An information-based accounting archive management method and system

Abstract

The present invention discloses an information-based accounting archive management method and system, which relates to the technical field of archive management, including constructing a double-layer intelligent analysis structure, realizing intelligent identification and classification of accounting archives, and adopting a deep learning model for feature extraction and fusion; constructing a subject relationship network based on temporal similarity, establishing an accounting entry association model through a variety of matrices, and forming a basic intelligent framework; based on the intelligent framework, constructing a multi-channel feature matrix and performing tensor projection operations to generate a multi-dimensional subgraph structure; constructing a double-layer knowledge graph based on the subject relationship network, and realizing automatic discovery of implicit associations between archives through inter-layer mapping; based on the double-layer knowledge graph, constructing a risk monitoring matrix to realize risk warning, integrating a multi-dimensional retrieval strategy for association analysis, and storing archives in a hierarchical manner. The present invention breaks the information island limitation of traditional archive management, establishes a collaborative sharing mechanism for archives, and greatly improves the efficiency of archive management.

Description

Technical Field

[0001] The present invention relates to the technical field of file management, and in particular to an information-based accounting file management method and system. Background Art

[0002] With the development of information technology and the digital transformation of accounting work, the digital management of accounting archives has become an inevitable trend for modern enterprises. The current accounting archive management system mainly adopts the traditional database management method, which realizes the storage and retrieval of archives by establishing archive directory indexes and classification codes. Although this management method has achieved the basic electronicization of archives, it has exposed many limitations in the face of increasingly complex business scenarios and continuously growing data scale. Traditional systems cannot effectively capture the deep correlation between archives and lack the ability to understand the semantics of archive content, resulting in low retrieval efficiency and insufficient accuracy.

[0003] At the same time, existing technologies only focus on static archive storage management, fail to establish a dynamic risk monitoring mechanism, and fail to timely identify and warn of potential business risks. In addition, existing systems generally have information island problems, lack effective data sharing and collaborative analysis mechanisms between business modules, and are difficult to meet the all-round needs of modern enterprises for accounting archive management. Especially in the big data environment, the traditional linear management model can no longer cope with the multi-dimensional and multi-level archive management challenges, and an intelligent and systematic management method is urgently needed to improve the management efficiency and risk prevention and control level of accounting archives. Summary of the invention

[0004] In view of the problems existing in the existing information-based accounting archive management methods, the present invention proposes an information-based accounting archive management method and system.

[0005] Therefore, the problem to be solved by the present invention is the technical problems of insufficient archive association analysis capability, lack of risk warning mechanism and low information utilization efficiency in the existing accounting archive management system.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In the first aspect, an embodiment of the present invention provides an information-based accounting archive management method, which includes: constructing a two-layer intelligent analysis structure to realize intelligent identification and classification of accounting archives, the first layer uses convolutional neural networks and bidirectional long short-term memory networks to extract features of images and texts, and realizes feature fusion through a multi-scale time series-aware attention mechanism; the second layer constructs a subject relationship network based on time series similarity, establishes an accounting entry association model through an association matrix, an adjacency matrix and a centrality vector, and forms a basic intelligent framework; based on the intelligent framework, a multi-channel feature matrix is ​​constructed, and a tensor projection operation is performed on the multi-channel feature matrix and the node centrality vector to generate a multi-dimensional subgraph structure; based on the subject relationship network, a two-layer knowledge graph containing bottom-level entity nodes and upper-level rule nodes is constructed, and automatic discovery of implicit associations between archives is realized through inter-layer mapping and graph completion; based on the two-layer knowledge graph, a real-time risk monitoring matrix is ​​constructed to calculate a risk warning index to realize real-time monitoring and warning of risks, integrate multi-dimensional retrieval strategies, and perform association analysis on the retrieval results, and hierarchically store archives according to the analysis results.

[0008] As a preferred solution of the information-based accounting archive management method of the present invention, the feature fusion is based on the weight coefficient Perform weighted fusion on the first eigenvector and the second eigenvector to generate a target eigenvector containing time series information ; The first feature vector is obtained by obtaining the image data and text data of the accounting file to be processed, and extracting the features of the image data through a preset convolutional neural network model; the second feature vector is obtained by extracting the features of the text data through a preset recurrent neural network model; a time series feature matrix is ​​constructed based on the timestamp information of the accounting file , the time series feature matrix Include day dimension , Monthly Dimension and year dimension Three time scales; Calculating the weight coefficient based on the three time scales , expressed as:

[0009] ;

[0010] in, is the eigenvector weight coefficient, is the basic weight coefficient, , , are the time intervals of day, month and year respectively. , , is the corresponding time decay factor, , , is the time scale weight, is a feature vector, including a first feature vector and a second feature vector; is the trainable parameter matrix, is the bias vector; the weighted fusion adopts an adaptive dynamic adjustment mechanism, and the adjustment rule is: if the time series feature matrix is ​​detected The time interval , , are smaller than the attenuation factor of the corresponding time scale , , , then the basic weight coefficient It increases according to the ratio of the time interval to the attenuation factor; if the time series feature matrix is ​​detected There are time scales where the time interval is greater than the corresponding decay factor , then the weight coefficient corresponding to this time scale is The time scale weights are redistributed according to the excess ratio and the reduction value is redistributed according to the weight ratio of the remaining time scales; if the first eigenvector and the second eigenvector are The difference in activation values ​​after function mapping is less than the parameter matrix The minimum eigenvalue of , then the weight coefficient is calculated according to the time series feature matrix The exponential decay term value of is adjusted proportionally.

[0011] As a preferred solution of the information-based accounting archive management method of the present invention, the archive association matrix is ​​constructed based on the target feature vector. :Use the improved time series similarity calculation method to calculate any two target feature vectors and The correlation between , expressed as:

[0012] ;

[0013] in, represents the cosine similarity of feature vectors, Indicates the time interval, is the time decay coefficient; based on the archive association matrix Build a subject relationship network , where the vertex set Represents accounting subjects, edge sets Represents the association relationship between subjects, and the edge weight is the corresponding association value; based on the archive association matrix Setting Thresholds Constructing the network adjacency matrix ; Based on the network adjacency matrix Describe the subject relationship network The connection relationship between the nodes in the network is calculated, and the centrality vector is constructed by calculating the centrality value of the node to describe the importance of the node; calculate each node The centrality value of is expressed as:

[0014] ;

[0015] in, is the total number of nodes, For Node The centrality value of each node Ordered by node number Dimensional vector, constructing node centrality vector ; Integrate network structure information and node centrality values ​​into the association model In the association model The parameters include: correlation matrix , network adjacency matrix and the node centrality vector ; By associating the model Conduct correlation analysis on accounting entries to form a basic intelligent framework.

