantic boundaries of each level of nodes in the index tree.

[0011] Record the semantic path of user access profiles and construct user behavior graphs. Identify potential abnormal behaviors through behavior graph analysis, and use time series analysis and fuzzy mining algorithms for risk prediction and behavior warning.

[0012] As a preferred embodiment of the intelligent monitoring and management method based on archive digitization described in this invention, the method includes: semantically tagging archive data using a multi-label classification model, constructing a semantic graph of archive content using a graph database to represent the relationships between archives, and constructing a semantic index tree based on the semantic graph, including:

[0013] The Label-GAT model is constructed by defining the set of labels required for file management, assigning an embedding vector to each label, and using the joint representation vector A and the label vector together as nodes to establish a heterogeneous graph structure.

[0014] Graph attention propagation is performed on heterogeneous graph structures, and each labeled node will obtain a semantic representation related to the current sample.

[0015] The joint representation vector A is split into three sub-vectors: image, text, and structured field, which serve as three modal information sources. By calculating attention weights, the matching degree of the label under different modalities is determined, and the final fused label representation is obtained.

[0016] The fused representation of all labels is input into a shared fully connected layer, and the prediction result is output through the Sigmoid function to train the Label-GAT model. For any new file, the corresponding joint representation vector is obtained and input into the trained Label-GAT model. After graph attention propagation and multimodal weighted interaction, a label probability value is output. According to the set threshold, the label with the probability value greater than the threshold is determined as the classification result of the file.

[0017] Based on the classification result set, the basic entity types in the graph are determined, semantic edge relationships are established between entities in the graph, all entities and their semantic connections are uniformly organized into triples in the graph structure, the graph structure is standardized, the initial graph structure is constructed, and semantic reasoning rules are defined.

[0018] The semantic graph structure is standardized, and a knowledge graph embedding model is used to train the graph, mapping each entity and relation to a unified low-dimensional vector space. A density-based clustering algorithm is used to cluster all archive entities. Based on the clustering results, the cluster centers and their corresponding entities are divided into a hierarchical semantic index tree. The upper-level nodes of the tree represent semantic cluster centers, and each center node is connected to its subordinate entity nodes or sub-cluster nodes. By controlling the number and depth of branches at each level, a multi-level index tree is constructed.

[0019] As a preferred embodiment of the intelligent monitoring and management method based on archive digitization described in this invention, the method of using a multimodal BERT model to fuse features from different data sources and generate a unified semantic representation refers to collecting multi-source archive data, including image data, text data, and structured data, and preprocessing the collected data.

[0020] The preprocessed image is uniformly divided into image blocks of fixed size. Based on the spatial position of the image block in the whole image, a corresponding positional information representation is assigned to each image block, thereby generating an image block feature matrix containing local image content and spatial position information. The image block sequence is modeled using a multi-layer coding structure based on an attention mechanism to extract semantic relevance and global feature expression in the image modality.

[0021] The preprocessed text data is decomposed into a series of words or sub-word units. The BERT model is used to map the segmented text into vector representations. Each token obtains a fixed-dimensional embedding vector, which forms the initial text feature matrix. A deep language encoder is used to model the text sequence to obtain text feature representations containing contextual semantics.

[0022] After encoding and mapping image feature sequences and text feature sequences to a unified semantic space, information interaction is achieved through a cross-modal attention mechanism, and finally they are aggregated into an image-text fusion feature vector F.

[0023] The preprocessed structured fields are concatenated into a structured feature vector. The preprocessed structured feature vector is then input into the hybrid modeling path. The optimal modeling strategy is selected for different data types of the fields. The feature results are concatenated and fused, and then input into the multilayer perceptron encoder to obtain the semantic embedding vector Q of the structured fields.

[0024] The image-text fusion feature vector F and the structured semantic embedding vector Q are aligned in terms of vector dimensions and then concatenated to construct a joint representation vector A.

[0025] As a preferred embodiment of the intelligent monitoring and management method based on archive digitization described in this invention, the following steps are taken: recording changes in each archive version, locating the change location based on a semantic index tree, identifying semantic changes in the archive through a semantic difference comparison algorithm and recording hash values; when an archive is detected to have been modified, rewritten, or updated, the constructed semantic index tree is immediately invoked, the changed content is encoded by embedding vectors, and semantic similarity is calculated with the embedding vectors of the path nodes where the original archive is located in the index tree to determine the closest path node of the change point in the index tree and locate the change anchor point of the archive's semantic position.

[0026] Extract the current modified content and the previous version of the file. Use a unified semantic encoding model to compare the distance between the two versions in the sentence vector or paragraph vector space. Combine the changes in the tag path of key fields to calculate the overall semantic difference. Generate a structured snapshot of the current modified content and perform full content hash calculation.

[0027] As a preferred embodiment of the intelligent monitoring and management method based on archive digitization described in this invention, the step of granularly dividing archive content and generating user access strategies by using the semantic boundaries of nodes at each level in the index tree refers to treating nodes at each level in the index tree as semantic classification boundaries, identifying semantic breakpoints and dividing semantic segments at different levels through similarity analysis of semantic vectors between nodes, binding each archive version to the corresponding semantic level in the corresponding index path, generating permission vectors based on the user's organization, role, and access scenario, comparing the vectors with each semantic segment, and determining whether the user has access rights through a similarity threshold.

[0028] As a preferred embodiment of the intelligent monitoring and management method based on archive digitization described in this invention, the step of identifying potential abnormal behaviors through behavioral graph analysis and using time series analysis and fuzzy mining algorithms for risk prediction and behavioral early warning refers to continuously constructing a user-centric behavioral graph by recording the semantic path nodes, access time, and archive version traversed by users when accessing archives. This involves modeling and analyzing the behavioral characteristics of users in each semantic category, including access frequency, path jump patterns, and access depth, identifying behavioral preferences and routine patterns. When user behavior exhibits characteristics of access paths or sudden changes in access density that differ from historical patterns, a time series model is constructed based on the behavioral graph, and fuzzy mining algorithms are used to identify potential abnormal behavioral patterns, generating corresponding behavioral early warning processes.

