Real-time synchronization method for personnel in cross-organization approval system based on public and private keys

Abstract

This invention provides a real-time synchronization method for cross-organizational approval system personnel based on public and private keys. This method relates to the field of information encryption and synchronization technology. The method involves generating unique public and private key pairs, constructing a multi-level hash tree of active nodes, forming a multidimensional association matrix based on distribution feature maps, performing dynamic programming to obtain optimal propagation paths, and utilizing network slicing for consistency checking and fault location. This method ensures the security, real-time performance, and consistency of cross-organizational personnel information synchronization, improves the overall system response efficiency, and enhances the reliability of data transmission.

Description

Technical Field

[0001] The present invention relates to the technical field of information encryption and synchronization, and in particular to a real-time synchronization method for personnel in a cross-organizational approval system based on public and private keys. Background Art

[0002] With increasingly close collaboration between organizations, cross-organizational approval systems have become a crucial tool for efficient collaboration across enterprises. In multi-organizational collaboration scenarios, the approval process involves the exchange of personnel information across different organizations, requiring the system to synchronize personnel changes across organizations in real time to ensure continuity and data consistency in the approval process. Traditional cross-organizational personnel information synchronization relies primarily on periodic batch synchronization or simple message push mechanisms, which have limitations in terms of data security, real-time performance, and handling of complex organizational structures.

[0003] Existing cross-organizational personnel information synchronization technology still has shortcomings, lacks effective identity authentication and data integrity protection mechanisms, and data tampering or identity impersonation are prone to occur during cross-organizational data transmission, making it difficult to ensure the authenticity and security of information synchronization; existing technologies mostly adopt static synchronization strategies, which are unable to dynamically adjust synchronization strategies according to changes in organizational structure and personnel activity, resulting in waste of system resources and low synchronization efficiency, and cannot meet the real-time synchronization needs of large-scale organizations; traditional synchronization methods lack effective error detection and fault recovery mechanisms. When network fluctuations or system anomalies cause synchronization failure, it is difficult to accurately locate the faulty node and automatically repair it, resulting in the persistence of data inconsistency and affecting the normal operation of cross-organizational business processes.

[0004] As organizations expand and their businesses become more complex, there is an urgent need for a cross-organizational personnel information synchronization method that is cryptographically secured, driven by intelligent prediction, and has self-healing capabilities to meet the security, real-time, and reliability requirements of modern enterprise collaboration. Summary of the Invention

[0005] The embodiment of the present invention provides a method for real-time synchronization of personnel in a cross-organizational approval system based on public and private keys, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for real-time synchronization of personnel in a cross-organizational approval system based on public and private keys, comprising:

[0007] Receive personnel information synchronization requests from organizations and obtain information request queues;

[0008] Generate a unique public-private key pair for each organization based on the information request queue, distribute the public key to other organizations, and encrypt and store the private key locally;

[0009] Obtaining personnel information from the information request queue to construct a node set, and using a long short-term memory neural network to predict and mark active nodes; constructing a multi-level hash tree based on active nodes, introducing a hierarchical perturbation factor to write to the hierarchical cache pool, calculating the data checksum and submitting it to the main storage pool, and constructing a distribution feature map based on the similarity of node hash values;

[0010] The received personnel information change content is digitally signed using the private key to generate a change request. A multi-dimensional association matrix is ​​constructed based on the distribution feature map. A directed propagation graph is constructed through feature decomposition. Dynamic programming using multi-dimensional attention is performed to obtain the optimal propagation path. A synchronization sequence is generated, and data integrity is verified through the main storage pool.

[0011] Use the public key to verify the digital signature of the change request, update the node information according to the change request synchronization sequence, divide the network slices based on the directed propagation graph, perform consistency checks based on the node hash values ​​in the hierarchical cache pool within the network slices, generate error detection values ​​based on the data check code of the main storage pool between network slices, and build an anomaly propagation graph based on the error detection value to locate the fault source node and correct the status information.

[0012] In an optional embodiment, obtaining personnel information from an information request queue to construct a node set, calculating node feature values, and predicting and marking active nodes using a long short-term memory neural network include:

[0013] Extract personnel identification information and organization identification information from the information request queue, construct a set of personnel information nodes, obtain the time window sequence of each personnel information node, calculate the number of personnel information changes in each time window to obtain the information change frequency, and calculate the time series feature value based on the time decay function of the time window and the information change frequency;

[0014] Obtain the organizational hierarchy difference of each personnel information node to obtain the organizational distance, count the number of information interactions between personnel information nodes to obtain the information interaction intensity, and calculate the correlation eigenvalue based on the organizational distance and information interaction intensity;

[0015] The time series eigenvalues ​​and the associated eigenvalues ​​are used to construct a feature vector, which is input into the pre-trained long short-term memory neural network to obtain the hidden state sequence. The attention mechanism is applied to the hidden state sequence to calculate the time series weight. The hidden state sequence is weighted according to the time series weight to obtain the predicted state. The predicted state is mapped through a fully connected layer to obtain a change trend prediction curve. The second-order derivative of the change trend prediction curve is calculated to determine the inflection point set. The dynamic threshold function is determined based on the inflection point set. The time series eigenvalues ​​and the associated eigenvalues ​​are input into the dynamic threshold function, and the personnel information nodes greater than the corresponding threshold are marked as active nodes.

[0016] In an optional embodiment, a multi-level hash tree is constructed based on active nodes, a hierarchical perturbation factor is introduced and the data is written into a hierarchical cache pool, a data checksum is calculated and submitted to a main storage pool, and a distribution feature map is constructed based on the similarity of node hash values, including:

[0017] Active nodes are used as leaf nodes of a multi-level hash tree. The total number of layers is determined based on the number of active nodes. The identification information and feature information of the active nodes are combined to calculate the leaf-level hash value. The leaf-level hash value is constructed as the first-level hash value. The first-level hash values ​​are combined in pairs to construct the second-level hash value. The hash values ​​of each layer are constructed upwards in sequence.

[0018] Based on the preset perturbation base and the layer number, the perturbation factor of each layer is determined. The hash value of the first layer node is calculated by combining the first layer hash value and the first layer perturbation factor. Starting from the second layer, the hash value of the lower layer node corresponding to each parent node is combined with the perturbation factor of this layer to obtain the hash value of the node in this layer.

[0019] Construct a hierarchical cache pool and determine the capacity of each level in the hierarchical cache pool based on the number of active nodes; write the hash value of each layer node to the corresponding level; generate a cache pool mask with the same length as the number of hierarchical cache pool levels; perform bitwise operations on the hash value of the top node and the hierarchical cache pool mask to obtain a data check code;

[0020] The similarity of hash values ​​between active nodes is calculated, and the active nodes are regarded as vertex sets, and the pairs of active nodes with similarity greater than the preset similarity threshold are regarded as edge sets to construct the distribution feature map of active nodes.

[0021] In an optional embodiment, a change request is generated by digitally signing the received personnel information change content using a private key, a multidimensional correlation matrix is ​​constructed based on the spatiotemporal correlation data of the distribution feature map, a directed propagation graph is constructed through feature decomposition, a dynamic programming algorithm based on multidimensional attention is executed to obtain the optimal propagation path, a synchronization sequence is generated, and a data check code is obtained from the main storage pool to verify data integrity, including:

[0022] Generate a digital signature for the received personnel information change using a private key, and combine the personnel information change, the digital signature, and a timestamp to generate a change request;

[0023] Calculate the spatiotemporal correlation data of the nodes in the distribution feature map, multiply and accumulate the spatiotemporal correlation data with the time weight that decays exponentially with the time window, and generate a multidimensional correlation matrix corresponding to the node feature dimension;

[0024] Performing eigenvalue decomposition on the multidimensional association matrix to obtain eigenvectors, selecting the eigenvectors corresponding to the maximum preset number of eigenvalues ​​to construct a low-dimensional representation vector, calculating the inner product of the low-dimensional representation vectors of different nodes, and normalizing them to obtain a node affinity matrix;

[0025] The node affinities in the node affinity matrix that are greater than a preset affinity threshold are set as edge weights to construct a directed propagation graph;

[0026] A dynamic programming algorithm based on multi-dimensional attention is executed in the directed propagation graph to calculate the propagation path of the change request. The path with the largest cumulative weight is obtained through multi-layer feature fusion and weight backtracking, and the optimal propagation path is determined.

[0027] The node identifiers and inter-node propagation delays in the optimal propagation path are combined to generate a change request synchronization sequence, a hash operation is performed on the change request synchronization sequence, and the sequence is compared with the data check code in the primary storage pool to verify data integrity.

[0028] In an optional embodiment, a dynamic programming algorithm based on multi-dimensional attention is executed in a directed propagation graph to calculate the propagation path of the change request, and a path with the maximum cumulative weight is obtained through multi-layer feature fusion and weight backtracking. Determining the optimal propagation path includes:

[0029] The propagation weights between nodes in the directed propagation graph are combined with the weighted historical synchronization delays to generate an initial propagation weight matrix;

[0030] Calculate the multi-dimensional attention value between nodes in the propagation weight matrix, normalize the multi-dimensional attention value to obtain the node propagation priority, and use the weighted product of the node propagation priority and the historical synchronization success rate as the state transition weight to generate the state transition matrix;

[0031] Based on the multi-channel structure, the state transition matrix is ​​transformed and weighted fusion is performed to obtain the cumulative propagation weight between nodes. The state transition probability is calculated based on the ratio of the cumulative propagation weight to the node processing capacity, and the optimal state transition sequence is constructed.

[0032] The historical synchronization features between nodes on the propagation path are used to extract deep features through a multi-layer skip connection network, and the deep features are combined with the accumulated propagation weights to obtain the path evaluation value;

[0033] The optimal number of propagation hops is selected based on the path evaluation value, and the prediction model is trained using the differential strategy optimization method. The prediction model is used to calculate the optimal preceding node of the node;

[0034] Starting from the target node, the optimal preceding nodes are connected sequentially until the source node, generating the optimal propagation path with the maximum cumulative propagation weight.