[0016] As a preferred solution of the information-based accounting archive management method described in the present invention, the two-layer knowledge graph includes a bottom-level graph and an upper-level graph, the entity nodes of the bottom-level graph are instantiated expressions of the subject relationship network nodes, and contain specific accounting archive entity information; the rule nodes of the upper-level graph are rule abstractions of the bottom-level entity nodes, reflecting business rules; the entity nodes of the bottom-level graph represent specific accounting archives, the edge weights between nodes are determined by the network adjacency matrix, and the attributes of the edges include the time dimension subgraph , spatial dimension subgraph and associated dimension subgraphs The corresponding association strength in the upper layer graph represents the business rules, the node importance is determined by the node centrality vector, and the connection relationship between the upper layer nodes is expressed as follows: when the number of connections between the underlying entity nodes corresponding to the two business rule nodes exceeds the preset threshold When an edge connection is established between these two rule nodes, the edge weight The calculation formula is expressed as:

[0017] ;

[0018] in, Represents the rule node in the upper graph and The edge weights between ; Represents a rule node and The number of connections between the corresponding underlying entity nodes; Represents a rule node and The geometric mean of the respective node centrality values; Indicates the maximum number of connections between all regular node pairs; represents the maximum centrality geometric mean of all regular node pairs, and is the weight coefficient.

[0019] As a preferred solution of the information-based accounting archive management method of the present invention, a multi-channel feature matrix is ​​constructed based on the parameters of the association model. , the accounting period identifier, business department code and business type number are mapped to different feature channels respectively, and the value distribution of each channel is standardized; the first channel maps the accounting period identifier, the second channel maps the business department code, and the third channel maps the business type number. The correlation matrix is ​​decomposed according to the time dimension, space dimension and correlation dimension. The time dimension calculates the first channel eigenvalue based on the temporal relationship of the accounting period identifier, the space dimension calculates the second channel eigenvalue based on the organizational hierarchy relationship of the business department code, and the correlation dimension calculates the third channel eigenvalue based on the classification system of the business type number; for the multi-channel feature matrix Perform tensor projection operation with the node centrality vector to generate a time dimension subgraph based on the first channel eigenvalue , Generate a spatial dimension subgraph based on the second channel eigenvalue , Generate the associated dimension subgraph based on the third channel eigenvalue .

[0020] As a preferred solution of the information-based accounting archive management method of the present invention, the tensor projection operation adopts a matrix decomposition method, multiplying the node centrality vector as a weight coefficient with the corresponding channel eigenvalue; the time dimension subgraph The edge weights in the graph use the temporal order of the accounting period identifier to calculate the association strength, and the spatial dimension subgraph The edge weights in the calculation use the direct and indirect superior-subordinate relationships coded by the business department to calculate the association strength, and the association dimension subgraph The edge weight in is calculated by the category to which the business type number belongs and the relationship between the associated categories; if the time dimension difference between the two accounting files is within the specified accounting period, the time association strength is set to 1, otherwise it is calculated according to the inverse of the time interval; if the business departments of the two accounting files have a direct superior-subordinate relationship, the spatial association strength is set to 1, if there is an indirect superior-subordinate relationship, it is set to 0.5, otherwise it is set to 0; if the business types of the two accounting files belong to the same category, the association dimension strength is set to 1, if they belong to a related category, it is set to 0.5, otherwise it is set to 0; the inter-layer mapping strength is jointly determined by the association strength of the three dimensions.

[0021] As a preferred solution of the information-based accounting archive management method described in the present invention, the row vectors of the risk monitoring matrix correspond to the characteristics of the underlying entity nodes, and the column vectors correspond to the constraints of the upper-level rule nodes; the edge weights in the time dimension subgraph, the space dimension subgraph and the association dimension subgraph are respectively mapped to the corresponding elements of the risk monitoring matrix, and when the edge weight of a certain dimension is lower than the characteristic fluctuation threshold of the dimension, the corresponding matrix element is marked as a risk item; based on the risk monitoring matrix Calculate the risk diffusion coefficient ; When there is a risk item, if the risk item satisfies the following conditions: the weight of the risk item has decreased for three consecutive accounting periods, or the abnormal frequency of the department to which it belongs is higher than the threshold, or the number of associated business nodes exceeds the preset value, the risk item will be marked as a propagation source node; if the propagation source node satisfies the condition that the node centrality is higher than the network average, the business type to which it belongs is a core type, and there is a direct connection with other propagation source nodes, the first-order neighbor nodes of the propagation source node will be included in the risk monitoring scope; if the set of nodes within the risk monitoring scope satisfies the condition that a complete business closed loop is formed between the nodes, or the node clustering coefficient exceeds the preset threshold, the node set will be marked as a risk propagation link; if the risk propagation link satisfies the condition that the link length is greater than the preset minimum value, the average weight of the nodes on the link deviates significantly, and the link spans multiple business domains, the risk level of all nodes on the link will be reassessed; if there are no new risk nodes, or the number of iterations reaches the upper limit, or the risk diffusion coefficient tends to be stable, the current round of risk propagation calculation will be terminated.

[0022] As a preferred solution of the information-based accounting archive management method of the present invention, based on the risk monitoring matrix and the risk diffusion coefficient , calculate the risk warning index ; The risk warning index Including risk components in time dimension, space dimension and association dimension; the calculation of risk components is through the weighted combination of edge weights and risk diffusion coefficients in the corresponding dimension subgraph; setting a sliding time window to predict the risk warning index Real-time tracking will be continuous Risk warning index sequence within a time window Input the trend prediction model to predict the risk warning index for the next time window , based on the confidence interval of historical distribution, the risk warning index It is divided into four intervals. The risk warning index reaches the maximum in the first interval, and decreases in other intervals in turn. The trend prediction model adopts the exponential smoothing method. If the time series shows drastic short-term fluctuations, obvious long-term trends, and significant seasonal characteristics, triple exponential smoothing is adopted. In the calculation process, the smoothing value is calculated once, the secondary value is calculated based on the primary value, and the tertiary value is calculated based on the secondary value to obtain the smoothing results of each stage.