[0029] As a preferred embodiment of the intelligent monitoring and management method based on archive digitization described in this invention, the behavior early warning process includes immediately freezing the user's current access permissions or forcibly downgrading them to read-only mode, automatically generating structured alarm information, pushing it to the administrator interface, recording the complete behavior chain, and appending risk events to the on-chain log.

[0030] Secondly, the present invention provides an intelligent monitoring and management system based on archive digitization, comprising,

[0031] The semantic fusion module is used to collect multi-source archival data and perform structured cleaning, and generate a unified semantic representation using the multimodal BERT model;

[0032] The graph construction module is used to semantically label archive content through a multi-label classification model and build a graph database to represent the semantic relationships between archives.

[0033] The semantic index building module is used to build a semantic index tree based on the semantic graph, enabling rapid file location and recording the path and hash information of each version change and semantic change.

[0034] The permission policy module is used to divide the file content granularity according to the semantic boundaries at each level in the index tree and generate fine-grained user access permission policies.

[0035] The risk warning module is used to record user access behavior to build a behavior graph, and combine time series analysis and fuzzy mining to identify abnormal behavior and generate risk warning signals.

[0036] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the intelligent monitoring and management method based on archive digitization as described in the first aspect of the present invention.

[0037] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent monitoring and management method based on archive digitization as described in the first aspect of the present invention.

[0038] The beneficial effects of this invention are as follows: This invention achieves a unified semantic fusion representation of multi-source archival data through a multimodal BERT model, and constructs an archival semantic graph and index tree by combining graph neural networks and graph databases, thereby achieving efficient and traceable content location and evolution tracking. At the same time, through continuous modeling of user behavior semantic paths, combined with time series analysis and fuzzy mining technology, it achieves dynamic perception and risk warning of user behavior, effectively improving the intelligence and security of the archival management system. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of the intelligent monitoring and management method based on archive digitization in Example 1.

[0041] Figure 2 This is a structural diagram of the intelligent monitoring and management system based on archive digitization in Example 1.

[0042] Figure 3 This is a schematic diagram of the process of constructing the semantic graph and index tree in Example 1. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0045] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0046] Example 1, refer to Figures 1-3 , which is the first embodiment of the present invention. This embodiment provides an intelligent monitoring and management method based on file digitization, including the following steps:

[0047] S1. Collect multi-source file data and perform preprocessing. Use a multi-modal BERT model to perform feature fusion on different data sources and generate a unified semantic representation;

[0048] Specifically, using a multi-modal BERT model to perform feature fusion on different data sources and generate a unified semantic representation means collecting multi-source file data, including image data, text data, and structured data, and performing preprocessing on the collected data. Among them, the image data comes from scanned paper documents, on-site photos, or charts, and needs to be uniformly collected from the file system, cloud storage, or scanning device. The text data comes from PDF, Word documents, or web page content, and the pure text content needs to be extracted through a parser or OCR tool. The structured data comes from Excel tables, CSV files, or database export results, and needs to be batch-extracted through a database interface or file reading tool;

[0049] Perform size unification, pixel standardization on the collected image data, and evenly divide each image into image blocks of a fixed size. Use NLP tools to disassemble the text data into word or sub-word units, and remove stop words, including "de", "le", "he", etc. Use the BERT model to map the segmented text into a vector representation, and each token gets an embedding vector of a fixed dimension (such as 768 dimensions). A token refers to the smallest processing unit after the text is segmented before entering the model. For structured data, perform missing value processing, numerical standardization, categorical field encoding, and field embedding;

[0050] The missing value processing refers to using mean filling or median filling to ensure data integrity. The numerical standardization refers to performing Z-score standardization (subtracting the mean and dividing by the standard deviation) on all numerical fields to ensure consistent dimensions. The categorical field encoding refers to using One-Hot Encoding to convert categorical fields into vector form. The field embedding refers to generating vectors for column names and corresponding values respectively;

[0051] Evenly divide the preprocessed images into image blocks of a fixed size. According to the spatial position of the image blocks in the entire image, assign corresponding position information representations to each image block, thereby generating an image block feature matrix containing local image content and spatial position information. Use a multi-layer encoding structure based on the attention mechanism to model the image block sequence and extract the semantic correlation and global feature expression in the image modality, specifically including:

[0052] Two-dimensional positional encoding is added to the feature sequence of image patches to explicitly represent the spatial position information of each image region in the original image. This positional encoding is implemented through a learnable embedding method, which enables the model to maintain the spatial structure of the image when processing unordered sequences.

[0053] The sequence of image patches with positional encoding is input into a multi-layer encoding structure based on an attention mechanism. This structure consists of several stacked self-attention encoding modules. Each module contains a multi-head self-attention mechanism and a feedforward neural network to simultaneously capture the local correlations and global semantic information between image patches. In each encoding module, the image patch feature sequence is mapped to query, key, and value vectors, respectively, and attention weights are calculated between each image patch to measure their semantic relevance. The image patch features are reweighted and combined according to the attention weights, so that the representation of each image patch not only contains its own information but also incorporates the contextual information of its associated regions.

[0054] Through a multi-layered stacked self-attention structure, the model gradually expands its receptive field, thereby achieving semantic construction from local to global. Early layers capture local dependencies in neighboring regions, while subsequent layers gradually enhance the connections between distant regions in the image, ultimately forming an image feature representation with global perception capabilities.