[0035] In an optional embodiment, using a public key to verify the digital signature of a change request, updating node information according to a change request synchronization sequence, dividing the network into slices based on a directed propagation graph, performing a consistency check based on node hash values ​​in a hierarchical cache pool within the network slice, generating an error detection value based on a data check code in a primary storage pool between network slices, constructing an anomaly propagation graph based on the error detection value to locate the fault source node and correct the status information includes:

[0036] Receive change requests, verify digital signatures using public keys, and update information in nodes according to the change request synchronization sequence;

[0037] Calculate the ratio of the sum of the node neighbor weights in the directed propagation graph to the global weight, aggregate the nodes that meet the preset slice threshold conditions into network slices, and establish cross-slice connection relationships between network slices;

[0038] In the network slice, the node status, timestamp and previous node identifier are concatenated to obtain a node feature sequence. A weighted hash operation is performed on the node feature sequence to generate a multidimensional hash value. The ratio of the intersection cardinality to the union cardinality of the multidimensional hash values ​​between nodes is calculated as the consistency measure.

[0039] Generate a check code for the network slice data, and perform an XOR operation on the check code and the check code in the main storage pool to obtain an error detection value;

[0040] Extract abnormal nodes whose error detection values ​​exceed the preset fault tolerance threshold. Use the intersection nodes of abnormal nodes and change request propagation paths as vertices of the abnormal propagation graph. Build an edge weight matrix based on delay correlation and construct the abnormal propagation graph through eigendecomposition.

[0041] Calculate the weighted out-degree sum of the nodes in the anomaly propagation graph and select the node with the largest weighted out-degree sum as the fault source node;

[0042] Correct the status information of the fault source node based on the checksum in the primary storage pool and the node backup data.

[0043] In an optional embodiment, abnormal nodes whose error detection values ​​exceed a preset fault tolerance threshold are extracted, intersection nodes of the abnormal nodes and the change request propagation path are used as vertices of an abnormal propagation graph, an edge weight matrix is ​​constructed based on delay correlation, and constructing the abnormal propagation graph through eigendecomposition includes:

[0044] Obtain the error detection value sequence of the node, and mark the node whose difference between the current error detection value and the expected error detection value exceeds the preset fault tolerance threshold as an abnormal node;

[0045] Construct a time-decay weight, multiply it with the error detection value sequence, and normalize it to obtain the abnormality degree of the abnormal node;

[0046] Obtaining the propagation path of the change request, and calculating the intersection node of the abnormal node and the propagation path as the vertex of the abnormal propagation graph;

[0047] Calculate the propagation delay between vertices, substitute it into the exponential decay function to get the delay correlation, and count the number of propagations between vertices to get the propagation probability;

[0048] The delay correlation, propagation probability and consistency measure are combined according to the preset weight coefficient to construct the edge weight matrix;

[0049] Perform eigendecomposition on the edge weight matrix to obtain eigenvalues ​​and eigenvectors, and select the product of the square root of the eigenvalue and the eigenvector to construct the node propagation eigenvector;

[0050] The node association strength is obtained by calculating the inner product of the node propagation feature vector and the modulus ratio, and the vertex pairs with node association strength greater than a preset strength threshold are taken as edges, and the consistency measure is used as the edge weight to construct the abnormal propagation graph.

[0051] A second aspect of an embodiment of the present invention provides a real-time synchronization system for cross-organizational approval system personnel based on public and private keys, including:

[0052] The first unit is used to receive personnel information synchronization requests from organizations and obtain information request queues;

[0053] The second unit is used to generate a unique public-private key pair for each organization based on the information request queue, distribute the public key to other organizations, and encrypt and store the private key locally;

[0054] The third unit is used to obtain personnel information from the information request queue to construct a node set, predict and mark active nodes through a long short-term memory neural network; construct a multi-level hash tree based on active nodes, introduce a hierarchical perturbation factor to write to the hierarchical cache pool, calculate the data check code and submit it to the main storage pool, and construct a distribution feature map based on the similarity of node hash values;

[0055] The fourth unit is used to digitally sign the received personnel information change content using the private key to generate a change request, construct a multidimensional association matrix based on the distribution feature map, construct a directed propagation graph through feature decomposition, perform dynamic programming with multidimensional attention to obtain the optimal propagation path, generate a synchronization sequence, and verify data integrity through the main storage pool;

[0056] The fifth unit is used to use the public key to verify the digital signature of the change request, update the node information according to the change request synchronization sequence, divide the network slices based on the directed propagation graph, perform consistency checks within the network slices based on the node hash values ​​in the hierarchical cache pool, generate error detection values ​​between network slices based on the data check code of the main storage pool, and construct an abnormal propagation graph based on the error detection value to locate the fault source node and correct the status information.

[0057] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0062] In an embodiment of the present invention, a public-private key mechanism is used to achieve secure exchange of cross-organizational personnel information, ensuring the confidentiality and integrity of data transmission. At the same time, digital signatures are used to ensure the credibility of information sources, effectively preventing unauthorized access and data tampering, and improving the security of information interaction between systems. A long short-term memory neural network is introduced to predict active nodes and construct a multi-level hash tree. Combined with a two-tier storage architecture of a hierarchical cache pool and a main storage pool, efficient management of massive personnel information is achieved, data query latency is significantly reduced, and system response speed and processing capabilities are improved. An optimization algorithm based on a distribution feature map and a multi-dimensional correlation matrix is ​​used, through directed propagation graphs and network slicing technology, accurate synchronization of personnel information and fault location are achieved, significantly improving the accuracy and stability of cross-organizational personnel information synchronization and reducing system maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flowchart of a method for real-time synchronization of personnel in a cross-organizational approval system based on public and private keys according to an embodiment of the present invention;

[0064] Figure 2 Build a schematic diagram for a multi-level hash tree;

[0065] Figure 3 A schematic diagram of the system architecture for synchronous sequence generation of change requests based on multi-dimensional attention dynamic programming;

[0066] Figure 4 Schematic diagram of the comparison of propagation path performance based on the multi-dimensional attention dynamic programming algorithm. DETAILED DESCRIPTION

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0068] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0069] Figure 1 Schematic diagram of the process of the real-time synchronization method of cross-organization approval system personnel based on public and private keys according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0070] Receive personnel information synchronization requests from organizations and obtain information request queues;

[0071] Generate a unique public-private key pair for each organization based on the information request queue, distribute the public key to other organizations, and encrypt and store the private key locally;

[0072] Obtaining personnel information from the information request queue to construct a node set, and using a long short-term memory neural network to predict and mark active nodes; constructing a multi-level hash tree based on active nodes, introducing a hierarchical perturbation factor to write to the hierarchical cache pool, calculating the data checksum and submitting it to the main storage pool, and constructing a distribution feature map based on the similarity of node hash values;

[0073] The received personnel information change content is digitally signed using the private key to generate a change request. A multi-dimensional association matrix is ​​constructed based on the distribution feature map. A directed propagation graph is constructed through feature decomposition. Dynamic programming using multi-dimensional attention is performed to obtain the optimal propagation path. A synchronization sequence is generated, and data integrity is verified through the main storage pool.

[0074] Use the public key to verify the digital signature of the change request, update the node information according to the change request synchronization sequence, divide the network slices based on the directed propagation graph, perform consistency checks based on the node hash values ​​in the hierarchical cache pool within the network slices, generate error detection values ​​based on the data check code of the main storage pool between network slices, and build an anomaly propagation graph based on the error detection value to locate the fault source node and correct the status information.

[0075] In an optional embodiment, obtaining personnel information from an information request queue to construct a node set, calculating node feature values, and predicting and marking active nodes using a long short-term memory neural network include:

[0076] Extract personnel identification information and organization identification information from the information request queue, construct a set of personnel information nodes, obtain the time window sequence of each personnel information node, calculate the number of personnel information changes in each time window to obtain the information change frequency, and calculate the time series feature value based on the time decay function of the time window and the information change frequency;

[0077] Obtain the organizational hierarchy difference of each personnel information node to obtain the organizational distance, count the number of information interactions between personnel information nodes to obtain the information interaction intensity, and calculate the correlation eigenvalue based on the organizational distance and information interaction intensity;

[0078] The time series eigenvalues ​​and the associated eigenvalues ​​are used to construct a feature vector, which is input into the pre-trained long short-term memory neural network to obtain the hidden state sequence. The attention mechanism is applied to the hidden state sequence to calculate the time series weight. The hidden state sequence is weighted according to the time series weight to obtain the predicted state. The predicted state is mapped through a fully connected layer to obtain a change trend prediction curve. The second-order derivative of the change trend prediction curve is calculated to determine the inflection point set. The dynamic threshold function is determined based on the inflection point set. The time series eigenvalues ​​and the associated eigenvalues ​​are input into the dynamic threshold function, and the personnel information nodes greater than the corresponding threshold are marked as active nodes.

[0079] In a specific embodiment, personnel identification information and organization identification information are extracted from the information request queue to construct a personnel information node set. The information request queue stores personnel information request data submitted by various organizations. Each request contains fields such as the request timestamp, request type, personnel identification code, and organization identification code. The personnel identification code uses a 32-bit unique identifier, and the organization identification code uses a 16-bit hierarchical encoding. For example, a certain request data contains: "2024-06-01T14:30:25", "Information update", "P2024060100125", "ORG2022A135" and other information. By parsing the data in the request queue, all personnel identification information and corresponding organization identification information are extracted, and a personnel information node set N = {n1, n2, ..., nm} is constructed, where each node ni contains the personnel identification pi and the organization identification oi.

[0080] Obtain the time window sequence for each person information node. For each node ni, divide the timeline into k consecutive time windows W = {w1, w2, ..., wk} based on its historical information request records. Each time window can be set to 7 days in length, with no overlap between adjacent windows. For example, for node n1, its time window sequence is {"2024-05-01 to 2024-05-07", "2024-05-08 to 2024-05-14", ...}.

[0081] The information change frequency is calculated by counting the number of personnel information changes for each time window. The number of changes refers to the total number of requests for personnel information changes within a specific time window. The information change frequency, f, is calculated by dividing the number of changes by the length of the time window. For example, if node n1 receives 15 information change requests during the time window from May 1, 2024, to May 7, 2024, its information change frequency, f, equals 15 / 7, or 2.14 times per day.