[0023] As a preferred solution of the information-based accounting archive management method of the present invention, the following is a method for constructing a multidimensional search engine by integrating the features of the time dimension subgraph, the space dimension subgraph and the association dimension subgraph. , multidimensional search engine The following search modes are supported: Time-based search: Enter the accounting period interval , returns the relevant archive nodes within the time interval; association-based retrieval: input the target node set, and return the archive nodes that are directly or indirectly connected to the target node in the three-dimensional subgraph; feature-based retrieval: input the feature vector, and return the archive node with the highest similarity to the query vector in the feature space; for multidimensional retrieval engines If the archive nodes in the retrieval results have rule associations in the two-layer knowledge graph, the set of these associated nodes is recorded as ; By analyzing the associated node set The distribution characteristics in sub-graphs of different dimensions are used to identify potential business models, and the identified business models are fed back to the rule nodes of the upper-level graph to dynamically update the knowledge graph structure.

[0024] In the second aspect, an embodiment of the present invention provides an information-based accounting archive management system, which includes: an identification and classification module, which realizes intelligent identification and classification of accounting archives by constructing a two-layer intelligent analysis structure, and extracts features of images and texts by using convolutional neural networks and bidirectional long short-term memory networks, and realizes feature fusion through a multi-scale time-series-aware attention mechanism; constructs a subject relationship network based on time-series similarity, and establishes an accounting entry association model through an association matrix, an adjacency matrix and a centrality vector to form a basic intelligent framework; an association analysis module, based on the intelligent framework, constructs a multi-channel feature matrix, and performs tensor projection operations on the multi-channel feature matrix and the node centrality vector to generate a multi-dimensional subgraph structure; constructs a two-layer knowledge graph containing underlying entity nodes and upper-layer rule nodes based on the subject relationship network, and realizes automatic discovery of implicit associations between archives through inter-layer mapping and graph completion; an early warning classification module, based on the two-layer knowledge graph, constructs a real-time risk monitoring matrix to calculate a risk early warning index, realizes real-time monitoring and early warning of risks, integrates multi-dimensional retrieval strategies, and performs association analysis on the retrieval results, and stores archives in a hierarchical manner according to the analysis results.

[0025] The beneficial effects of the present invention are that the present invention realizes the deep mining and visual expression of the complex correlation relationship between accounting archives by constructing a two-layer architecture, thereby improving the intelligent level of archive management; through the multi-dimensional feature analysis mechanism, the archive information is comprehensively analyzed from the three dimensions of time, space and association, thereby enhancing the utilization value of archive data; and with the help of a dynamic risk monitoring and early warning system, the real-time identification and hierarchical management of potential risks are realized, significantly improving the risk prevention and control capabilities. The present invention breaks the information island limitation of traditional archive management, establishes a collaborative sharing mechanism for archive information, and greatly improves the efficiency of archive management and data analysis capabilities. At the same time, the hierarchical storage strategy ensures the security of archives, provides a new solution for the intelligent management of corporate accounting archives, and has significant practical value and promotion significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them:

[0027] Figure 1 A flowchart of the information-based accounting records management method;

[0028] Figure 2 It is an intelligent framework diagram of the information-based accounting archive management method;

[0029] Figure 3A two-layer knowledge graph structure for information-based accounting archive management methods;

[0030] Figure 4 A hierarchical storage strategy for information-based accounting archive management methods. DETAILED DESCRIPTION

[0031] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0032] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0033] 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 term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0034] Example 1, reference Figure 1~Figure 4 , which is the first embodiment of the present invention, and provides an information-based accounting archive management method, including:

[0035] S1: Construct a two-layer intelligent analysis structure to realize intelligent recognition and classification of accounting files. The first layer uses convolutional neural networks and bidirectional long short-term memory networks to extract features of images and texts, and realizes feature fusion and automatic classification through multi-scale time series-aware attention mechanism; the second layer constructs a subject relationship network based on time series similarity, and establishes an accounting entry association model through association matrix, adjacency matrix and centrality vector to form a basic intelligent framework.

[0036] Acquire image data and text data of the accounting file to be processed, and extract features of the image data through a preset convolutional neural network model to obtain a first feature vector;

[0037] The text data is subjected to feature extraction by a preset recurrent neural network model to obtain a second feature vector;

[0038] Among them, the preset convolutional neural network model is a convolutional neural network model containing three convolutional layers and two fully connected layers. The convolution kernel size of the first convolutional layer is 3×3, the convolution step size is 1, and the number of output channels is 64. The convolution kernel size of the second convolutional layer is 3×3, the convolution step size is 1, and the number of output channels is 128. The convolution kernel size of the third convolutional layer is 3×3, the convolution step size is 1, and the number of output channels is 256. Each convolutional layer is followed by a maximum pooling layer and a ReLU activation function. The output dimensions of the two fully connected layers are 512 and preset dimensions respectively. , and get the first eigenvector ;

[0039] The preset recurrent neural network model is a recurrent neural network model that includes an embedding layer and two layers of bidirectional long short-term memory networks. The word vector dimension of the embedding layer is 300, and the number of hidden units of the two layers of bidirectional long short-term memory networks is 256. The output is compressed to the preset dimension through the temporal pooling layer. , and get the second eigenvector .

[0040] Based on the first eigenvector and the second eigenvector , construct a multi-scale temporal perception attention weight mechanism, and calculate the weight coefficients of the first eigenvector and the second eigenvector;

[0041] Constructing a time series feature matrix based on timestamp information of accounting archives , the time series feature matrix Include day dimension , Monthly Dimension and year dimension Three time scales, based on which the weight coefficients are calculated, are expressed as:

[0042] ;

[0043] in, is the feature vector weight, is the basic weight coefficient, , , are the time intervals of day, month and year respectively. , , is the corresponding time decay factor, , , is the time scale weight, is a feature vector, including a first feature vector and a second feature vector; is the trainable parameter matrix, is the bias vector.

[0044] According to the weight coefficient Perform weighted fusion on the first eigenvector and the second eigenvector to generate a target eigenvector containing time series information .

[0045] The weighted fusion adopts an adaptive dynamic adjustment mechanism, and the specific adjustment rules are as follows:

[0046] If the time series feature matrix is ​​detected The time interval , , are smaller than the attenuation factor of the corresponding time scale , , , then the basic weight coefficient Increase according to the ratio of the time interval to the attenuation factor;

[0047] If the time series feature matrix is ​​detected There are time scales where the time interval is greater than the corresponding decay factor , then the weight coefficient corresponding to this time scale is Decrease the time scale weights in proportion to the excess, and redistribute the reduced values ​​in proportion to the weights of the remaining time scales;

[0048] If the first eigenvector and the second eigenvector are The difference in activation values ​​after function mapping is less than the parameter matrix The minimum eigenvalue of , then the weight coefficient is calculated according to the time series feature matrix The exponential decay term value of is adjusted proportionally.