[0055] This modeling process not only preserves the independent semantic representation of each image patch, but also constructs the contextual structure between regions in the hidden layer, enabling the system to understand the role and relationship of each part in the overall semantic structure of the image, thereby extracting structured global image information;

[0056] The output from the multi-layer coding structure is still an image patch feature sequence, but this sequence has been remodeled and has a higher level of semantic relevance and cross-region information fusion capability, providing a deep representation basis for image modality for subsequent multimodal interaction and fusion.

[0057] The preprocessed text data is decomposed into a series of words or sub-word units. The BERT model is used to map the segmented text into vector representations. Each token obtains a fixed-dimensional embedding vector, which forms the initial text feature matrix. A deep language encoder (such as the Transformer encoder) is used to model the text sequence to obtain text feature representations containing contextual semantics.

[0058] After encoding and mapping image and text feature sequences to a unified semantic space, information interaction is achieved through a cross-modal attention mechanism, ultimately aggregating them into an image-text fusion feature vector F. The specific steps include:

[0059] Image patch feature sequences and text feature sequences are input into independent modal encoders. The image modal encoder models the spatial relationships between image patches and outputs a deep feature representation sequence of the image. The text modal encoder models the contextual relationships between text tokens and outputs a deep semantic feature representation sequence of the text. Both types of features are kept in the same order as the original input and are accompanied by positional encoding to ensure that the relative order of semantic information in the sequence is preserved. The image modal feature sequences and text modal feature sequences are dimensional alignment processed separately, mapping their respective vector dimensions to a unified semantic space. This mapping process introduces learnable projection parameters to adjust the image features and text features into vector representations of the same dimension, thus providing a foundation for subsequent fusion operations.

[0060] After semantic space alignment is completed, a cross-modal attention mechanism for image-guided text is established, using image feature sequences as the query source and text feature sequences as the response source. Through this mechanism, the semantic similarity between each image patch feature and all text token features is calculated, thereby generating an attention weight matrix. This weight reflects the degree of attention that the image patch pays to different text fragments, and is used to capture the correspondence between potential content in the image and text description. The text feature sequence is weighted and summed according to the attention weight to generate a text augmented representation that corresponds one-to-one with each image patch feature. The original features of the image patch and its corresponding text augmented representation are combined to construct new image-text joint features. This process enables the semantic expressive power of the image modality to be further supplemented and guided by the text content, thereby improving its understanding of the actual archival content.

[0061] The image-text joint feature sequence is uniformly aggregated, and the dominant semantic information in the whole set of features is extracted by global average pooling. This aggregation operation can compress the original joint sequence and form a compact, multi-dimensional representation vector, which represents the overall semantics of the current archive sample in the image and text modalities. The aggregated image-text fusion representation vector is output as the final result of the multimodal feature fusion stage.

[0062] The preprocessed structured fields are concatenated into structured feature vectors, which are then input into the hybrid modeling path. Optimal modeling strategies are selected for different data types of the fields. Specifically, numerical fields are fed into a Gradient Boosting Tree (GBDT) model to extract the nonlinear mapping relationship between the numerical and target semantic distributions and output probabilistic features with semantic bias information. Categorical fields are input into a Support Vector Machine (SVM) model, and the discrete prediction boundaries output by Platt Scaling are mapped to continuous probability distributions to better capture similarity information between categories. The probabilistic features output by the two models represent the semantic positions of the numerical and categorical fields in the target label space, respectively. These two feature results are concatenated and fused, and then input into a Multilayer Perceptron (MLP) encoder to further extract the global expressive power of structured information in a high-dimensional semantic space. This process can be seen as a deep modeling of the implicit contextual patterns and cross-field dependencies within the structured fields.

[0063] The output after MLP modeling is the semantic embedding vector Q of the structured field, whose dimension is consistent with the image-text fusion representation vector. This structured feature representation retains the key information of the original structured field, while also having the ability to express information aligned with unstructured information.

[0064] The Platt Scaling technique, also known as Platt calibration or Sigmoid calibration, is a technique used to convert the output of a classifier into a probability estimate.

[0065] The image-text fusion feature vector F and the structured semantic embedding vector Q are aligned in terms of vector dimensions. If the two dimensions are inconsistent, a single or multi-layer linear transformation module is introduced to map the structured features to the same dimensional space as the image-text features, ensuring that the three modalities are additive and splicable in the fusion operation.

[0066] The image-text fusion feature vector F and the structured semantic embedding vector Q are concatenated to construct a joint representation vector A. This vector retains the semantic distribution features of image, text and structural fields in the feature space, and is a complete mapping representation of the archive sample in the multimodal semantic space.

[0067] By constructing a feature fusion architecture based on a multimodal BERT model, deep modeling and efficient fusion of image, text, and structured data in a unified semantic space were achieved, significantly improving the understanding and processing capabilities of multi-source archival data. After spatial partitioning and location encoding, image data retained local details and global structural information; text data used BERT to obtain contextual semantic features, enhancing the accuracy of semantic expression; and structured data used GBDT and SVM modeling to extract nonlinear and semantic distribution features from numerical and categorical fields, ensuring the integrity and interpretability of structural information. The cross-modal attention mechanism enabled image-guided text interaction, allowing image and text information to complement each other and strengthen semantic alignment, further enhancing the deep understanding of complex archival content. Finally, a joint representation vector was constructed through unified projection and vector concatenation, integrating the semantic features of the three types of data, providing high-quality feature support for downstream tasks such as classification, retrieval, archiving, and risk assessment.

[0068] S2. Semantically label the archive data using a multi-label classification model, and construct a semantic graph of the archive content using a graph database to represent the relationships between archives. Based on the semantic graph, construct a semantic index tree to optimize the rapid location and retrieval of archive content.