[0082] The time series eigenvalue is calculated based on the time decay function and the information change frequency of the time window. The time decay function represents the influence of time distance on the current prediction, with the change frequency closer to the current time having a greater impact. Time decay is implemented using an exponentially decreasing method. For historical data within t windows from the current time, the weight is an exponentially decreasing value. The time series eigenvalue of each node is calculated as the weighted sum of the information change frequency in each time window and the corresponding time decay weight. For example, for node n1, assuming the information change frequencies in the last three time windows were 2.14, 1.57, and 0.86 times per day, respectively, and the corresponding time decay weights are 0.9, 0.81, and 0.729, the time series eigenvalue is 2.14 × 0.9 + 1.57 × 0.81 + 0.86 × 0.729 = 3.71.

[0083] Obtain the organizational level difference for each personnel information node to obtain the organizational distance. The organizational distance is determined by calculating the distance between two organizations in the organizational hierarchy tree. For example, the organizational level difference between "ORG2022A135" and "ORG2022A246" is 2, indicating that two parent-level jumps are required to reach the common parent node.

[0084] The information interaction intensity is calculated by counting the number of information interactions between human information nodes. The number of information interactions refers to the number of related requests that appear in the information request queue between two human nodes. The information interaction intensity is calculated by dividing the number of information interactions by the length of the statistical period. For example, if nodes n1 and n2 had 45 information interactions in the past 30 days, the information interaction intensity is 45 / 30 = 1.5 times / day.

[0085] The association eigenvalue is calculated based on organizational distance and information interaction intensity. The association eigenvalue represents the degree of association between person nodes and is calculated by weighting the information interaction intensity divided by the organizational distance. The smaller the organizational distance and the greater the information interaction intensity, the higher the association eigenvalue. For example, if the organizational distance between nodes n1 and n2 is 2 and the information interaction intensity is 1.5 times per day, the association eigenvalue is 1.5 / (2 × 0.5 + 1) = 0.75.

[0086] The time series eigenvalues ​​and the associated eigenvalues ​​are combined to construct a eigenvector. The eigenvector of each person information node ni is Vi = [ts_i, ri_i], where ts_i is the time series eigenvalue and ri is the associated eigenvalue. For example, the eigenvector of node n1 is [3.71, 0.75].

[0087] The feature vector is input into a pretrained LSTM neural network to generate a hidden state sequence. The pretrained LSTM neural network consists of an input layer, a hidden layer, and an output layer. The hidden layer consists of 128 LSTM units. After the input feature vector is processed by the LSTM network, a hidden state sequence H = {h1, h2, ..., ht} is generated. For example, for the feature vector [3.71, 0.75] of node n1, after processing by the LSTM network, the generated hidden state sequence is [0.65, 0.72, 0.83, 0.91, 0.87].

[0088] An attention mechanism is applied to the hidden state sequence to calculate temporal weights. The attention mechanism assigns different weights to different time steps in the hidden state sequence based on the current prediction task. The temporal weight α is calculated by calculating the similarity between each hidden state and the query vector and then performing softmax normalization. For example, for the hidden state sequence of node n1, the calculated temporal weights are [0.05, 0.15, 0.25, 0.35, 0.20].

[0089] The predicted state is obtained by weighting the hidden state sequence according to the time series weights. The predicted state c is calculated as the weighted sum of the hidden state sequence and the corresponding time series weights. For example, for node n1, its predicted state c = 0.65 × 0.05 + 0.72 × 0.15 + 0.83 × 0.25 + 0.91 × 0.35 + 0.87 × 0.20 = 0.84.

[0090] The predicted state is mapped through a fully connected layer to generate a change trend prediction curve. The fully connected layer maps the predicted state to predicted change trend values ​​at multiple future time points. For example, for node n1, the predicted change trend values ​​for the next seven days are [0.82, 0.85, 0.89, 0.92, 0.87, 0.83, 0.81].

[0091] The second-order derivative of the change trend prediction curve is calculated to determine the set of inflection points. The second-order derivative represents the change in the convexity of the curve, and the point where the second-order derivative is zero is the inflection point. The second-order derivative of the change trend prediction curve is approximated using numerical calculation methods to determine the set of inflection points. For example, for the change trend prediction curve of node n1, the calculated inflection points are located on the 4th and 6th days, with corresponding inflection point values ​​of 0.92 and 0.83.

[0092] The dynamic threshold function is determined based on the inflection point set. The dynamic threshold function T is constructed based on the inflection point value and the distribution characteristics of historical data, and adopts a piecewise function form. For the time series eigenvalue ts and the correlation eigenvalue r, the value of the dynamic threshold function T(ts, r) increases with the increase of ts and r, and has obvious changes at the inflection point. The dynamic threshold function is implemented based on the benchmark threshold obtained by historical data statistics, combined with the weighted adjustment of the time series eigenvalue and the correlation eigenvalue. The benchmark threshold is 0.75, the weight of the time series eigenvalue is 0.6, and the weight of the correlation eigenvalue is 0.4. When the time series eigenvalue is less than the first inflection point value, the dynamic threshold function value is the benchmark threshold; when the time series eigenvalue is between the first inflection point value and the second inflection point value, the dynamic threshold function value is the benchmark threshold multiplied by the adjustment coefficient; when the time series eigenvalue is greater than the second inflection point value, the dynamic threshold function value is further adjusted. For example, for a time series eigenvalue of 3.71 and a correlation eigenvalue of 0.75, the dynamic threshold function value is 0.75×(1+0.6×3.71 / 10+0.4×0.75 / 5)=0.97.

[0093] The time series feature values ​​and associated feature values ​​are input into the dynamic threshold function, and personnel information nodes with a value greater than the corresponding threshold are marked as active nodes. If the node's predicted state value is greater than the threshold calculated by the dynamic threshold function, the node is marked as active. For example, if the predicted state value of node n1 is 0.84 and the dynamic threshold function value is 0.97, since 0.84 < 0.97, node n1 is not marked as active. Conversely, if the predicted state value of a node is 0.98 and the dynamic threshold function value is 0.95, the node is marked as active.

[0094] Monitor information request queues in real time and prioritize synchronization of personnel information marked as active nodes to improve the efficiency and accuracy of personnel information synchronization across cross-organizational approval systems.

[0095] In this embodiment, by counting the number of information changes of each node in different time windows and quantifying the change frequency with a time decay function, the activity changes of personnel information can be captured dynamically; the combination of LSTM and attention mechanism allows the model to automatically focus on the most critical historical moments for prediction, thereby significantly improving the accuracy of change trend prediction; jointly modeling the differences in organizational hierarchies and the intensity of information interaction between nodes can reveal the path and intensity of mutual influence among personnel in a collaborative network; the fusion of this correlation feature with its own temporal feature can effectively reflect the pattern of information dissemination and collaborative change within the group, providing a more comprehensive perspective for macro-decision-making; by calculating the second-order derivative of the predicted change trend curve to automatically extract the inflection point, and determining the dynamic threshold function based on the inflection point, the rigid limitations of the static threshold can be avoided; the dynamic threshold can be flexibly adjusted according to the predicted trend of each node, accurately marking the active nodes that truly "suddenly" or "accelerated" changes.

[0096] In an optional embodiment, a multi-level hash tree is constructed based on active nodes, hierarchical perturbation factors are introduced and written into a hierarchical cache pool, a data checksum is calculated and submitted to a main storage pool, and a distribution feature map is constructed based on the similarity of node hash values, including:

[0097] Active nodes are used as leaf nodes of a multi-level hash tree. The total number of layers is determined based on the number of active nodes. The identification information and feature information of the active nodes are combined to calculate the leaf-level hash value. The leaf-level hash value is constructed as the first-level hash value. The first-level hash values ​​are combined in pairs to construct the second-level hash value. The hash values ​​of each layer are constructed upwards in sequence.

[0098] Based on the preset perturbation base and the layer number, the perturbation factor of each layer is determined. The hash value of the first layer node is calculated by combining the first layer hash value and the first layer perturbation factor. Starting from the second layer, the hash value of the lower layer node corresponding to each parent node is combined with the perturbation factor of this layer to obtain the hash value of the node in this layer.

[0099] Construct a hierarchical cache pool and determine the capacity of each level in the hierarchical cache pool based on the number of active nodes; write the hash value of each layer node to the corresponding level; generate a cache pool mask with the same length as the number of hierarchical cache pool levels; perform bitwise operations on the hash value of the top node and the hierarchical cache pool mask to obtain a data check code;

[0100] The similarity of hash values ​​between active nodes is calculated, and the active nodes are regarded as vertex sets, and the pairs of active nodes with similarity greater than the preset similarity threshold are regarded as edge sets to construct the distribution feature map of active nodes.

[0101] Figure 2 This is a diagram illustrating the construction of a multi-level hash tree. In one specific embodiment, the process of constructing a multi-level hash tree of active nodes begins by identifying active nodes in the network. Assume that 12 active nodes are detected in a certain network environment, labeled N1 through N12. Each active node contains identification information (such as MAC address and IP address) and characteristic information (such as traffic characteristics and behavioral characteristics). For node N1, its identification information is "192.168.1.101" and its characteristic information is "HTTP access frequency 85 times / minute." For node N2, its identification information is "192.168.1.102" and its characteristic information is "data transmission volume 2.5 GB / hour." After concatenating the identification information and characteristic information, the SHA-256 algorithm is used to calculate the leaf-level hash value. For example, the leaf-level hash value of node N1 is "7a8b9c0d1e2f3g4h5i6j7k8l9m0n1o2p."

[0102] Based on the number of active nodes (12), the total number of levels in the hash tree is determined to be 4 (leaf level is level 0, and hash levels are counted starting from level 1). The 12 leaf-level hash values ​​are organized into the first-level hash value. The first-level hash value consists of 6 nodes, formed by concatenating the hash values ​​of two adjacent leaf-level nodes and recalculating the hash value. For example, the hash value of the first node in level 1 (labeled L1-1) is calculated by concatenating the leaf-level hash values ​​of nodes N1 and N2 to obtain "3c4d5e6f7g8h9i0j1k2l3m4n5o6p7q". Similarly, the leaf-layer hash values ​​of nodes N3 and N4 are concatenated to obtain the hash value of the second node in the first layer (marked as L1-2); the leaf-layer hash values ​​of nodes N5 and N6 are concatenated to obtain the hash value of the third node in the first layer (marked as L1-3); the leaf-layer hash values ​​of nodes N7 and N8 are concatenated to obtain the hash value of the fourth node in the first layer (marked as L1-4); the leaf-layer hash values ​​of nodes N9 and N10 are concatenated to obtain the hash value of the fifth node in the first layer (marked as L1-5); and the leaf-layer hash values ​​of nodes N11 and N12 are concatenated to obtain the hash value of the sixth node in the first layer (marked as L1-6).