[0049] The target feature vector is input into the pre-trained subject classification model, and the accounting files are classified by the support vector machine algorithm to obtain the accounting subject labels. The subject classification model is constructed using the conventional support vector machine training method, including:

[0050] Perform data cleaning and standardization on historical accounting data, and build a training sample set based on subject label information },in represents the target feature vector of the first sample, Indicates the corresponding subject label; Indicates The target feature vector of samples, Indicates the corresponding subject label;

[0051] Select the optimal kernel function type from linear kernel function, polynomial kernel function and radial basis kernel function through cross-validation method;

[0052] Grid search method is used to determine penalty parameters and kernel function parameters , where the value range of the penalty parameter is adaptively divided according to the sample variance, and the kernel function parameter The value range of is determined based on the feature dimension;

[0053] Solve the dual problem using the sequential minimal optimization algorithm to obtain the Lagrange multiplier , The training samples corresponding to values ​​greater than zero are support vectors;

[0054] Based on Lagrange multipliers And the support vector calculates the normal vector of the classification hyperplane and the bias term , construct a classification decision function;

[0055] The model performance is evaluated by the holdout method, and when the classification accuracy is lower than the preset threshold, the parameter optimization step is repeated;

[0056] After the support vector machine model training is completed, the normal vector obtained by training is , bias term And the kernel function parameters are stored in the parameter matrix middle;

[0057] Through the parameter matrix The newly input accounting files are classified by subject, and the feature vectors of the new files are substituted into the trained support vector machine model to achieve automatic subject classification.

[0058] The archive association matrix is ​​constructed based on the target feature vector, including: using an improved temporal similarity calculation method to calculate any two target feature vectors and The correlation between It is expressed as:

[0059] ;

[0060] in, represents the cosine similarity of feature vectors, Indicates the time interval, is the time decay coefficient.

[0061] The correlation values ​​between all the accounting files to be processed are composed of The correlation matrix ; Based on the correlation matrix Build a subject relationship network , where the vertex set Represents accounting subjects, edge sets It represents the association relationship between subjects, and the edge weight is the corresponding association value.

[0062] Furthermore, the network adjacency matrix is ​​used to describe the connection relationship between nodes in the subject relationship network, and the centrality value of the node is calculated to construct a centrality vector to describe the importance of the node.

[0063] Specifically, according to the correlation matrix Setting Thresholds Constructing the network adjacency matrix , specifically, the network adjacency matrix for A binary matrix with elements Representation Node and nodes Is there a connection relationship? ≥ When , the corresponding adjacency matrix element =1, otherwise =0.

[0064] Subject Relationship Network Each node in Calculate its centrality value, expressed as:

[0065] ;

[0066] in, is the total number of nodes, For Node The centrality value of .

[0067] The centrality value of each node Ordered by node number Dimensional vector, constructing node centrality vector .

[0068] Integrate network structure information and node centrality values ​​into association models In the association model The parameters include: correlation matrix , network adjacency matrix and the node centrality vector .

[0069] By associating the model Conduct correlation analysis on accounting entries and recommend relevant historical files and accounting subjects for newly input accounting files.

[0070] S2: Based on the intelligent framework, a multi-dimensional accounting archive association network and knowledge graph are constructed. A multi-channel feature matrix is ​​constructed through the three dimensions of time dimension, space dimension and association dimension. The multi-channel feature matrix and the node centrality vector are subjected to tensor projection operation to generate a subgraph structure of time dimension, space dimension and association dimension. A two-layer knowledge graph containing bottom-level entity nodes and upper-level rule nodes is constructed based on the subject relationship network. The automatic discovery of implicit associations between archives is realized through inter-layer mapping and graph completion.

[0071] Constructing a multi-channel feature matrix based on correlation model parameters , the accounting period identifier, business department code and business type number are mapped to different feature channels respectively, and the value distribution of each channel is standardized.

[0072] Among them, the first channel maps the accounting period identifier, the second channel maps the business department code, and the third channel maps the business type number. The correlation matrix is ​​decomposed according to the time dimension, space dimension, and correlation dimension. The time dimension calculates the first channel characteristic value based on the temporal relationship of the accounting period identifier, the space dimension calculates the second channel characteristic value based on the organizational hierarchy relationship of the business department code, and the correlation dimension calculates the third channel characteristic value based on the classification system of the business type number.

[0073] The eigenvalue is obtained by calculating the row and column projection of the correlation matrix, and when the corresponding element of the network adjacency matrix is ​​1, the corresponding eigenvalue is kept unchanged, and when the corresponding element of the network adjacency matrix is ​​0, the corresponding eigenvalue is set to zero;

[0074] If the value fluctuation in a certain feature channel exceeds the feature fluctuation threshold, the dynamic sampling mechanism is triggered: the feature value of the channel is resampled according to the exponential step size until the value distribution tends to be stable; if periodic fluctuations in the feature value are detected, the sampling interval is adjusted to 1 / 100 of the period. , is the preset sampling factor.

[0075] For multi-channel feature matrix Perform tensor projection operation with the node centrality vector to generate a time dimension subgraph based on the first channel eigenvalue , Generate a spatial dimension subgraph based on the second channel eigenvalue , Generate the associated dimension subgraph based on the third channel eigenvalue .

[0076] The tensor projection operation uses matrix decomposition to multiply the node centrality vector with the corresponding channel eigenvalue as the weight coefficient.

[0077] Time dimension subgraph The edge weights in the graph use the temporal order of the accounting period identifier to calculate the association strength, and the spatial dimension subgraph The edge weights in the calculation use the direct and indirect superior-subordinate relationships coded by the business department to calculate the association strength, and the association dimension subgraph The edge weight in is calculated by using the category to which the business type number belongs and the relationship between the associated categories to calculate the association strength.

[0078] If the time dimension difference between two accounting files is within the specified accounting period, the time association strength is set to 1, otherwise it is calculated according to the inverse of the time interval; if the business departments of the two accounting files have a direct superior-subordinate relationship, the spatial association strength is set to 1, if there is an indirect superior-subordinate relationship, it is set to 0.5, otherwise it is set to 0; if the business types of the two accounting files belong to the same category, the association dimension strength is set to 1, if they belong to related categories, it is set to 0.5, otherwise it is set to 0; the inter-layer mapping strength is jointly determined by the association strength of the three dimensions.