[0069] Specifically, a multi-label classification model is used to semantically label archival data, and a semantic graph of archival content is constructed using a graph database to represent the relationships between archives. Based on the semantic graph, a semantic index tree is built to optimize the rapid location and retrieval of archival content, including:

[0070] The Label-GAT model is constructed by defining the set of labels required for archive management, assigning an embedding vector to each label, and maintaining the same semantic dimension as the joint representation vector A.

[0071] By using the joint representation vector A and the label vector together as nodes, a heterogeneous graph structure is established. In this graph, the sample nodes are connected to all label nodes, and the label nodes are also connected by edges through semantic similarity. In this way, the model can capture the connection between samples and labels, and also model the correlation between labels.

[0072] Graph attention propagation is performed on a heterogeneous graph structure. For each labeled node, its representation is updated based on the information of its neighbors. During the update process, the contribution of the neighbors is dynamically determined by the attention weights, enabling each labeled representation to integrate key information from the sample node and other labeled nodes. After multiple layers of propagation, each labeled node obtains a semantic representation related to the current sample.

[0073]

[0074]

[0075] In the formula, N(t) is the vector representation of node t after the g-th layer graph attention network, and N(t) is the set of neighbors of node t (including nodes connected to it by edges). It is the attention weight of node r to the center node t. It is the linear mapping matrix of the g-th layer, used to transform the input features. This represents node r in the previous layer. σ() is an activation function (such as ReLU or ELU) used to increase non-linear expressiveness. a is a learnable attention vector that controls the weights of the concatenated vectors. LeakyReLU is an activation function that processes the linear transformation result after concatenation. exp() is the exponential operation of Softmax used to normalize the attention weights. It is the representation of node t in the previous layer. It is the representation of node j in the previous layer;

[0076] To further integrate multimodal features, after the label representation is updated, it needs to be weighted and interacted with different parts of the three-modal vector. The joint representation vector A is split into three sub-vectors: image, text, and structured field, which serve as the three modal information sources respectively. By calculating attention weights, the matching degree of the label under different modalities can be determined, resulting in the final fused label representation.

[0077]

[0078]

[0079] In the formula, It is the attention weight of label k to modality p (such as image F, text T, structural field Q). It is the final representation vector of label k in modality p. It is the semantic vector representation of modality p. It is the fused label k representation vector;

[0080] The fused representation of all labels is input into a shared fully connected layer, and the prediction result is output through the Sigmoid function. Each label will get a probability value between 0 and 1, representing the likelihood that the file belongs to that label.

[0081] The Label-GAT model is trained by taking the historical joint representation vector A as input and the real multi-label set as output. The model calculates the prediction result through forward propagation and compares it with the real labels to obtain the loss value. The binary cross-entropy is used as the loss function. Backpropagation updates the parameters in the model, including label embedding, weight matrix and attention parameters. The training process is iterated repeatedly until the maximum number of iterations is reached to obtain the trained Label-GAT model.

[0082] For any new file, the corresponding joint representation vector is obtained and then input into the trained Label-GAT model. After graph attention propagation and multimodal weighted interaction, a label probability value is output. According to the set threshold (based on experience), the labels with probability values ​​greater than the threshold are determined as the classification results of the file (such as: file type, topic, time, sensitivity, etc.).

[0083] Based on the tag set, determine the basic entity types in the graph, including archive entities, tag entities, and time, event, and organization auxiliary semantic entities;

[0084] The archive entity is used to represent each archive data sample;

[0085] The label entity represents the label result obtained from the semantic classification stage;

[0086] Time-based entities represent the time period to which an archive belongs;

[0087] Event and organization entities are used to label the specific events involved in the archives and the organizations to which they belong, respectively. All types of entities are named and coded in a structured manner.

[0088] Semantic edge relationships are established between entities in the graph. These edge relationships are mainly derived directly from the information extracted in the previous step. For example, if a file sample is assigned several labels through multi-label classification, then an "attribution" relationship is established between the file and the corresponding label entity.

[0089] If the file's text or structure fields indicate that it references another file, then a "reference" relationship is established between the two file entities;

[0090] If the archive contains clear archiving time or occurrence time information, a "time relationship" is established between it and the time-related entity. In addition, if the structure field contains related events or affiliated organizations, a corresponding relationship is also established between them and the event and organization-related entities, respectively.

[0091] All entities and their semantic connections are organized into triplets in a graph structure. Each relation consists of a starting entity, a relation type, and a target entity. In this way, information such as archives and their semantic tags, related archives, occurrence time, and event attribution are all abstracted into nodes and edges in the graph structure.

[0092] The graph structure is standardized to meet the data format requirements of the graph database, and then imported into the graph database in batches for unified storage. In the import stage, the entry of various entity nodes is completed first, and various edge relationships are established in sequence. The connection direction and attribute type of the edges are marked. At this point, various semantic units such as archives, tags, time, references, events, and organizations are integrated into a unified graph database structure, and all archive semantic relationships have a structural foundation that can be queried and reasoned.

[0093] After completing the initial graph structure construction, semantic reasoning rules are defined to identify indirect relationships in the graph that are not explicitly present but can be logically deduced. The reasoning rules are constructed based on the connection patterns of the graph structure. For example, if a file C references a file B and file B belongs to a certain tag, then it can be inferred that there is also a certain relationship between file C and the tag. If two files C and B appear in the same time period, it can be inferred that they are concurrent or related to some extent. If both files C and B involve the same event or the same organization, it can be inferred that there is a certain relationship between them. These rules are uniformly defined and executed in batches in the graph structure to generate a batch of new semantic relationship edges to enrich the semantic structure of the graph.

[0094] Once the graph is built and reasoning is complete, users can retrieve the graph through the graph database's query interface, especially the Cypher query language. Through the query interface, users can quickly find information of interest such as files, tags, times, and reference relationships from the graph.