[0103] The second layer has three nodes and is formed by concatenating the hash values ​​of two adjacent nodes in the first layer and recalculating the hash value. Specifically, the hash values ​​of nodes L1-1 and L1-2 are concatenated to obtain the hash value of the first node in the second layer (labeled L2-1); the hash values ​​of nodes L1-3 and L1-4 are concatenated to obtain the hash value of the second node in the second layer (labeled L2-2); and the hash values ​​of nodes L1-5 and L1-6 are concatenated to obtain the hash value of the third node in the second layer (labeled L2-3).

[0104] The third layer has two nodes and is constructed using a similar method. The hash values ​​of nodes L2-1 and L2-2 are concatenated to form the hash value of the first node in the third layer (labeled L3-1). The hash value of node L2-3 (since the second layer has only three nodes and the last node has no paired node) is directly used as the hash value of the second node in the third layer (labeled L3-2).

[0105] The number of nodes in the fourth layer is 1 (root node), which is calculated by concatenating the hash values ​​of the two nodes L3-1 and L3-2 in the third layer, forming the root node of the entire hash tree.

[0106] Layered perturbation factors are introduced to enhance the security of the hash tree. The default perturbation base is "0x5a3c2e1d," and the perturbation factors at each layer are determined by the layer number. The perturbation factor at the first layer is the bitwise operation of "0x5a3c2e1d" and "1," resulting in "0x5a3c2e1e"; the perturbation factor at the second layer is the bitwise operation of "0x5a3c2e1d" and "2," resulting in "0x5a3c2e1f"; the perturbation factor at the third layer is the bitwise operation of "0x5a3c2e1d" and "3," resulting in "0x5a3c2e18"; and the perturbation factor at the fourth layer is the bitwise operation of "0x5a3c2e1d" and "4," resulting in "0x5a3c2e19."

[0107] Each hash value in layer 1 is combined with the layer 1 perturbation factor to produce the hash value for the layer 1 node. This combination is performed by XORing the hash value and the perturbation factor, followed by a SHA-256 hash. For example, the hash value "3c4d5e6f7g8h9i0j1k2l3m4n5o6p7q" of the first node in layer 1 (L1-1) is XORed with the layer 1 perturbation factor "0x5a3c2e1e" to produce the node hash value "4b5c6d7e8f9g0h1i2j3k4l5m6n7o8p" for the first node in layer 1. The same method is applied to all nodes in layers 1-2 through 1-6.

[0108] Starting at layer 2, the hash value of each parent node is calculated by combining the hash value of the node in the lower layer with the perturbation factor of that layer. For example, the first node in layer 2 (L2-1) corresponds to the first and second nodes in layer 1 (L1-1 and L1-2). Its hash value is calculated by concatenating the hash values ​​of these two nodes and combining them with the perturbation factor of layer 2. The same method is applied to nodes L2-2 and L2-3, as well as higher-level nodes L3-1, L3-2, and the root node.

[0109] The hierarchical cache pool is constructed with the capacity of each level determined by the number of active nodes (12). The first-level cache capacity is 6 (corresponding to the number of nodes in the first layer), the second-level cache capacity is 3 (corresponding to the number of nodes in the second layer), the third-level cache capacity is 2 (corresponding to the number of nodes in the third layer), and the fourth-level cache capacity is 1 (corresponding to the number of nodes in the fourth layer). The hash values ​​of the nodes in each layer are written to the corresponding cache pool in hierarchical order. The hash values ​​of the six nodes in the first layer (L1-1 to L1-6) are written to the first-level cache, the hash values ​​of the three nodes in the second layer (L2-1 to L2-3) are written to the second-level cache, the hash values ​​of the two nodes in the third layer (L3-1 and L3-2) are written to the third-level cache, and the hash value of the root node in the fourth layer is written to the fourth-level cache.

[0110] Generate a cache pool mask with a length of 4 (the same as the number of hierarchical cache pool levels). Assuming the mask value is "1011," this enables the first, third, and fourth levels of cache and disables the second level. Perform a bitwise operation on the hash value "9a8b7c6d5e4f3g2h1i" of the top-level node (the fourth-level root node) and the hierarchical cache pool mask "1011" to obtain the data checksum "8b7c6d5e4f3g2h1i." This checksum is submitted to the primary storage pool for persistent storage.

[0111] The construction of the active node distribution feature map is based on the similarity calculation of node hash values. The similarity of the hash values ​​of the 12 active nodes is calculated based on the number of identical bits in the hash values ​​divided by the total number of bits in the hash values. For example, the hash value similarity between nodes N1 and N3 is 0.75, the hash value similarity between nodes N2 and N5 is 0.62, and the hash value similarity between nodes N4 and N8 is 0.81.

[0112] A similarity threshold of 0.7 is set, and node pairs with similarity greater than the threshold are considered to be associated. An undirected graph is constructed, with active nodes as the vertex set and node pairs with similarity greater than the threshold as the edge set. For example, there is an edge between nodes N1 and N3, an edge between nodes N4 and N8, and no edge between nodes N2 and N5. The resulting graph shows that nodes N1, N3, N6, and N9 form one cluster, nodes N4, N8, and N11 form another cluster, and nodes N2, N5, N7, N10, and N12 are relatively independent. By analyzing the graph structure, potential node clusters and abnormal nodes can be identified, providing a basis for network management and security analysis.

[0113] In this embodiment, active nodes are organized into a multi-level hash tree and the hash values ​​of each level are calculated from the bottom up. This can complete positioning and retrieval in massive nodes in logarithmic time, greatly improving data access and query efficiency. A disturbance factor dynamically generated according to the layer number is introduced, and the disturbance is bound to the hash value, so that the hash value of each layer of nodes has controllable randomness, which not only reduces the risk of hash collision, but also can quickly detect data tampering through the top-level data verification code when verifying the hierarchical cache pool. The capacity of each level of the cache pool is dynamically determined according to the number of active nodes, and combined with the hierarchical hash value writing strategy, the hot and cold data are automatically separated by level, which not only ensures the rapid hit of high-frequency access nodes, but also avoids the waste of cache resources.

[0114] In an optional embodiment, a change request is generated by digitally signing the received personnel information change content using a private key, a multidimensional correlation matrix is ​​constructed based on the spatiotemporal correlation data of the distribution feature map, a directed propagation graph is constructed through feature decomposition, a dynamic programming algorithm based on multidimensional attention is executed to obtain the optimal propagation path, a synchronization sequence is generated, and a data check code is obtained from the main storage pool to verify data integrity, including:

[0115] Generate a digital signature for the received personnel information change using a private key, and combine the personnel information change, the digital signature, and a timestamp to generate a change request;

[0116] Calculate the spatiotemporal correlation data of the nodes in the distribution feature map, multiply and accumulate the spatiotemporal correlation data with the time weight that decays exponentially with the time window, and generate a multidimensional correlation matrix corresponding to the node feature dimension;

[0117] Performing eigenvalue decomposition on the multidimensional association matrix to obtain eigenvectors, selecting the eigenvectors corresponding to the maximum preset number of eigenvalues ​​to construct a low-dimensional representation vector, calculating the inner product of the low-dimensional representation vectors of different nodes, and normalizing them to obtain a node affinity matrix;

[0118] The node affinities in the node affinity matrix that are greater than a preset affinity threshold are set as edge weights to construct a directed propagation graph;

[0119] A dynamic programming algorithm based on multi-dimensional attention is executed in the directed propagation graph to calculate the propagation path of the change request. The path with the largest cumulative weight is obtained through multi-layer feature fusion and weight backtracking, and the optimal propagation path is determined.

[0120] The node identifiers and inter-node propagation delays in the optimal propagation path are combined to generate a change request synchronization sequence, a hash operation is performed on the change request synchronization sequence, and the sequence is compared with the data check code in the primary storage pool to verify data integrity.

[0121] Figure 3This is a schematic diagram of the architecture of a change request synchronization sequence generation system based on multi-dimensional attention dynamic programming. In one specific implementation, the secure transmission of personnel information changes is first addressed. After receiving the personnel information changes, a digital signature is generated for the changes using an RSA-2048-bit private key. The personnel information changes are calculated using the SHA-256 algorithm to generate a message digest, which is then encrypted using the private key to produce a 256-byte digital signature. For example, when receiving the change information "User Zhang Gong's position has changed from 'Engineer' to 'Senior Engineer'," the message digest "7f83b1657ff1fc53b92dc18148a1d65dfc2d4b1fa3d677284addd200126d9069" is generated and encrypted using the private key to generate a digital signature. The system then obtains a UTC timestamp accurate to milliseconds (such as "2025-06-09T08:15:23.456Z") and combines the change content, digital signature, and timestamp into a change request in JSON format, such as {"content":"User Zhang Gong's position changed from 'Engineer' to 'Senior Engineer'", "signature":"a1b2c3...", "timestamp":"2025-06-09T08:15:23.456Z"}.

[0122] Then, the spatiotemporal correlation data of the nodes in the distribution feature map are calculated. Each node represents an organizational unit, and the interaction data between nodes in the past 30 days are collected, including three dimensions: information transmission frequency, average response time, and interaction content similarity. The time window exponential decay function is used to process historical data, and the decay coefficient is set to 0.95, so that the more recent interaction data has a greater weight. For each dimension of spatiotemporal correlation data, multiply it by the corresponding time weight and accumulate it to generate an N×N×3 multidimensional correlation matrix (N is the number of nodes). For example, for a system with 5 nodes, the generated matrix size is 5×5×3, where the matrix [2][3][0] represents the weighted correlation value between node 2 and node 3 in the dimension of information transmission frequency.