[0079] Based on the expansion of the nodes in the subject relationship network, a two-layer knowledge graph structure is constructed, including the bottom graph and the upper graph. The entity nodes of the bottom graph are instantiations of the subject relationship network nodes, containing specific accounting file entity information; the rule nodes of the upper graph are rule abstractions of the bottom entity nodes, reflecting business rules;

[0080] Specifically, the bottom-level nodes represent specific accounting files, the edge weights between nodes are determined by the network adjacency matrix, and the attributes of the edges include the time dimension subgraph , spatial dimension subgraph and associated dimension subgraphs The corresponding association strength in ;

[0081] The upper-level nodes represent business rules. The node importance is determined by the node centrality vector. The connection relationship between the upper-level nodes is expressed as follows: when the number of connections between the underlying entity nodes corresponding to the two business rule nodes exceeds the preset threshold When an edge connection is established between these two rule nodes, the edge weight The calculation formula is expressed as:

[0082] ;

[0083] in, Represents the rule node in the upper graph and The edge weights between ; Represents a rule node and The number of connections between the corresponding underlying entity nodes; Represents a rule node and The geometric mean of the respective node centrality values; Indicates the maximum number of connections between all regular node pairs; represents the maximum centrality geometric mean of all regular node pairs, and is the weight coefficient.

[0084] Inter-layer mapping strength The calculation formula is:

[0085] ;

[0086] in, Represents the underlying entity node With the upper rule node The mapping strength between Represents the underlying entity node The sum of the association strengths in the three-dimensional subgraphs; Represents the underlying entity node With the upper rule node The cosine similarity of the feature vectors; is a balance parameter.

[0087] Based on the structure of the subject relationship network, the knowledge graph is completed: when it is detected that two entity nodes are connected in the subject relationship network, but there is no direct connection in the current graph, the comprehensive association strength between the entity nodes is calculated according to the edge weights in the subject relationship network and the time dimension feature vectors of the entity nodes, the corresponding connection relationship is added to the underlying graph, and the association relationship strength between the corresponding upper-level rule nodes is updated to optimize the overall structure of the hierarchical knowledge graph.

[0088] The two-layer knowledge graph is dynamically maintained and updated, the changing trend of the graph structure is monitored through a sliding time window, and the weights of nodes and edges are adjusted in a timely manner.

[0089] Specifically, if the correlation between the feature vector of the newly added accounting file and the existing node meets the requirements of the business rules, it will be included in the corresponding local subgraph; if the correlation strength of a node continues to weaken within a continuous time window, the weight coefficient of the node will be reduced; if a new stable business association pattern is found, the corresponding rule node will be added to the upper-level graph.

[0090] Among them, "comprehensive correlation strength" refers to the weighted average of the correlation strengths of the three dimensions, and the weight coefficient is pre-set according to the business scenario; "local subgraph" refers to a subgraph structure centered on a core node and consisting of nodes reachable within a preset step size; "stable business correlation pattern" refers to a node relationship pattern that appears repeatedly in multiple consecutive time windows and whose correlation strength remains stable.

[0091] S3: Based on the two-layer knowledge graph, a real-time risk monitoring matrix is ​​constructed to calculate the risk warning index, realize real-time monitoring and warning of risks, integrate multi-dimensional retrieval strategies, and perform correlation analysis on the retrieval results, and store the archives in a hierarchical manner according to the analysis results.

[0092] Constructing a real-time risk monitoring matrix based on a two-layer knowledge graph structure , risk monitoring matrix The row vectors correspond to the features of the underlying entity nodes, and the column vectors correspond to the constraints of the upper-level rule nodes; the edge weights in the time dimension subgraph, space dimension subgraph, and association dimension subgraph are mapped to the risk monitoring matrix When the edge weight of a dimension is lower than the characteristic fluctuation threshold of the dimension, the corresponding matrix element is marked as a risk item.

[0093] Based on the risk monitoring matrix Calculate the risk diffusion coefficient ; When there is a risk item, if the risk item meets the requirement of weight decline for three consecutive accounting periods, or the abnormal frequency of the department to which it belongs is higher than the threshold, or the number of associated business nodes exceeds the preset value, the risk item will be marked as a propagation source node;

[0094] If the node centrality of the propagation source node is higher than the network average, and its business type is a core type, and there is a direct connection with other propagation source nodes, the first-order neighbor nodes of the propagation source node will be included in the risk monitoring scope.

[0095] If the node set within the risk monitoring range satisfies the requirement to form a complete business closed loop between nodes, or the node clustering coefficient exceeds the preset threshold, the node set will be marked as a risk propagation link.

[0096] If the risk propagation link satisfies the condition that the link length is greater than the preset minimum value, the average weight of the nodes on the link deviates significantly, and the link spans multiple business domains, the risk level of all nodes on the link will be reassessed.

[0097] If there are no new risk nodes, or the number of iterations reaches the upper limit, or the risk diffusion coefficient tends to be stable, the current round of risk propagation calculation is terminated.

[0098] Based on the risk monitoring matrix and the risk diffusion coefficient , calculate the risk warning index , risk warning index It includes risk components in time dimension, space dimension and association dimension; the risk components are calculated through the weighted combination of edge weights in the corresponding dimension subgraph and the risk diffusion coefficient.

[0099] If all the elements in a row or column in the risk monitoring matrix rise and the diagonal elements change, it is recorded as an initial warning signal, the risk source corresponding to the abnormal row is locked, the impact domain corresponding to the abnormal column is locked, the diagonal element change area is monitored in detail, and an associated monitoring group for the abnormal elements is established.

[0100] If the diffusion coefficient between nodes in the risk diffusion coefficient increases, and the diffusion range continues to expand and the diffusion speed accelerates, the warning level will be raised, the high diffusion coefficient nodes will be isolated, the main diffusion path will be cut off, a diffusion buffer area will be set up, and a diffusion inhibition mechanism will be constructed.

[0101] If the matrix structure changes during the monitoring process and multiple monitoring points change at the same time, a comprehensive warning will be triggered, the key connections of the closed-loop path will be cut off, the multi-point monitoring will be decomposed into independent processing units, and the matrix reconstruction process will be started.

[0102] If multiple types of warnings are triggered at the same time, the warning priorities should be clearly ranked, an interaction mechanism between warnings should be established, various processing resources should be coordinated, and the processing schedule should be coordinated.

[0103] If the matrix state continues to be abnormal after the warning is triggered, the diffusion effect is not weakened, and the warning points are interrelated, then the warning is confirmed to be valid.

[0104] Set a sliding time window to predict the risk Real-time tracking will be continuous Risk warning index sequence within a time window Input the trend prediction model to predict the risk warning index for the next time window , based on the confidence interval of historical distribution, the risk warning index It is divided into four intervals. The risk warning index reaches the maximum in the first interval and decreases in other intervals.