[0095] The semantic graph structure is standardized by uniformly numbering all nodes such as archive entities, tags, events, organizations, and times in the graph. At the same time, triples (such as archive C, belonging to, project contract) are formatted into the standard knowledge graph input form, and entity vocabulary, relation set and adjacency structure matrix are established.

[0096] The knowledge graph is trained using a knowledge graph embedding model (such as TransE, TransR, or ComplEx) to map each entity and relation into a unified low-dimensional vector space. The embedding results preserve the semantic proximity between entities, making semantically similar files or tags close to each other in the vector space, and obtaining the embedding vector (semantic coordinates) corresponding to each graph node.

[0097] All archival entities are clustered using a density-based clustering algorithm, specifically including:

[0098] Using the semantic embedding vectors of all archival entities as input data, a similarity matrix between entities is constructed by calculating the distance between vectors (such as Euclidean distance or cosine similarity). Based on the parameters set by the clustering algorithm (e.g., eps represents the neighborhood radius and min_samples represents the minimum number of points in the cluster in DBSCAN), regions with high local density in the vector space are identified, and multiple entity clusters are automatically formed. Each cluster represents a class of archival entities with similar semantics. The algorithm automatically marks noise points (i.e., abnormal entities that cannot be classified into any cluster) as independent samples. After clustering, for each cluster, a center vector representing the semantic cluster is obtained by calculating the average of all entity vectors in the cluster. This center vector can be regarded as the semantic "centroid" of the class in the semantic space. This clustering algorithm can automatically divide semantically similar archival groups according to the density relationship of entities in the vector space, while removing isolated points or abnormal samples. The cluster center represents the semantic center of a class of archives.

[0099] Based on the clustering results, a hierarchical semantic index tree is constructed by dividing the cluster centers and their corresponding entities. The upper-level nodes of the tree represent semantic cluster centers (i.e., semantic categories). Each center node connects to its subordinate entity nodes or sub-cluster nodes. By controlling the number and depth of branches at each level, a multi-level index tree is constructed, specifically including:

[0100] Based on the semantic clusters output by clustering, each cluster is regarded as a semantic category node, and the embedding vector of its cluster center is used as the semantic representation of the node, forming the first-level structure of the index tree. On this basis, the distribution of entities within each cluster is further analyzed:

[0101] If the number of entities in a cluster is large, then the cluster is further subdivided into sub-clusters, and the entities are further subdivided into multiple sub-nodes, which serve as the next-level branch nodes of the cluster center. Conversely, if the cluster is small or the entities within it are semantically close, then these entities are directly attached as leaf nodes to the cluster center node. As all clusters expand layer by layer and connect to specific file entities, a multi-level index tree is gradually formed.

[0102] To ensure the stability of the tree structure in terms of query efficiency and maintainability, the number of nodes at each level, the depth of the tree, and the capacity of leaf nodes need to be controlled during the construction process. When the number of branches of a node exceeds the set threshold (based on experience), local re-clustering or structural reconstruction is automatically triggered to maintain the overall structural balance. While attaching nodes, a complete semantic path label is constructed for each node starting from the root node, and its embedding vector and path position in the semantic space are recorded.

[0103] Finally, the entire index tree is serialized into a structured data model, with each node containing its ID, semantic representation, parent-child node relationship, path information, and related file attributes, thus unifying the semantic structure and entity index.

[0104] Through the above consecutive steps, the clustering result naturally transitions into an index tree structure with semantic hierarchy awareness, providing a structural foundation for subsequent semantic query optimization and access strategy formulation;

[0105] To enable user queries to be efficiently mapped to semantic paths in the index tree, a semantic query encoder (such as BERT or MPNet as the encoder backbone) is further trained to map query content (keywords, tag combinations, etc.) to the same vector space as the tree structure. Based on this, the embedding representations of the query encoder and the index path are updated simultaneously through a joint optimization method, so that semantically similar queries can more efficiently hit the correct tree path.

[0106] By constructing the Label-GAT model, label nodes and sample nodes are modeled together, and semantic edge relationships between labels are introduced, enabling the model to have stronger label semantic perception capabilities in multi-label classification tasks, significantly improving classification accuracy and collaborative expression effects between labels. Secondly, by combining multimodal features (images, text, and structural fields) for weighted interaction, the semantic loss problem caused by a single modality is effectively alleviated, enhancing the system's understanding of complex file content and robustness of label matching.

[0107] In terms of knowledge organization, the constructed semantic graph connects various entities such as archives, tags, time, events, and organizations in the form of triples, realizing the explicit expression and unified management of semantic relationships. Through the graph reasoning mechanism, the system can automatically discover the relationships between archives that are not explicitly recorded but are logically differentiable, improving the semantic association and retrieval breadth of archives. Furthermore, with the help of knowledge graph embedding and density clustering algorithms, semantically similar archive entities are divided into several semantic clusters, and a hierarchical semantic index tree is constructed, enabling the system to have the ability of fast positioning, semantic clustering retrieval, and path navigation.

[0108] The semantic encoder maps the query content to the same semantic space as the index tree, which significantly improves the matching efficiency and semantic accuracy of the query. Overall, this invention achieves full-process optimization from semantic understanding to semantic organization and then to semantic retrieval. It has structural scalability, retrieval intelligence and semantic reasoning capabilities, and is suitable for large-scale archive management scenarios in government affairs, law, enterprises and other fields.

[0109] S3. Record the changes of each file version, locate the change location based on the semantic index tree, identify the semantic changes of the file through the semantic difference comparison algorithm and record the hash, and divide the file content into granular parts and generate user access strategies through the semantic boundaries of each level of nodes in the index tree.