[0123] The obtained multidimensional association matrix is ​​subjected to eigenvalue decomposition, and the singular value decomposition (SVD) algorithm is used to decompose the three-dimensional matrix separately, extracting the eigenvalues ​​and corresponding eigenvectors. According to the preset parameters, the eigenvectors corresponding to the first 8 largest eigenvalues ​​are selected to construct a low-dimensional representation vector for each node. For the above example of 5 nodes, each node will obtain an 8-dimensional representation vector, such as the representation vector of node 1 is [0.253, 0.142, -0.368, 0.097, 0.421, -0.183, 0.276, 0.119]. The inner product between the representation vectors of different nodes is calculated. For example, the inner product of the vectors of node 1 and node 2 is calculated to be 0.687. Then, through MinMax normalization, all inner product values ​​are mapped to the interval [0, 1] to form a node affinity matrix.

[0124] Construct a directed propagation graph based on the node affinity matrix. Set the preset affinity threshold to 0.65 and traverse each element in the affinity matrix. When the affinity between nodes is greater than 0.65, a directed edge is established between the corresponding nodes, with the edge weight set to the affinity value. For example, if the affinity between nodes 1 and 2 is 0.71, the affinity between nodes 1 and 3 is 0.58, and the affinity between nodes 2 and 3 is 0.83, then directed edges are established between nodes 1 and 2 and between nodes 2 and 3, with weights of 0.71 and 0.83, respectively. No direct connection is established between nodes 1 and 3.

[0125] A dynamic programming algorithm based on multi-dimensional attention is executed on the constructed directed propagation graph to calculate the propagation path of the change request. Three attention weights are assigned to each node in the graph: information sensitivity, processing power, and historical response weight, with initial values ​​of [0.4, 0.3, 0.3], respectively. Starting from the source node, the cumulative weight to each reachable node is calculated as the dot product of the current path weight and the target node's attention weight. Multiple rounds of iteration are performed, with each round adjusting the attention weight distribution based on the characteristics of the visited nodes. For example, when a path passes through multiple nodes with high information sensitivity, the processing power weight is dynamically increased to balance propagation efficiency. After the iterations are complete, the weights of the three dimensions are combined through multi-layer feature fusion. A weighted backtracking algorithm is used to trace back from the end node to the source node, and the path with the largest cumulative weight is determined as the optimal propagation path. For a system with five nodes, the optimal propagation path might be 1→2→5→4, with a cumulative weight of 0.927.

[0126] The change request synchronization sequence is generated by combining the node identifiers and inter-node propagation delays in the optimal propagation path. Node identifiers are in UUID format, such as "f47ac10b-58cc-4372-a567-0e02b2c3d479". Inter-node propagation delays are calculated based on historical data and are accurate to the millisecond. The change request synchronization sequence is formatted as a JSON array, where each element contains the node identifier, estimated arrival time, and processing completion time, such as [{"nodeId":"f47ac10b-58cc-4372-a567-0e02b2c3d479","arriveTime":"2025-06-09T08:15:23.789Z","completeTime":"2025-06-09T08:15:24.123Z"}, ...]. The system performs a SHA-256 hash operation on this synchronization sequence, generating a 64-bit hexadecimal string. This is then compared with the data checksum in the primary storage pool to verify data integrity. If the hash values ​​match, the change request synchronization sequence is confirmed to be valid, and the actual synchronization operation begins. If not, the system logs an error and triggers a recalculation process.

[0127] In this embodiment, by using a private key to generate a digital signature for the changed content and attaching a timestamp, not only can the authenticity and non-repudiation of the request source be ensured, but also the data can be prevented from being tampered with during transmission, greatly improving the security and trustworthiness of the system; a multidimensional correlation matrix is ​​calculated based on spatiotemporal correlation data and exponentially decaying time weights, and eigenvalue decomposition is performed to generate a low-dimensional representation vector, so that the similarity and correlation between nodes can be accurately characterized, providing solid data support for subsequent propagation path calculations; the normalized node affinity matrix is ​​converted into a directed propagation graph, and a dynamic programming algorithm with multidimensional attention is introduced into it, which can not only efficiently discover the optimal propagation path with the largest cumulative weight, but also achieve accurate identification and priority synchronization of key nodes through weight backtracking; the node identifiers and propagation delays in the optimal propagation path generate a synchronization sequence, and through hash operations and comparison with the main storage pool check code, the integrity and consistency of data synchronization can be verified in real time, ensuring that the data status of each node always remains highly reliable.

[0128] In an optional embodiment, a dynamic programming algorithm based on multi-dimensional attention is executed in a directed propagation graph to calculate the propagation path of the change request, and a path with the largest cumulative weight is obtained through multi-layer feature fusion and weight backtracking. Determining the optimal propagation path includes:

[0129] The propagation weights between nodes in the directed propagation graph are combined with the weighted historical synchronization delays to generate an initial propagation weight matrix;

[0130] Calculate the multi-dimensional attention value between nodes in the propagation weight matrix, normalize the multi-dimensional attention value to obtain the node propagation priority, and use the weighted product of the node propagation priority and the historical synchronization success rate as the state transition weight to generate the state transition matrix;

[0131] Based on the multi-channel structure, the state transition matrix is ​​transformed and weighted fusion is performed to obtain the cumulative propagation weight between nodes. The state transition probability is calculated based on the ratio of the cumulative propagation weight to the node processing capacity, and the optimal state transition sequence is constructed.

[0132] The historical synchronization features between nodes on the propagation path are used to extract deep features through a multi-layer skip connection network, and the deep features are combined with the accumulated propagation weights to obtain the path evaluation value;

[0133] The optimal number of propagation hops is selected based on the path evaluation value, and the prediction model is trained using the differential strategy optimization method. The prediction model is used to calculate the optimal preceding node of the node;

[0134] Starting from the target node, the optimal preceding nodes are connected sequentially until the source node, generating the optimal propagation path with the maximum cumulative propagation weight.

[0135] In one specific implementation, basic data for a directed propagation graph is obtained, including node information and inter-node propagation relationships. Based on historical propagation records, direct inter-node propagation weights and historical synchronization delays can be obtained. For a change request, suppose source node A needs to propagate to target node F, potentially passing through nodes B, C, D, and E. The propagation weight from node A to B is 0.82, with a historical synchronization delay of 45ms; the propagation weight from A to C is 0.75, with a historical synchronization delay of 62ms; and so on, to construct an initial propagation weight matrix.

[0136] Perform a weighted combination on the initial propagation weight matrix. Set the propagation weight factor α to 0.7 and the delay factor β to 0.3. For node pair (i, j), calculate the combined weight W(i, j) = α × propagation weight (i, j) - β × normalized delay (i, j). For example, the combined weight from A to B is 0.7 × 0.82 - 0.3 × (45 / 100) = 0.574 - 0.135 = 0.439, where the delay is normalized to 100 ms. A similar calculation is performed to obtain the complete initial propagation weight matrix.

[0137] Calculate the multi-dimensional attention values ​​between nodes. For each node pair (i, j), consider the attention characteristics of three dimensions: propagation frequency, data size, and processing priority. For example, the propagation frequency attention from A to B is 0.85, the data size attention is 0.72, and the processing priority attention is 0.91. Normalize these three attention values ​​to obtain the node propagation priority P(i, j) = (0.85 + 0.72 + 0.91) / 3 = 0.827. Simultaneously, obtain the historical synchronization success rate S(i, j). Assume that the historical synchronization success rate from A to B is 0.94. The weighted product of the node propagation priority and the historical synchronization success rate is used as the state transition weight T(i, j) = P(i, j) × S(i, j) = 0.827 × 0.94 = 0.777. A similar calculation is used to obtain the complete state transition matrix.

[0138] A multi-channel structure is used to perform feature transformation on the state transition matrix. Three feature channels are designed: a time sensitivity channel, a resource consumption channel, and a service importance channel. For the node pair (i, j), the eigenvalue of the time sensitivity channel is 0.88, the eigenvalue of the resource consumption channel is 0.65, and the eigenvalue of the service importance channel is 0.79. The weights of the three channels are set to 0.4, 0.25, and 0.35, respectively. Through weighted fusion, the cumulative inter-node propagation weight C(i, j) = 0.4 × 0.88 + 0.25 × 0.65 + 0.35 × 0.79 = 0.799 is obtained. The state transition probability is calculated based on the ratio of the cumulative propagation weight to the node processing capacity. Assuming that the processing capacity of node B is 0.85, the state transition probability is 0.799 / 0.85 = 0.94. When constructing the optimal state transition sequence, the node with the highest state transition probability is selected as the next hop node.

[0139] When evaluating propagation paths, historical synchronization features between nodes are extracted. These features are processed through a multi-layer skip connection network to extract deep features. For example, for path ABDF, the extracted deep feature value is 0.86. Combining the deep features with the cumulative propagation weight yields a path evaluation value of E = 0.6 × 0.86 + 0.4 × average cumulative propagation weight = 0.6 × 0.86 + 0.4 × 0.78 = 0.516 + 0.312 = 0.828. The optimal number of propagation hops is determined based on the path evaluation value, with 3-4 hops typically being the optimal balance.

[0140] The prediction model is trained using a differential strategy optimization method. A training set is constructed using historical propagation data, consisting of feature vectors and labels of optimal predecessor nodes. The feature vectors include metrics such as node load, propagation delay, and link stability. The trained prediction model is used to calculate the optimal predecessor node for each node. For example, for target node F, the prediction model outputs the optimal predecessor node as D, the optimal predecessor node of node D as B, and the optimal predecessor node of node B as A.

[0141] Starting from the target node F, the optimal preceding nodes D and B are sequentially connected to the source node A, generating the optimal propagation path A→→B→D→F with the maximum cumulative propagation weight. This path has a cumulative propagation weight of 0.842, significantly higher than other possible paths such as A→C→E→F (0.763) and A→C→D→F (0.791). This approach achieves optimal propagation path planning for change requests in the directed propagation graph, ensuring maximum propagation efficiency and success rate.