[0105] The trend prediction model uses the exponential smoothing method. If the time series shows sharp short-term fluctuations, obvious long-term trends, and significant seasonal characteristics, triple exponential smoothing will be used.

[0106] During the calculation process, the primary smoothing value is calculated first, the secondary value is calculated based on the primary value, and the tertiary value is calculated based on the secondary value to obtain the smoothing results of each stage.

[0107] The calculation of a smoothing value is expressed as:

[0108] ;

[0109] in, for A smoothed value at time, for The actual observed value at time, for A smoothed value at time, is the smoothing coefficient.

[0110] Integrate the features of time dimension subgraph, space dimension subgraph and association dimension subgraph to build a multidimensional search engine , multidimensional search engine The following search modes are supported:

[0111] Time-based search: Enter the accounting period interval , returns the relevant archive nodes within the time interval;

[0112] Association-based retrieval: input the target node set and return the archive nodes that are directly or indirectly connected to the target nodes in the three-dimensional subgraph;

[0113] Feature-based retrieval: Input a feature vector and return the archive node with the highest similarity to the query vector in the feature space.

[0114] The search results are sorted according to the matching degree with the query conditions and the risk warning index. Sort by weighted scores.

[0115] Multidimensional Search Engine If the archive nodes in the retrieval results have rule associations in the two-layer knowledge graph, the set of these associated nodes is recorded as ; By analyzing the associated node set The distribution characteristics in sub-graphs of different dimensions are used to identify potential business models, and the identified business models are fed back to the rule nodes of the upper-level graph to dynamically update the knowledge graph structure.

[0116] If the results of the association analysis show that the risk level of the archive node is determined, the business model characteristics are obvious, and the association relationship is stable, the archives will be stored in a graded manner according to the risk level.

[0117] Furthermore, the risk level of the archives is divided into four levels based on the interval range of the risk warning index and the risk diffusion coefficient. The first level of risk warning index is in the first interval, the risk diffusion coefficient reaches the maximum diffusion range, and the warning level is the highest. The archives need to be independently encrypted and stored, with multiple backup mechanisms and strict access control.

[0118] If the risk warning index is in the second interval and the risk diffusion coefficient reaches the secondary diffusion range, the risk level is at the second level, and the files need to be encrypted and stored, double backed up, and hierarchical access control set up.

[0119] If the risk warning index is in the third interval and the risk diffusion coefficient reaches the basic diffusion range, the risk level is at the third level. The files need to be encrypted regularly, backed up individually, and basic access control needs to be set.

[0120] If the risk warning index is in the fourth interval and the risk diffusion coefficient reaches the minimum diffusion range, the risk level is at the fourth level, and the files need to be stored normally, backed up regularly, and set up standard access control.

[0121] If the archive correlation analysis finds that there is a business chain involving multiple links that requires tracing and verification, an archive traceability index will be established.

[0122] If new business models are identified during the archive processing, the knowledge graph structure will be updated, new management rules will be formed, and the archive archiving process will be optimized.

[0123] In summary, the present invention achieves in-depth mining and visual expression of complex correlations between accounting archives by constructing a two-layer architecture, thereby improving the intelligent level of archive management; through a multi-dimensional feature analysis mechanism, archival information is comprehensively analyzed from the three dimensions of time, space and correlation, thereby enhancing the utilization value of archival data; and with the help of a dynamic risk monitoring and early warning system, real-time identification and hierarchical management of potential risks are achieved, significantly improving risk prevention and control capabilities. The present invention breaks the information island limitations of traditional archive management, establishes a collaborative sharing mechanism for archival information, and greatly improves archive management efficiency and data analysis capabilities. At the same time, the hierarchical storage strategy ensures the security of archives, provides a new solution for the intelligent management of corporate accounting archives, and has significant practical value and promotion significance.

[0124] This embodiment further provides an information-based accounting archive management system, including:

[0125] The recognition and classification module realizes intelligent recognition and classification of accounting files by building a two-layer intelligent analysis structure. The convolutional neural network and the bidirectional long short-term memory network are used to extract the features of images and texts, and feature fusion is realized through the multi-scale time series perception attention mechanism. The subject relationship network is constructed based on the time series similarity, and the accounting entry association model is established through the association matrix, adjacency matrix and centrality vector to form a basic intelligent framework.

[0126] The association analysis module builds a multi-channel feature matrix based on an intelligent framework, and performs tensor projection operations on the multi-channel feature matrix and the node centrality vector to generate a multi-dimensional subgraph structure; a two-layer knowledge graph containing bottom-level entity nodes and upper-level rule nodes is constructed based on the subject relationship network, and the implicit associations between archives are automatically discovered through inter-layer mapping and graph completion;

[0127] The early warning classification module, based on a two-layer knowledge graph, constructs a real-time risk monitoring matrix to calculate the risk early warning index, realizes real-time monitoring and early warning of risks, integrates multi-dimensional retrieval strategies, and performs correlation analysis on the retrieval results, and stores the archives in a hierarchical manner according to the analysis results.

[0128] This embodiment also provides a computer device suitable for the information-based accounting archive management method, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the information-based accounting archive management method proposed in the above embodiment.

[0129] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0130] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method for realizing information-based accounting archive management as proposed in the above embodiment is implemented.

[0131] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiment belong to the same inventive concept. The technical details not fully described in this embodiment can be found in the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0132] Example 2: This example provides an information-based accounting archive management method. In order to verify the beneficial effects of the present invention, a scientific demonstration is carried out through simulation experiments.

[0133] In order to verify the effectiveness of the present invention, the accounting file data of a large enterprise from January to December 2023 were selected for simulation experiments. The experimental data set contains 50,000 files of various types such as invoices, vouchers, and reports, involving 300 business units and 2,000 related items.

[0134] The 50,000 archive data were randomly divided into training and test sets in a ratio of 7:3.

[0135] The knowledge graph model of the present invention is trained using the training set data, the archive association network is constructed, the risk warning rules are set, and the system parameters are optimized.

[0136] Use test set data to conduct system testing and verify performance indicators, including: correlation analysis performance testing and risk warning performance testing.

[0137] 1,000 files were randomly selected from the test set as query samples to conduct association analysis performance tests. The test results showed that the accuracy of association path recognition reached 92.8%, which is a significant improvement over the 65.3% of the traditional system.