[0110] Specifically, the changes for each version of the archive are recorded, the location of the change is located based on the semantic index tree, and the semantic changes of the archive are identified by the semantic difference comparison algorithm and the hash index is recorded. When the archive is detected to have been modified, rewritten or updated, whether the title content, body fields, tag attributes or supplementary descriptions have changed, the completed semantic index tree is immediately invoked, the changed content is embedded with vector encoding, and semantic similarity is calculated with the embedding vector of the path node where the original archive is located in the index tree. The closest path node of the change point in the index tree is determined by a method such as cosine similarity, so as to accurately locate the change anchor point of the archive's semantic position.

[0111] Extract the current modified content and the previous version of the file. Use a unified semantic coding model (such as MPNet or BGE) to compare the distance between the two versions in the sentence vector or paragraph vector space. Combine the changes in the label path of key fields to calculate the overall semantic difference. Specifically, for each pair of modified fields or paragraphs, calculate their vector cosine similarity. If the similarity is lower than the threshold (based on experience) (such as 0.85), it is determined that "semantic changes have occurred". At the same time, the proportion of paragraphs with semantic changes, field coverage and their hierarchical changes in the index path are counted to obtain a "structural semantic change score".

[0112] After the comparison is completed, a structured snapshot of the current changes is generated, and a full content (including text, metadata, and semantic anchor path) hash calculation is performed. An irreversible hash fingerprint is generated using the SHA-256 algorithm. This hash value, along with the version ID, timestamp, user identity, and change path ID, is submitted as version metadata to the on-chain storage system, such as the version ledger of Hyperledger Fabric. At the same time, a "sequential connection relationship" is established between the historical version nodes and the current new version nodes in the local semantic graph, clearly recording which old version the new version evolved from. This connection relationship is used to construct a clear version evolution path and to attach new version child nodes under the original leaf nodes in the semantic index tree to form version branches.

[0113] If this change results in a change to the tag system (e.g., "Purchase Contract" becomes "Entrustment Contract"), the file content is re-embedded and its semantic cluster vector representation is locally updated, thereby triggering the clustering adjustment of related subtrees in the index tree. This ensures that structural consistency and subsequent retrieval capabilities are not affected. The entire process achieves an integrated closed loop from file change detection to semantic positioning, difference comparison, version snapshot generation, hash signature, and structural update, ensuring that every file change has a clear semantic anchor and an immutable record. It also supports tracing back the evolution path and semantic change process of any version.

[0114] By constructing a semantic index tree and introducing a semantic difference comparison algorithm, accurate detection and location of changes in archival content are achieved. Embedded vector encoding and cosine similarity calculation ensure efficient identification of semantic changes, while generating structured snapshots and calculating hash fingerprints to achieve tamper-proof version records. Writing hash values ​​and version metadata into the blockchain system gives the archives traceability and auditability. By constructing version evolution paths, the historical changes of the archives are clearly presented. The system also supports dynamic adjustment of the tag system through semantic clustering, ensuring that the index structure and retrieval performance are not affected. Furthermore, combining user operation records enhances access control and behavior auditing, comprehensively improving the security, intelligence, and reliability of the archive management system.

[0115] Furthermore, each node of the semantic index tree is regarded as a semantic classification boundary, and each node represents a relatively independent semantic domain (such as "project contract", "internal approval document", "financial statement" etc.). The archive entity is located at its last leaf node. By performing clustering distance analysis on the embedding vector of each path from the root node to the leaf node, the breakpoint of the hierarchical semantic boundary in the path is automatically identified, that is, the position where the semantic change between two levels of nodes suddenly increases. For example, if the cosine similarity of the embedding vector of a parent node and its child node is significantly lower than that of its adjacent layer (such as lower than 0.7), then the node is marked as an "access control breakpoint", indicating that the content after this level has more fine-grained sensitivity and semantic isolation value. Based on this, the archive content is bound to the path it is on, and the content is divided into several semantic segments (such as top-level semantic segment, middle-level semantic segment, and bottom-level detail segment). Each segment is labeled to identify its classification, sensitivity level, and the semantic path level number to which it is attached.

[0116] A user role-based access control model is constructed, encoding user organizational attributes (such as affiliated unit), job roles (such as project manager, auditor), and access scenarios (such as daily queries, cross-departmental retrieval, joint editing) into a set of access control vectors. These access control vectors are then compared with the semantic segment vectors on the index tree path. The comparison adopts a threshold-based authorization logic: if the cosine similarity between a user's access control vector and a semantic segment vector exceeds a set threshold (such as 0.85, based on experience), the user is deemed to have the semantic ability and organizational qualifications to access that semantic segment; if it is lower, access is directly denied; if it is equal, an upper-level manual approval mechanism or alternative fuzzy query results (such as desensitized fragments or tag prompts) are triggered.

[0117] The permission policy is ultimately recorded as an access mapping relationship consisting of path ID + hierarchy number + role label. For example, "path Y-node 3 can be accessed by role X in organization Y". This relationship will be mapped into smart contract rules on the blockchain to realize on-chain verification of access control.

[0118] Specifically, during the actual access process, the access request submitted by the user is first transformed into a query vector and located to the target path in the semantic index tree. Then, based on the user's identity, the access policy is queried to see if the access permission for the semantic segment under the path is matched. If the match is successful, the requested file segment content is returned, and the access operation (including time, user, path ID, version number, etc.) is recorded and added to the chain. If the match is not successful, the access policy returns a denial result or an alternative path suggestion.

[0119] This approach enables hierarchical, granular access control based on semantic structure, ensuring that the information obtained by each organization and each user in each access scenario is consistent with their authorization, and that all access processes are auditable, verifiable, and traceable, thereby maximizing cross-organizational collaborative sharing while ensuring data security.

[0120] S4. Record the semantic path of user access profiles and construct user behavior graphs. Identify potential abnormal behaviors through behavior graph analysis and use time series analysis and fuzzy mining algorithms for risk prediction and behavior warning.