[0142] Existing organizational personnel information synchronization technologies primarily utilize simple broadcast or fixed routing models for data dissemination. These methods often struggle to adapt to complex and changing organizational network structures and dynamically changing network conditions. Traditional shortest path algorithms, such as the Dijkstra algorithm, can find the shortest path between nodes, but they fail to comprehensively consider multi-dimensional factors such as dissemination efficiency, data processing capacity, and business importance. This leads to problems such as high synchronization latency and failure rates in practical applications. Using a single-dimensional evaluation metric fails to effectively address the complex interactions between organizations. For example, existing path planning methods based on dissemination weights consider the strength of connections between nodes but ignore the impact of latency. Delay-based path planning methods overly focus on dissemination speed and overlook the importance of dissemination success rate. Furthermore, existing technologies typically employ static planning methods that fail to dynamically adjust dissemination strategies based on network conditions, making them difficult to address abnormal situations such as traffic bursts or node failures.

[0143] The method of this embodiment starts from the perspective of multi-dimensional evaluation of the propagation path, and proposes a method for generating the initial propagation weight matrix based on a weight-delay weighted combination, organically combining the two key factors of propagation efficiency and reliability. By introducing a multi-dimensional attention mechanism, this method can comprehensively evaluate the complex characteristics of the propagation relationship between nodes and more accurately characterize the collaboration model between organizations. In particular, by processing the state transition matrix through a multi-channel structure, this method can comprehensively evaluate propagation decisions from three key dimensions: time sensitivity, resource consumption, and business importance, greatly improving the accuracy of propagation path planning.

[0144] The differential strategy optimization method employed in this embodiment enables the system to adaptively learn optimal decision rules from historical propagation data and continuously optimize the propagation strategy. Compared with traditional methods, the proposed multi-layer skip connection network can more effectively extract deep features of historical synchronization between nodes and capture hidden propagation patterns. The propagation path is constructed by backtracking the optimal preceding node, ensuring the acquisition of a global optimal solution and avoiding the trap of local optimality.

[0145] Figure 4This figure shows a performance comparison of propagation paths based on the multi-dimensional attention dynamic programming algorithm. In terms of cumulative propagation weight, the multi-dimensional attention dynamic programming algorithm achieved a performance value of 1.4, significantly higher than both the traditional broadcast model and the Dijkstra algorithm. This demonstrates that the algorithm successfully constructs a superior propagation weight system through its multi-channel structure, processing the state transition matrix and combining weights and delays.

[0146] The multi-dimensional attention dynamic programming algorithm also achieved an excellent transmission success rate of 1.4, 44% higher than traditional methods and the Dijkstra algorithm. This is attributed to the algorithm's introduction of a multi-dimensional attention mechanism and the prediction model trained using differential strategy optimization, which effectively improved the reliability of inter-node transmission.

[0147] In terms of average propagation delay, this algorithm significantly outperforms the other two methods with low latency. This is due to the deep features extracted from the multi-layer skip connection network and the optimal number of propagation hops selected based on path evaluation values, which enables change requests to propagate along the most efficient path.

[0148] In terms of resource utilization efficiency, this algorithm can allocate and utilize network resources more reasonably after comprehensively considering the node processing capacity and state migration probability.

[0149] In an optional implementation, using a public key to verify the digital signature of a change request, updating node information according to a change request synchronization sequence, dividing the network into slices based on a directed propagation graph, performing a consistency check based on node hash values ​​in a hierarchical cache pool within the network slice, generating an error detection value based on a data check code in a primary storage pool between network slices, constructing an anomaly propagation graph based on the error detection value to locate the fault source node and correct the status information includes:

[0150] Receive change requests, verify digital signatures using public keys, and update information in nodes according to the change request synchronization sequence;

[0151] Calculate the ratio of the sum of the node neighbor weights in the directed propagation graph to the global weight, aggregate the nodes that meet the preset slice threshold conditions into network slices, and establish cross-slice connection relationships between network slices;

[0152] In the network slice, the node status, timestamp and previous node identifier are concatenated to obtain a node feature sequence. A weighted hash operation is performed on the node feature sequence to generate a multidimensional hash value. The ratio of the intersection cardinality to the union cardinality of the multidimensional hash values ​​between nodes is calculated as the consistency measure.

[0153] Generate a check code for the network slice data, and perform an XOR operation on the check code and the check code in the main storage pool to obtain an error detection value;

[0154] Extract abnormal nodes whose error detection values ​​exceed the preset fault tolerance threshold. Use the intersection nodes of abnormal nodes and change request propagation paths as vertices of the abnormal propagation graph. Build an edge weight matrix based on delay correlation and construct the abnormal propagation graph through eigendecomposition.

[0155] Calculate the weighted out-degree sum of the nodes in the anomaly propagation graph and select the node with the largest weighted out-degree sum as the fault source node;

[0156] Correct the status information of the fault source node based on the checksum in the primary storage pool and the node backup data.

[0157] In a specific embodiment, each node in the distributed system is configured with a unique identifier and a public-private key pair. When a change request is received, the node uses the sender's public key to verify the digital signature attached to the request. After the verification is passed, according to the instructions in the change request, the node updates the local information in a synchronization sequence. For example, node A receives a change request "{operation: 'Update user information', data: {user ID: 12345, new address: 'a certain street'}, timestamp: 1634567890, signature: '0x8a7b...3d2f'}", and after verifying the signature with the sender's public key, it updates the address information of the user ID 12345 to "a certain street".

[0158] During network slicing, the ratio of the sum of neighbor weights to the global weight is calculated for each node. For example, if a node's sum of neighbor weights is 82 and its global weight is 200, the ratio is 0.41. When this ratio exceeds a preset slice threshold (e.g., 0.4), the node is included in the current network slice. For example, in a financial transaction system, multiple network slices might be formed, such as a "payment processing slice," an "account management slice," and a "risk control and review slice." The system also establishes cross-slice connections by selecting node pairs with the highest cross-slice connectivity, ensuring efficient communication between slices.

[0159] To assess the consistency of nodes within a network slice, each node's state, timestamp, and previous node identifier are concatenated to generate a node signature sequence. For example, node B's signature sequence might be "ACTIVE_1634568920_NODE_A." A weighted hash operation is performed on this signature sequence to generate a multidimensional hash value, such as [0x4a2f, 0x8b1c, 0x3d7e]. The system calculates the ratio of the intersection cardinality to the union cardinality of the multidimensional hash values ​​between nodes as a consistency metric. For example, the multidimensional hash values ​​of nodes B and C are [0x4a2f, 0x8b1c, 0x3d7e] and [0x4a2f, 0x9d2e, 0x3d7e], respectively. Their intersection cardinality is 2, their union cardinality is 4, and their consistency metric is 0.5.

[0160] For each network slice's data, a checksum is generated using methods such as Reed-Solomon coding. For example, if the slice data is [25, 63, 88, 42], the generated checksum is [173, 29]. The system performs an XOR operation on this checksum with the standard checksum in the main storage pool to obtain an error detection value. If the standard checksum in the main storage pool is [173, 29], the error detection value is [0, 0], indicating that the data is completely consistent. If the error detection value is [4, 2], it indicates a data anomaly.

[0161] When the error detection value exceeds a preset tolerance threshold (e.g., a non-zero value), an abnormal node is identified. These abnormal nodes are intersected with the change request propagation path to obtain the vertex set of the abnormal propagation graph. For example, if the abnormal node set is {node D, node E, node F, node G}, and the nodes involved in the change request propagation path are {node C, node D, node E, node H}, then the vertex of the abnormal propagation graph is {node D, node E}.

[0162] An edge weight matrix is ​​constructed based on the latency correlation between nodes. For example, if the latency correlation between node D and node E is 0.8, this means that a change in node D's state will cause a change in node E's state within a specific time window with an 80% probability. By performing eigendecomposition on this matrix, a complete anomaly propagation graph is constructed.

[0163] In the anomaly propagation graph, the weighted sum of out-degrees is calculated for each node. For example, if node D is connected to nodes E and F with edge weights of 0.8 and 0.6, respectively, then the weighted sum of out-degrees for node D is 1.4. The system selects the node with the largest weighted sum of out-degrees as the fault source node. In this example, if the weighted sum of out-degrees for node D is 1.4 and the weighted sum of out-degrees for node E is 0.9, node D is identified as the fault source node.

[0164] After determining the source node, the state information of the source node is corrected based on the checksum in the primary storage pool and the node's backup data. For example, if the correct data corresponding to the checksum of node D in the primary storage pool is [25, 63, 88, 42], and the current data of node D is [25, 67, 88, 42], the second data item of node D is corrected from 67 to 63, restoring normal system operation.

[0165] In this embodiment, the public key is used to verify the digital signature to ensure that only authorized change requests can be synchronized between nodes, preventing malicious tampering and unauthorized operations; the network slices are dynamically divided based on the ratio of the node neighbor weight to the global weight to achieve on-demand resource aggregation and isolation, enhancing the scalability and load balancing capabilities of the system; by performing weighted hashing on the node feature sequence and calculating the intersection and union ratio of the multi-dimensional hash values, a fine-grained consistency measurement is provided to improve the accuracy of data synchronization and consistency management; a check code is generated within the slice and XORed with the main storage check code, which can quickly detect abnormal nodes that exceed the fault tolerance threshold and provide accurate clues for subsequent diagnosis; an abnormal propagation graph based on delay correlation is constructed, and the source of the fault is quickly located by weighted out-degree and selecting the maximum node, shortening the fault diagnosis time; the state of the fault source node is corrected using the main storage check code and node backup data to ensure that the system can automatically recover to a consistent and correct state.

[0166] In an optional embodiment, abnormal nodes whose error detection values ​​exceed a preset fault tolerance threshold are extracted, intersection nodes of the abnormal nodes and the change request propagation path are used as vertices of the abnormal propagation graph, an edge weight matrix is ​​constructed based on the delay correlation, and the abnormal propagation graph is constructed by eigendecomposition, including:

[0167] Obtain the error detection value sequence of the node, and mark the node whose difference between the current error detection value and the expected error detection value exceeds the preset fault tolerance threshold as an abnormal node;

[0168] Construct a time-decay weight, multiply it with the error detection value sequence, and normalize it to obtain the abnormality degree of the abnormal node;

[0169] Obtaining the propagation path of the change request, and calculating the intersection node of the abnormal node and the propagation path as the vertex of the abnormal propagation graph;

[0170] Calculate the propagation delay between vertices, substitute it into the exponential decay function to get the delay correlation, and count the number of propagations between vertices to get the propagation probability;

[0171] The delay correlation, propagation probability and consistency measure are combined according to the preset weight coefficient to construct the edge weight matrix;

[0172] Perform eigendecomposition on the edge weight matrix to obtain eigenvalues ​​and eigenvectors, and select the product of the square root of the eigenvalue and the eigenvector to construct the node propagation eigenvector;

[0173] The node association strength is obtained by calculating the inner product of the node propagation feature vector and the modulus ratio, and the vertex pairs with node association strength greater than a preset strength threshold are taken as edges, and the consistency measure is used as the edge weight to construct the abnormal propagation graph.