[0138] On average, 25 valid association relationships were successfully established for each file, the maximum depth of association analysis could reach 6 layers, the accuracy rate of cross-departmental association identification reached 90.5%, and the accuracy rate of cross-time association identification reached 88.7%.

[0139] The risk warning performance test was conducted for 90 consecutive days in the test set. The test results showed that the system issued 85 risk warnings in total, of which 76 were accurate warnings with an accuracy rate of 89.3%. The average warning was 3.5 working days earlier than the traditional method.

[0140] The level distribution of risk warning is as follows:

[0141] Level 1 (high risk) files: approximately 320 (2.1%);

[0142] Level 2 (higher risk) files: approximately 1,180 (7.9%);

[0143] Level 3 (medium risk) files: approximately 3,850 (25.7%);

[0144] Level 4 (low risk) files: approximately 9,650 (64.3%).

[0145] Comparison of test data shows that the present invention has significant technical advantages in intelligent archive management, with significantly improved correlation analysis capabilities, and can accurately identify and restore complex business correlation relationships; the risk warning mechanism is efficient and reliable, and can detect and warn of potential risks in advance, providing strong support for enterprise risk management.

[0146] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An information-based accounting archive management method, characterized in that: include: A two-layer intelligent analysis structure is constructed to realize intelligent recognition and classification of accounting files. The first layer uses convolutional neural networks and bidirectional long short-term memory networks to extract features of images and texts, and realizes feature fusion through a multi-scale time-series-aware attention mechanism. The second layer builds a subject relationship network based on time series similarity, and establishes an accounting entry association model through the association matrix, adjacency matrix and centrality vector to form a basic intelligent framework; Based on the intelligent framework, a multi-channel feature matrix is ​​constructed, and a tensor projection operation is performed on the multi-channel feature matrix and the node centrality vector to generate a multi-dimensional subgraph structure; Based on the subject relationship network, a two-layer knowledge graph including bottom-level entity nodes and upper-level rule nodes is constructed, and the implicit associations between archives are automatically discovered through inter-layer mapping and graph completion; Based on the two-layer knowledge graph, a real-time risk monitoring matrix is ​​constructed to calculate the risk warning index, so as to realize real-time monitoring and warning of risks, integrate multi-dimensional retrieval strategies, and perform correlation analysis on the retrieval results, and hierarchically store the archives according to the analysis results; The two-layer knowledge graph includes a bottom-layer graph and an upper-layer graph. The entity nodes of the bottom-layer graph are instantiations of the subject relationship network nodes, including specific accounting file entity information; the rule nodes of the upper-layer graph are rule abstractions of the bottom-layer entity nodes, reflecting business rules. The entity nodes of the underlying graph represent specific accounting files. The edge weights between nodes are determined by the network adjacency matrix. The attributes of the edges include the time dimension subgraph. , spatial dimension subgraph and associated dimension subgraphs The corresponding association strength in ; The rule nodes of the upper-level graph represent business rules, and the node importance is determined by the node centrality vector. The connection relationship between the upper-level nodes is expressed as follows: when the number of connections between the underlying entity nodes corresponding to the two business rule nodes exceeds the preset threshold When an edge connection is established between these two rule nodes, the edge weight The calculation formula is expressed as: ； in, Represents the rule node in the upper graph and The edge weights between Represents a rule node and The number of connections between the corresponding underlying entity nodes; Represents a rule node and The geometric mean of the respective node centrality values; Indicates the maximum number of connections between all regular node pairs; represents the maximum centrality geometric mean of all regular node pairs, and is the weight coefficient.

2. The information-based accounting archive management method according to claim 1, characterized in that: The feature fusion is based on the weight coefficient Perform weighted fusion on the first eigenvector and the second eigenvector to generate a target eigenvector containing time series information ; The first feature vector is obtained by obtaining image data and text data of the accounting file to be processed, and extracting features from the image data through a preset convolutional neural network model; The second feature vector is obtained by extracting features from the text data through a preset recurrent neural network model; Constructing a time series feature matrix based on timestamp information of accounting archives , the time series feature matrix Include day dimension , Monthly Dimension and year dimension Three time scales; The weight coefficients are calculated based on three time scales , expressed as: ； in, is the eigenvector weight coefficient, is the basic weight coefficient, , , are the time intervals of day, month and year respectively. , , is the corresponding time decay factor, , , is the time scale weight, is a feature vector, including a first feature vector and a second feature vector; is the trainable parameter matrix, is the bias vector; The weighted fusion adopts an adaptive dynamic adjustment mechanism, and the adjustment rule is: If the time series feature matrix is ​​detected The time interval , , are smaller than the decay factor of the corresponding time scale , , , then the basic weight coefficient Increase according to the ratio of the time interval to the attenuation factor; If the time series feature matrix is ​​detected There are time scales where the time interval is greater than the corresponding decay factor , then the weight coefficient corresponding to this time scale is Decrease the time scale weights in proportion to the excess, and redistribute the reduced values ​​in proportion to the weights of the remaining time scales; If the first eigenvector and the second eigenvector are The difference in activation values ​​after function mapping is less than the parameter matrix The minimum eigenvalue of , then the weight coefficient is calculated according to the time series feature matrix The exponential decay term value of is adjusted proportionally.

3. The information-based accounting archive management method according to claim 2, characterized in that: Constructing a profile association matrix based on the target feature vector : The improved time series similarity calculation method is used to calculate the feature vectors of any two targets. and The correlation between , expressed as: ； in, represents the cosine similarity of feature vectors, Indicates the time interval, is the time attenuation coefficient; Based on the archive relevance matrix Build a subject relationship network , where the vertex set Represents accounting subjects, edge sets It represents the association relationship between subjects, and the edge weight is the corresponding association value; Based on the archive relevance matrix Setting Thresholds Constructing the network adjacency matrix ; Based on the network adjacency matrix Describe the subject relationship network The connection relationship between nodes in the network is calculated, and the centrality vector is constructed by calculating the centrality value of the node to describe the importance of the node; Calculate each node The centrality value of is expressed as: ； in, is the total number of nodes, For Node The centrality value of The centrality value of each node Ordered by node number Dimensional vector, constructing node centrality vector ; Integrate network structure information and node centrality values ​​into association models In the association model The parameters include: correlation matrix , network adjacency matrix and the node centrality vector ; By associating the model Conduct correlation analysis on accounting entries to form a basic intelligent framework.