[0121] Specifically, whenever a user performs any operation, such as accessing a file, downloading an attachment, editing or modifying it, submitting a version, forwarding or deleting a file, the key elements of the action event are recorded, including user identity, operation type, operation time, target file path (corresponding to the node ID in the semantic index tree), version number, access method (interface, web, API), etc. The action record is encapsulated in real time into a standardized event sequence, with each user as the subject, and organized in chronological order to form a continuous operation trajectory.

[0122] A "user behavior graph" is constructed based on continuous operation trajectories: each node in the graph represents an operation (such as "download-financial statement-v3" or "modify-project contract-v2"), and the node is accompanied by operation meta-information (such as the tag, semantic path, and organizational level). A directed edge is established between any two consecutive operation nodes to express the operation sequence, duration interval, or frequency statistics between behaviors. The weight of the edge can be modeled according to access density, operation inertia, and semantic jump degree. This graph can not only intuitively show the user's behavior preferences and path concentration, but also provide structured input for downstream behavior prediction.

[0123] The constructed behavior graph is then transformed into training samples for a time series model to predict subsequent user actions or identify potential anomalies. LSTM is used to model the user's action sequence, with the input being a continuous sequence of action vectors within a time window (obtained by vectorizing behavior graph nodes) and the output being the probability distribution of the predicted action category or action path.

[0124] To enhance anomaly detection capabilities, a fuzzy behavior mining mechanism is further introduced: the "semantic mutation degree" (i.e., the path jump magnitude between the current operation and the previous operation in the semantic index tree), "time interval anomalies" (such as sudden operations late at night), and "operation intensity" (such as operation frequency per unit time) of each operation are used as features to construct a comprehensive anomaly scoring function. For example, if a user suddenly jumps from the "Project Progress Documents" directory to the "Contract Performance Records" and "Financial Statements" directories in succession, and initiates a large number of download or share operations in a short period of time, then the behavior trajectory has multiple anomaly features of "path cross-domain + excessive operation density + semantic jump", and finally raises its anomaly score to above the set threshold (experimental optimization setting);

[0125] Once the abnormal score exceeds the set threshold (set in the experiment and optimization), the control response process bound to the user permission contract will be triggered immediately, including: immediately freezing the user's current access permissions or forcibly downgrading to read-only mode to prevent the risk behavior from spreading further, automatically generating structured alarm information, pushing it to the administrator interface and security audit module, recording the complete behavior chain, and appending the risk event to the on-chain log, including behavior graph fragments, score details, triggering nodes, control actions, etc.

[0126] Through this risk control mechanism that combines behavioral mapping with time-series prediction, the system not only achieves dynamic learning and anomaly detection of user operation patterns, but also establishes a complete closed loop from access control to risk response, effectively ensuring data security and behavioral controllability of the multi-party archive system in an open and collaborative environment.

[0127] This embodiment also provides an intelligent monitoring and management system based on archive digitization, including:

[0128] The semantic fusion module is used to collect multi-source archival data and perform structured cleaning, and generate a unified semantic representation using the multimodal BERT model;

[0129] The graph construction module is used to semantically label archive content through a multi-label classification model and build a graph database to represent the semantic relationships between archives.

[0130] The semantic index building module is used to build a semantic index tree based on the semantic graph, enabling rapid file location and recording the path and hash information of each version change and semantic change.

[0131] The permission policy module is used to divide the file content granularity according to the semantic boundaries at each level in the index tree and generate fine-grained user access permission policies.

[0132] The risk warning module is used to record user access behavior to build a behavior graph, and combine time series analysis and fuzzy mining to identify abnormal behavior and generate risk warning signals.

[0133] This embodiment also provides a computer device applicable to the intelligent monitoring and management method based on archive digitization, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the intelligent monitoring and management method based on archive digitization as proposed in the above embodiment.

[0134] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0135] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent monitoring and management method based on archive digitization as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent monitoring and management method based on archive digitization, characterized in that: include, Collect and preprocess multi-source archival data, and use a multimodal BERT model to fuse features from different data sources to generate a unified semantic representation; The archival data is semantically labeled using a multi-label classification model, and a semantic graph of the archival content is constructed using a graph database to represent the relationships between archives. A semantic index tree is then built based on the semantic graph to optimize the rapid location and retrieval of archival content. Record changes for each version of the archive, locate the changes based on the semantic index tree, identify semantic changes in the archive through a semantic difference comparison algorithm and record the hash, and divide the archive content into granular parts and generate user access policies through the semantic boundaries of nodes at each level in the index tree. Record the semantic path of user access profiles and construct user behavior graphs. Identify potential abnormal behaviors through behavior graph analysis, and use time series analysis and fuzzy mining algorithms for risk prediction and behavior warning. The construction of the semantic index tree based on the semantic graph includes: constructing the Label-GAT model, which specifically includes defining the set of labels required for file management, assigning an embedding vector to each label, and using the joint representation vector A and the label vector together as nodes to establish a heterogeneous graph structure; Graph attention propagation is performed on heterogeneous graph structures, and each labeled node will obtain a semantic representation related to the current sample. The joint representation vector A is split into three sub-vectors: image, text, and structured field, which serve as three modal information sources. By calculating attention weights, the matching degree of the label under different modalities is determined, and the final fused label representation is obtained. The fused representation of all labels is input into a shared fully connected layer, and the prediction result is output through the Sigmoid function to train the Label-GAT model. For any new file, the corresponding joint representation vector is obtained and input into the trained Label-GAT model. After graph attention propagation and multimodal weighted interaction, a label probability value is output. According to the set threshold, the label with the probability value greater than the threshold is determined as the classification result of the file. Based on the classification result set, the basic entity types in the graph are determined, semantic edge relationships are established between entities in the graph, all entities and their semantic connections are uniformly organized into triples in the graph structure, the graph structure is standardized, the initial graph structure is constructed, and semantic reasoning rules are defined. The semantic graph structure is standardized, and a knowledge graph embedding model is used to train the graph, mapping each entity and relation to a unified low-dimensional vector space. A density-based clustering algorithm is used to cluster all archive entities. Based on the clustering results, the cluster centers and their corresponding entities are divided into a hierarchical semantic index tree. The upper-level nodes of the tree represent semantic cluster centers, and each center node is connected to its subordinate entity nodes or sub-cluster nodes. By controlling the number and depth of branches at each level, a multi-level index tree is constructed.