[0174] In a specific embodiment, during the real-time synchronization of personnel in the cross-organizational approval system, the system continuously monitors the error detection value sequence of each node to identify abnormal nodes. The error detection value includes five indicators: network delay, packet loss rate, CPU utilization, memory utilization and request response time. The system collects error detection values ​​for each node in the past 24 hours, with a total of 288 data points at intervals of 5 minutes. For the network delay indicator of a certain node, the current value is 120ms, and the expected value is 50ms. The difference of 70ms between the two exceeds the preset fault tolerance threshold of 40ms, and the system marks the node as an abnormal node. Similarly, if the packet loss rate of the node is 4.2%, which exceeds the preset fault tolerance threshold of 3%, it will also trigger an abnormal mark. In fact, if three or more indicators of a node are marked as abnormal, the node will be determined as an abnormal node.

[0175] Time-decay weights are constructed to assess the severity of anomaly nodes. These weights use an exponential decay function with a half-life of 4 hours, giving recent false positives a greater impact on anomaly assessment. A time-decay weight sequence is constructed for 288 data points, with the most recent time point receiving a weight of 1, the previous time point receiving a weight of 0.9826, and so on. The time-decay weights are multiplied by the false positive sequence. For example, if the network latency of a node at the last three time points was 120ms, 115ms, and 105ms, the corresponding time-decay weights are 1, 0.9826, and 0.9656, resulting in 120ms, 112.999ms, and 101.388ms, respectively. The five metrics are calculated and normalized, resulting in a node abnormality score of 0.78 (out of a maximum score of 1). Nodes with an abnormality score greater than 0.5 are included in the construction of the anomaly propagation graph.

[0176] Obtain the propagation path of the change request, which represents the order in which changes to organizational personnel information propagate across the organization. For example, if the propagation path of a change request is A→B→C→D→E→F, nodes B and E are marked as abnormal nodes. The intersection of the abnormal nodes and the propagation path is calculated, resulting in B and E as vertices of the abnormal propagation graph. Simultaneously, the system analyzes other nodes that directly interact with B and E. Nodes G and H frequently interact with B, and node I frequently interacts with E. Therefore, G, H, and I are also included in the vertex set of the abnormal propagation graph. The final vertex set of the abnormal propagation graph is {B, E, G, H, I}.

[0177] Calculate the propagation delay between vertices, including network transmission delay and node processing delay. For example, the network transmission delay from node B to node E is 85ms, the processing delay at node E is 120ms, and the total propagation delay is 205ms. Substituting the propagation delay into an exponential decay function with a parameter of 0.005, the delay correlation between B and E is 0.359. Similarly, calculate the delay correlation between all pairs of vertices. Also count the number of propagations between vertices over the past 30 days. For example, the number of propagations from B to E is 127, the number of propagations from E to B is 42, and the total number of propagations is 169. Based on the number of propagations, the probability of propagation from B to E is 0.751 (127 / 169), and the probability of propagation from E to B is 0.249 (42 / 169).

[0178] Calculate the consistency metric between vertices to reflect the consistency level of data synchronization between nodes. The consistency metric is based on historical data verification results and calculates the ratio of correct synchronization times to the total synchronization times. For example, if the number of correct synchronizations from B to E is 118 and the total number of synchronizations is 127, the consistency metric is 0.929 (118 / 127). Set the weight coefficients for delay correlation, propagation probability, and consistency metric to 0.3, 0.3, and 0.4, respectively, and construct an edge weight matrix. The edge weight from B to E is calculated as: 0.3 × 0.359 + 0.3 × 0.751 + 0.4 × 0.929 = 0.1077 + 0.2253 + 0.3716 = 0.7046. Similarly, calculate the edge weights between all vertex pairs to form a 5 × 5 edge weight matrix.

[0179] Perform eigendecomposition on the edge weight matrix to calculate the eigenvalues ​​and corresponding eigenvectors. Assume the first three eigenvalues ​​calculated are 3.425, 1.782, and 0.946, and the corresponding eigenvectors are three 5-dimensional vectors. The square root of the eigenvalue multiplied by the corresponding eigenvector is used to construct the node propagation eigenvector. For example, the propagation eigenvector of node B is [1.851, 0.966, 0.512], and the propagation eigenvector of node E is [1.782, -0.845, 0.376]. Calculate the inner product of the node propagation eigenvectors and the ratio of their modulo lengths to obtain the node association strength. The inner product of the eigenvectors of B and E is 1.851×1.782 +0.966×(-0.845) + 0.512×0.376 = 3.299 - 0.816 + 0.193 = 2.676. The modulus of B's ​​eigenvector is 2.147, and the modulus of E's eigenvector is 2.014. Therefore, the node association strength between B and E is 2.676 / (2.147×2.014) = 0.619.

[0180] Vertex pairs with node association strength greater than the preset strength threshold of 0.5 are used as edges, and the consistency measure is used as the edge weight to construct an anomaly propagation graph. In this example, the node association strength between B and E is 0.619, which is greater than the preset strength threshold, so an edge is established between B and E with an edge weight of 0.929, the consistency measure. Similarly, the system determines other vertex pairs that meet the conditions and constructs a complete anomaly propagation graph. The final constructed anomaly propagation graph contains vertices {B, E, G, H, I} and edges {(B, E), (B, G), (B, H), (E, I), (G, H)}, with edge weights {0.929, 0.876, 0.912, 0.894, 0.832} respectively.

[0181] Typical anomalous node detection methods rely primarily on static thresholds or simple statistical analysis, failing to effectively capture the dynamic relationships between nodes. These methods typically independently evaluate the status of each node, ignoring its position and influence within the propagation network. This leads to inaccurate anomaly assessments and incomplete assessments of propagation impact. Existing graph structure analysis methods often employ a single weighting metric to construct networks, making it difficult to reflect the multi-dimensional nature of node interactions and limiting their effectiveness in identifying anomalous propagation paths.

[0182] The method of this embodiment starts from the perspective of multi-dimensional anomaly detection and dynamic correlation analysis, introduces a time decay weight mechanism, and can more accurately capture the time evolution characteristics of the node state. By comprehensively considering the three key indicators of delay correlation, propagation probability and consistency measurement, this method constructs a more comprehensive and accurate edge weight matrix, which can better reflect the actual interaction relationship between nodes. In particular, the characteristic decomposition method is used to extract the node propagation feature vector, which can extract the key propagation pattern from the high-dimensional interaction data. The embodiment of the present invention is based on the cross-organizational approval system and the real-time synchronization system of institutional personnel based on public and private keys. It includes:

[0183] The first unit is used to receive personnel information synchronization requests from organizations and obtain information request queues;

[0184] The second unit is used to generate a unique public-private key pair for each organization based on the information request queue, distribute the public key to other organizations, and encrypt and store the private key locally;

[0185] The third unit is used to obtain personnel information from the information request queue to construct a node set, predict and mark active nodes through a long short-term memory neural network; construct a multi-level hash tree based on active nodes, introduce a hierarchical perturbation factor to write to the hierarchical cache pool, calculate the data check code and submit it to the main storage pool, and construct a distribution feature map based on the similarity of node hash values;

[0186] The fourth unit is used to digitally sign the received personnel information change content using the private key to generate a change request, construct a multidimensional association matrix based on the distribution feature map, construct a directed propagation graph through feature decomposition, perform dynamic programming with multidimensional attention to obtain the optimal propagation path, generate a synchronization sequence, and verify data integrity through the main storage pool;

[0187] The fifth unit is configured to verify the digital signature of the change request using a public key, update node information according to the change request synchronization sequence, divide the network into slices based on a directed propagation graph, perform consistency checks within the network slices based on the node hash values ​​in the hierarchical cache pool, generate error detection values ​​between network slices based on the data checksum of the main storage pool, construct an anomaly propagation graph based on the error detection values ​​to locate the fault source node and correct the status information. A third aspect of an embodiment of the present invention provides an electronic device comprising:

[0188] processor;

[0189] a memory for storing processor-executable instructions;

[0190] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0191] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0192] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0193] Finally, 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time synchronization method for personnel in a cross-organizational approval system based on public and private keys, characterized in that: include: Receive personnel information synchronization requests from organizations and obtain information request queues; Generate a unique public-private key pair for each organization based on the information request queue, distribute the public key to other organizations, and encrypt and store the private key locally; Obtain personnel information from the information request queue to build a node set, and use the long short-term memory neural network to predict and mark active nodes; A multi-level hash tree is constructed based on active nodes. A hierarchical perturbation factor is introduced to write data into the hierarchical cache pool. The data checksum is calculated and submitted to the main storage pool. A distribution feature map is constructed based on the similarity of node hash values. The received personnel information change content is digitally signed using the private key to generate a change request. A multidimensional correlation matrix is ​​constructed based on the spatiotemporal correlation data of the distribution feature map. A directed propagation graph is constructed through feature decomposition. A dynamic programming algorithm based on multidimensional attention is executed to obtain the optimal propagation path, generate a synchronization sequence, and obtain a data check code from the main storage pool to verify data integrity. Use the public key to verify the digital signature of the change request, update node information according to the change request synchronization sequence, divide the network into slices based on the directed propagation graph, perform consistency checks based on the node hash values ​​in the hierarchical cache pool within the network slice, generate error detection values ​​based on the data checksum of the main storage pool between network slices, and build an anomaly propagation graph based on the error detection values ​​to locate the fault source node and correct the status information; A multi-level hash tree is constructed based on active nodes. After introducing a hierarchical perturbation factor, the data is written into the hierarchical cache pool. The data checksum is calculated and submitted to the main storage pool. A distribution feature map is constructed based on the similarity of node hash values, including: Active nodes are used as leaf nodes of a multi-level hash tree. The total number of layers is determined based on the number of active nodes. The identification information and feature information of the active nodes are combined to calculate the leaf-level hash value. The leaf-level hash value is constructed as the first-level hash value. The first-level hash values ​​are combined in pairs to construct the second-level hash value. The hash values ​​of each layer are constructed upwards in sequence. Based on the preset perturbation base and the layer number, the perturbation factor of each layer is determined. The hash value of the first layer node is calculated by combining the first layer hash value and the first layer perturbation factor. Starting from the second layer, the hash value of the lower layer node corresponding to each parent node is combined with the perturbation factor of this layer to obtain the hash value of the node in this layer. Construct a hierarchical cache pool and determine the capacity of each level in the hierarchical cache pool based on the number of active nodes; write the hash value of each layer node to the corresponding level; generate a cache pool mask with the same length as the number of hierarchical cache pool levels; perform bitwise operations on the hash value of the top node and the hierarchical cache pool mask to obtain a data check code; Submitting the data verification code to the main storage pool for persistent storage; The similarity of hash values ​​between active nodes is calculated, and the active nodes are regarded as vertex sets, and the pairs of active nodes with similarity greater than the preset similarity threshold are regarded as edge sets to construct the distribution feature map of active nodes.