4. The information-based accounting archive management method according to claim 3 is characterized in that: Construct a multi-channel feature matrix based on the parameters of the association model , the accounting period identifier, business department code and business type number are mapped to different feature channels respectively, and the value distribution of each channel is standardized; The first channel maps the accounting period identifier, the second channel maps the business department code, and the third channel maps the business type number. The correlation matrix is ​​decomposed according to the time dimension, space dimension, and correlation dimension. The time dimension calculates the first channel eigenvalue based on the temporal relationship of the accounting period identifier, the space dimension calculates the second channel eigenvalue based on the organizational hierarchy relationship of the business department code, and the correlation dimension calculates the third channel eigenvalue based on the classification system of the business type number. For multi-channel feature matrix Perform tensor projection operation with the node centrality vector to generate a time dimension subgraph based on the first channel eigenvalue , Generate a spatial dimension subgraph based on the second channel eigenvalue , Generate the associated dimension subgraph based on the third channel eigenvalue .

5. The information-based accounting archive management method according to claim 4, characterized in that: The tensor projection operation adopts a matrix decomposition method, and multiplies the node centrality vector as a weight coefficient with the corresponding channel eigenvalue; Time dimension subgraph The edge weights in the graph use the temporal order of the accounting period identifier to calculate the association strength, and the spatial dimension subgraph The edge weights in the calculation use the direct and indirect superior-subordinate relationships coded by the business department to calculate the association strength, and the association dimension subgraph The edge weight in the calculation uses the relationship between the category to which the business type number belongs and the associated category to calculate the association strength; If the time dimension difference between two accounting files is within the specified accounting period, the time association strength is set to 1, otherwise it is calculated according to the inverse of the time interval; if the business departments of the two accounting files have a direct superior-subordinate relationship, the spatial association strength is set to 1, if there is an indirect superior-subordinate relationship, it is set to 0.5, otherwise it is set to 0; if the business types of the two accounting files belong to the same category, the association dimension strength is set to 1, if they belong to related categories, it is set to 0.5, otherwise it is set to 0; the inter-layer mapping strength is jointly determined by the association strength of the three dimensions.

6. The information-based accounting archive management method according to claim 5, characterized in that: The row vectors of the risk monitoring matrix correspond to the features of the underlying entity nodes, and the column vectors correspond to the constraints of the upper rule nodes; The edge weights in the time dimension subgraph, space dimension subgraph and association dimension subgraph are mapped to the corresponding elements of the risk monitoring matrix respectively. When the edge weight of a certain dimension is lower than the characteristic fluctuation threshold of the dimension, the corresponding matrix element is marked as a risk item. Based on the risk monitoring matrix Calculate the risk diffusion coefficient ; When there is a risk item, if the risk item meets the requirement of weight decline for three consecutive accounting periods, or the abnormal frequency of the department to which it belongs is higher than the threshold, or the number of associated business nodes exceeds the preset value, the risk item will be marked as a propagation source node; If the node centrality of the transmission source node is higher than the network average, the business type is a core type, and there is a direct connection with other transmission source nodes, the first-order neighbor nodes of the transmission source node will be included in the risk monitoring scope; If the node set within the risk monitoring range satisfies the requirement to form a complete business closed loop between nodes, or the node clustering coefficient exceeds the preset threshold, the node set will be marked as a risk propagation link; If the risk propagation link meets the requirement that the link length is greater than the preset minimum value, the average weight of the nodes on the link deviates significantly, and the link spans multiple business domains, the risk level of all nodes on the link will be reassessed; If there are no new risk nodes, or the number of iterations reaches the upper limit, or the risk diffusion coefficient tends to be stable, the current round of risk propagation calculation is terminated.

7. The information-based accounting archive management method according to claim 6, characterized in that: Based on the risk monitoring matrix and the risk diffusion coefficient , calculate the risk warning index ; The risk warning index Including risk components in time dimension, space dimension and association dimension; the calculation of risk components is a weighted combination of edge weights in the corresponding dimension subgraph and risk diffusion coefficient; Set a sliding time window to predict the risk Real-time tracking will be continuous Risk warning index sequence within a time window Input the trend prediction model to predict the risk warning index for the next time window , based on the confidence interval of historical distribution, the risk warning index It is divided into four intervals. The risk warning index reaches the maximum in the first interval, and decreases in other intervals; The trend prediction model adopts the exponential smoothing method. If the time series shows sharp short-term fluctuations, obvious long-term trends, and significant seasonal characteristics, triple exponential smoothing is used; During the calculation process, the primary smoothing value is calculated first, the secondary value is calculated based on the primary value, and the tertiary value is calculated based on the secondary value to obtain the smoothing results of each stage.

8. The information-based accounting archive management method according to claim 7, characterized in that: Integrate the features of time dimension subgraph, space dimension subgraph and association dimension subgraph to build a multidimensional search engine , multidimensional search engine The following search modes are supported: Time-based search: Enter the accounting period interval , returns the relevant archive nodes within the time interval; Association-based retrieval: input the target node set and return the archive nodes that are directly or indirectly connected to the target nodes in the three-dimensional subgraph; Feature-based retrieval: Input a feature vector and return the archive node with the highest similarity to the query vector in the feature space; Multidimensional Search Engine If the archive nodes in the retrieval results have rule associations in the two-layer knowledge graph, the set of these associated nodes is recorded as ; By analyzing the associated node set The distribution characteristics in sub-graphs of different dimensions are used to identify potential business models, and the identified business models are fed back to the rule nodes of the upper-level graph to dynamically update the knowledge graph structure.

9. An information-based accounting archive management system, based on the information-based accounting archive management method according to any one of claims 1 to 8, characterized in that: include: The recognition and classification module realizes intelligent recognition and classification of accounting files by building a two-layer intelligent analysis structure. The convolutional neural network and the bidirectional long short-term memory network are used to extract the features of images and texts, and feature fusion is realized through the multi-scale time series perception attention mechanism. The subject relationship network is constructed based on the time series similarity, and the accounting entry association model is established through the association matrix, adjacency matrix and centrality vector to form a basic intelligent framework. The association analysis module builds a multi-channel feature matrix based on an intelligent framework, and performs tensor projection operations on the multi-channel feature matrix and the node centrality vector to generate a multi-dimensional subgraph structure; a two-layer knowledge graph containing bottom-level entity nodes and upper-level rule nodes is constructed based on the subject relationship network, and the implicit associations between archives are automatically discovered through inter-layer mapping and graph completion; The early warning classification module, based on a two-layer knowledge graph, constructs a real-time risk monitoring matrix to calculate the risk early warning index, realizes real-time monitoring and early warning of risks, integrates multi-dimensional retrieval strategies, and performs correlation analysis on the retrieval results, and stores the archives in a hierarchical manner according to the analysis results.

Images