2. The intelligent monitoring and management method based on archive digitization as described in claim 1, characterized in that: The use of a multimodal BERT model to fuse features from different data sources and generate a unified semantic representation refers to collecting multi-source archival data, including image data, text data, and structured data, and preprocessing the collected data. The preprocessed image is uniformly divided into image blocks of fixed size. Based on the spatial position of the image block in the whole image, a corresponding positional information representation is assigned to each image block, thereby generating an image block feature matrix containing local image content and spatial position information. The image block sequence is modeled using a multi-layer coding structure based on an attention mechanism to extract semantic relevance and global feature expression in the image modality. The preprocessed text data is decomposed into a series of words or sub-word units. The BERT model is used to map the segmented text into vector representations. Each token obtains a fixed-dimensional embedding vector, which forms the initial text feature matrix. A deep language encoder is used to model the text sequence to obtain text feature representations containing contextual semantics. After encoding and mapping image feature sequences and text feature sequences to a unified semantic space, information interaction is achieved through a cross-modal attention mechanism, and finally they are aggregated into an image-text fusion feature vector F. The preprocessed structured fields are concatenated into a structured feature vector. The preprocessed structured feature vector is then input into the hybrid modeling path. The optimal modeling strategy is selected for different data types of the fields. The feature results are concatenated and fused, and then input into the multilayer perceptron encoder to obtain the semantic embedding vector Q of the structured fields. The image-text fusion feature vector F and the structured semantic embedding vector Q are aligned in terms of vector dimensions and then concatenated to construct a joint representation vector A.

3. The intelligent monitoring and management method based on archive digitization as described in claim 2, characterized in that: The system records changes for each version of the archive, locates the change location based on the semantic index tree, identifies semantic changes in the archive through a semantic difference comparison algorithm, and records hash values. When a modified, rewritten, or updated version of an archive is detected, the constructed semantic index tree is immediately invoked, the changed content is embedded with vector encoding, and semantic similarity is calculated with the embedding vector of the path node where the original archive is located in the index tree. The closest path node of the change point in the index tree is determined, and the change anchor point of the archive's semantic location is located. Extract the current modified content and the previous version of the file. Use a unified semantic encoding model to compare the distance between the two versions in the sentence vector or paragraph vector space. Combine the changes in the tag path of key fields to calculate the overall semantic difference. Generate a structured snapshot of the current modified content and perform full content hash calculation.

4. The intelligent monitoring and management method based on archive digitization as described in claim 3, characterized in that: The process of granularly dividing file content and generating user access strategies by using the semantic boundaries of nodes at each level in the index tree refers to treating nodes at each level in the index tree as semantic classification boundaries. Through similarity analysis of semantic vectors between nodes, semantic breakpoints are identified and semantic segments at different levels are divided. Each file version is bound to the corresponding semantic level in the corresponding index path. Permission vectors are generated based on the user's organization, role, and access scenario, and compared with each semantic segment. A similarity threshold is used to determine whether the user has access rights.

5. The intelligent monitoring and management method based on archive digitization as described in claim 4, characterized in that: The process of identifying potential abnormal behaviors through behavioral graph analysis and using time-series analysis and fuzzy mining algorithms for risk prediction and behavioral early warning involves continuously constructing a user-centric behavioral graph by recording the semantic path nodes, access times, and file versions traversed when a user accesses a file. This graph models and analyzes the user's behavioral characteristics in each semantic category, including access frequency, path jump patterns, and access depth, identifying behavioral preferences and routine patterns. When a user's behavior exhibits characteristics such as access paths or sudden changes in access density that differ from historical patterns, a time-series model is constructed based on the behavioral graph. This model is then combined with fuzzy mining algorithms to identify potential abnormal behavioral patterns and generate corresponding behavioral early warning processes.

6. The intelligent monitoring and management method based on archive digitization as described in claim 5, characterized in that: The behavior warning process includes immediately freezing the user's current access permissions or forcibly downgrading them to read-only mode, automatically generating structured alarm information, pushing it to the administrator interface, recording the complete behavior chain, and appending risk events to the on-chain log.

7. An intelligent monitoring and management system based on digitized archives, based on the intelligent monitoring and management method based on digitized archives as described in any one of claims 1 to 6, characterized in that: include, The semantic fusion module is used to collect multi-source archival data and perform structured cleaning, and generate a unified semantic representation using the multimodal BERT model; The graph construction module is used to semantically label archive content through a multi-label classification model and build a graph database to represent the semantic relationships between archives. The semantic index building module is used to build a semantic index tree based on the semantic graph, enabling rapid file location and recording the path and hash information of each version change and semantic change. The permission policy module is used to divide the file content granularity according to the semantic boundaries at each level in the index tree and generate fine-grained user access permission policies. The risk warning module is used to record user access behavior to build a behavior graph, and combine time series analysis and fuzzy mining to identify abnormal behavior and generate risk warning signals.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the intelligent monitoring and management method based on archive digitization as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent monitoring and management method based on archive digitization as described in any one of claims 1 to 6.

Images