2. The method according to claim 1, characterized in that Obtain personnel information from the information request queue to build a node set. Use the long short-term memory neural network to predict and mark active nodes, including: Extract personnel identification information and organization identification information from the information request queue, construct a set of personnel information nodes, obtain the time window sequence of each personnel information node, calculate the number of personnel information changes in each time window to obtain the information change frequency, and calculate the time series feature value based on the time decay function of the time window and the information change frequency; Obtain the organizational hierarchy difference of each personnel information node to obtain the organizational distance, count the number of information interactions between personnel information nodes to obtain the information interaction intensity, and calculate the correlation eigenvalue based on the organizational distance and information interaction intensity; The time series eigenvalues ​​and the associated eigenvalues ​​are used to construct a feature vector, which is input into the pre-trained long short-term memory neural network to obtain the hidden state sequence. The attention mechanism is applied to the hidden state sequence to calculate the time series weight. The hidden state sequence is weighted according to the time series weight to obtain the predicted state. The predicted state is mapped through a fully connected layer to obtain a change trend prediction curve. The second-order derivative of the change trend prediction curve is calculated to determine the inflection point set. The dynamic threshold function is determined based on the inflection point set. The time series eigenvalues ​​and the associated eigenvalues ​​are input into the dynamic threshold function, and the personnel information nodes greater than the corresponding threshold are marked as active nodes.

3. The method according to claim 1, characterized in that The received personnel information change content is digitally signed using the private key to generate a change request. A multidimensional correlation matrix is ​​constructed based on the spatiotemporal correlation data of the distribution feature map. A directed propagation graph is constructed through feature decomposition. A dynamic programming algorithm based on multidimensional attention is executed to obtain the optimal propagation path. A synchronization sequence is generated, and a data check code is obtained from the main storage pool to verify data integrity. This includes: Generate a digital signature for the received personnel information change using a private key, and combine the personnel information change, the digital signature, and a timestamp to generate a change request; Calculate the spatiotemporal correlation data of the nodes in the distribution feature map, multiply and accumulate the spatiotemporal correlation data with the time weight that decays exponentially with the time window, and generate a multidimensional correlation matrix corresponding to the node feature dimension; Performing eigenvalue decomposition on the multidimensional association matrix to obtain eigenvectors, selecting the eigenvectors corresponding to the maximum preset number of eigenvalues ​​to construct a low-dimensional representation vector, calculating the inner product of the low-dimensional representation vectors of different nodes, and normalizing them to obtain a node affinity matrix; The node affinities in the node affinity matrix that are greater than a preset affinity threshold are set as edge weights to construct a directed propagation graph; A dynamic programming algorithm based on multi-dimensional attention is executed in the directed propagation graph to calculate the propagation path of the change request. The path with the largest cumulative weight is obtained through multi-layer feature fusion and weight backtracking, and the optimal propagation path is determined. The node identifiers and inter-node propagation delays in the optimal propagation path are combined to generate a change request synchronization sequence, a hash operation is performed on the change request synchronization sequence, and the sequence is compared with the data check code in the primary storage pool to verify data integrity.

4. The method according to claim 3, characterized in that A dynamic programming algorithm based on multi-dimensional attention is executed in the directed propagation graph to calculate the propagation path of the change request. The path with the largest cumulative weight is obtained through multi-layer feature fusion and weight backtracking. Determining the optimal propagation path includes: The propagation weights between nodes in the directed propagation graph are combined with the weighted historical synchronization delays to generate an initial propagation weight matrix; Calculate the multi-dimensional attention value between nodes in the propagation weight matrix, normalize the multi-dimensional attention value to obtain the node propagation priority, and use the weighted product of the node propagation priority and the historical synchronization success rate as the state transition weight to generate the state transition matrix; Based on the multi-channel structure, the state transition matrix is ​​transformed and weighted fusion is performed to obtain the cumulative propagation weight between nodes. The state transition probability is calculated based on the ratio of the cumulative propagation weight to the node processing capacity, and the optimal state transition sequence is constructed. The historical synchronization features between nodes on the propagation path are used to extract deep features through a multi-layer skip connection network, and the deep features are combined with the accumulated propagation weights to obtain the path evaluation value; The optimal number of propagation hops is selected based on the path evaluation value, and the prediction model is trained using the differential strategy optimization method. The prediction model is used to calculate the optimal preceding node of the node; Starting from the target node, the optimal preceding nodes are connected sequentially until the source node, generating the optimal propagation path with the maximum cumulative propagation weight.

5. The method according to claim 1, wherein Use the public key to verify the digital signature of the change request, update node information according to the change request synchronization sequence, divide the network into slices based on the directed propagation graph, perform consistency checks based on the node hash values ​​in the hierarchical cache pool within the network slice, generate error detection values ​​based on the data check code of the main storage pool between network slices, build an anomaly propagation graph based on the error detection value to locate the fault source node and correct the status information, including: Receive change requests, verify digital signatures using public keys, and update information in nodes according to the change request synchronization sequence; Calculate the ratio of the sum of the node neighbor weights in the directed propagation graph to the global weight, aggregate the nodes that meet the preset slice threshold conditions into network slices, and establish cross-slice connection relationships between network slices; In the network slice, the node status, timestamp and previous node identifier are concatenated to obtain a node feature sequence. A weighted hash operation is performed on the node feature sequence to generate a multidimensional hash value. The ratio of the intersection cardinality to the union cardinality of the multidimensional hash values ​​between nodes is calculated as the consistency measure. Generate a check code for the network slice data, and perform an XOR operation on the check code and the check code in the main storage pool to obtain an error detection value; Extract abnormal nodes whose error detection values ​​exceed the preset fault tolerance threshold. Use the intersection nodes of abnormal nodes and change request propagation paths as vertices of the abnormal propagation graph. Build an edge weight matrix based on delay correlation and construct the abnormal propagation graph through eigendecomposition. Calculate the weighted out-degree sum of the nodes in the anomaly propagation graph and select the node with the largest weighted out-degree sum as the fault source node; Correct the status information of the fault source node based on the checksum in the primary storage pool and the node backup data.

6. The method according to claim 5, characterized in that Extract abnormal nodes whose error detection values ​​exceed the preset fault tolerance threshold. Use the intersection nodes of abnormal nodes and change request propagation paths as vertices of the abnormal propagation graph. Build an edge weight matrix based on the delay correlation. Construct the abnormal propagation graph through eigendecomposition, including: Obtain the error detection value sequence of the node, and mark the node whose difference between the current error detection value and the expected error detection value exceeds the preset fault tolerance threshold as an abnormal node; Construct a time-decay weight, multiply it with the error detection value sequence, and normalize it to obtain the abnormality degree of the abnormal node; Obtaining the propagation path of the change request, and calculating the intersection node of the abnormal node and the propagation path as the vertex of the abnormal propagation graph; Calculate the propagation delay between vertices, substitute it into the exponential decay function to get the delay correlation, and count the number of propagations between vertices to get the propagation probability; The delay correlation, propagation probability and consistency measure are combined according to the preset weight coefficient to construct the edge weight matrix; Perform eigendecomposition on the edge weight matrix to obtain eigenvalues ​​and eigenvectors, and select the product of the square root of the eigenvalue and the eigenvector to construct the node propagation eigenvector; The node association strength is obtained by calculating the inner product of the node propagation feature vector and the modulus ratio, and the vertex pairs with node association strength greater than a preset strength threshold are taken as edges, and the consistency measure is used as the edge weight to construct the abnormal propagation graph.

7. A real-time synchronization system for personnel in a cross-organizational approval system based on public and private keys, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to receive personnel information synchronization requests from organizations and obtain information request queues; The second unit is used to generate a unique public-private key pair for each organization based on the information request queue, distribute the public key to other organizations, and encrypt and store the private key locally; The third unit is used to obtain personnel information from the information request queue to build a node set, and predict and mark active nodes through the long short-term memory neural network; A multi-level hash tree is constructed based on active nodes. A hierarchical perturbation factor is introduced to write data into the hierarchical cache pool. The data checksum is calculated and submitted to the main storage pool. A distribution feature map is constructed based on the similarity of node hash values. The fourth unit is used to digitally sign the received personnel information change content using the private key to generate a change request, construct a multidimensional association matrix based on the distribution feature map, construct a directed propagation graph through feature decomposition, perform dynamic programming with multidimensional attention to obtain the optimal propagation path, generate a synchronization sequence, and verify data integrity through the main storage pool; The fifth unit is used to use the public key to verify the digital signature of the change request, update the node information according to the change request synchronization sequence, divide the network slices based on the directed propagation graph, perform consistency checks within the network slices based on the node hash values ​​in the hierarchical cache pool, generate error detection values ​​between network slices based on the data check code of the main storage pool, and construct an abnormal propagation graph based on the error detection value to locate the fault source node and correct the status information.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images